Sahill Poddar is the Co-Founder and CEO of Parafin, the embedded capital platform behind DoorDash, Jobber, Mindbody. Most recently, Parafin raised a $100M Series C at a $750M valuation, led by Notable Capital, and joined by Ribbit (backer since the seed), Thrive, Redpoint, and Pathlight. Since then, they’ve crossed $100M in revenue and raised >$500M in additional lending capital, including a recent $150M facility led by Goldman Sachs.

Before Parafin, Sahill led machine learning at Robinhood and ran revenue growth at Meta. He’s no stranger to quantitative undertakings — he began his career as a physicist at CERN, where he contributed to the work that discovered the Higgs boson. Thanks to Parafin’s success, Sahill now has a view into dozens of vertical platforms from the inside — an ideal vantage point for pattern-matching what’s working in Vertical AI.

Watch the Full Episode Here

Today’s Episode

In software, channel strategies tend to be most effective once you’ve hit ~$10-15M ARR. You build the product, acquire users, prove unit economics and then consider channels — learn how to sell it yourself before asking others to. Parafin did the opposite. From day one, the company took the embedded route — well before vertical embedded was a thing. With its capital products lively natively inside vertical platforms, it is able to touch hundreds of small businesses that ultimately receive the capital for each channel partner it adds. No consumer brand, no Google Ads, no outbound sales team targeting SMBs.

This is an unusual strategic choice: in early days, it can be expensive and risky. You need large platforms to scale quickly, yet they may hesitate to bet on a startup. Once in, you require partner buy-in to drive adoption amongst end-users. Each subsequent partner’s needs — and user bases — may not mirror the next.

In Sahill’s case, the problem was clear, universal, and non-core enough for its customers that it worked. But the real payoff of the channel-first GTM lies in information asymmetry. Parafin now underwrites across dozens of vertical platforms spanning restaurants, fitness studios, auto shops, home services, and payroll. That cross-platform view reveals patterns about which types of SMBs are healthy and why. And importantly for our episode, along with that comes an understanding of which platforms are building real defensibility and loyalty in their verticals, and which may be exposed to AI disruption that remains in its earliest stages.

The Audacity of a Channel-First GTM

Sahill went after the white whale first: DoorDash. “Everyone’s on the call,” Sahill says about the earliest pitch. The platform’s product, risk, legal, and executive teams all had to say yes. DoorDash was public. Parafin was a seed-stage company.

The reason Sahill was willing to take this harder path because he’d developed a thesis on why direct-to-SMB lending had been a graveyard — Cabbage, OnDeck, BlueVine — to date.

1. Adverse selection. “You don’t really want to give money to people searching for money,” Sahill says. Pull-based distribution systematically attracts lower-credit borrowers.

2. Cost of marketing. Paid ads in the lending space are costly, and perform poorly in many verticals.

3. Commoditization. Every SMB lending competitor underwrites on the same public data, leaving price as the only differentiator, inviting a race to the bottom. There’s always someone willing to cut corners to win on price.

Channel-first would eliminate all three problems at once. For Parafin, customer acquisition cost is zero1 because the platform already owns the merchant relationship. Underwriting runs on proprietary data (real-time transaction volume, retention patterns, seasonal behavior) that no one outside the platform can access. And offers are pushed to merchants based on actual performance, not pulled by the desperate. Parafin has had zero enterprise churn since inception.

The product advantage of true embedding is extreme, especially in lending. Parafin doesn’t just give its partners as referral link to a form, as did Cabbage with Airbnb, and many other SMB lenders before it. The platform-native nature of Parafin’s product enables a different experience: pre-approved offers and 10-second decisioning. By using the platform’s granular transactional data to inform underwriting, they deliver better loan performance and a lower-friction, higher-trust merchant experience. And that matters for channel partners. As we explored in a prior episode — Vertical FinTech, with Rahul Hampole at ServiceTitan — fintech products at the right moment can dramatically expand platform ARPA, NPS, and retention. The better Parafin does, the happier its platform partners.

A personal note
This week, the partner that makes this show possible is also our guest. Parafin has been the presenting sponsor of Verticals since we launched. Working with Sahill and his team has been a lesson in founder foresight, and we are incredibly grateful for their support. We’ve shared many call-outs highlighting Parafin’s success — the revenue & retention they can bring to your platform. But working with a partner on something like lending is just as much about trust and relationship as it is performance. In our experience, Parafin offers a rare combination of both. So whether or not you’re ready for the product today — reach out. They’re a long-term partner who want to see you (and the larger Vertical AI ecosystem) succeed.

Learn more about Parafin today →

What the Best Vertical Platforms Look Like from the Inside

Because Parafin sits inside the financial plumbing of dozens of platforms, Sahill has had an inside view into what works and what doesn’t in the vertical business model.

The best Vertical SaaS are collective bargainers for all their businesses when it comes to negotiating with their vendors for services. And the harder they bargain, the better they get.

The strongest platforms he’s seen don’t just sell software — they aggregate fragmented SMBs into an effective Group Purchasing Organization. They negotiate on behalf of their merchants with payments processors, lenders, insurance providers, and suppliers, then pass those savings through. More merchants means more bargaining leverage, which means better terms, which attracts more merchants. It’s a flywheel purchasing power, a virtual Group Purchasing Organization.

Product-wise, Sahill sees a bimodal distribution of user engagement across the platforms he works with. The majority of merchants either use 1-2 features or 6+. Few in the middle. Lending turns out to be one trigger that can move a merchant from the point-solution cluster to the platform cluster. A restaurant that takes a Parafin advance to buy a second oven starts paying more attention to the platform’s other tools — driving ARPA and retention.

If it’s Written Down, AI Is Coming for It

Sahill has a fairly straightforward view on the positioning of various verticals against model labs.

If it’s white-collar, AI is coming for it in a very hard way. But if it’s a platform meant for blue-collar businesses or that industry, it’s very difficult for AI to come for it. [In legal in finance] everything worth knowing is written in some document that AI can read.

The likes of Harvey, Legora, or Norm AI have had a faster path to efficacy precisely because most first-order data in legal and compliance is already digitized. But that availability is a double-edged sword.

Home services, fitness studios, and restaurants are a different story. The critical insights — a plumber’s no-show rate spikes on Fridays, pizza box logistics cost more than the POS system, gym class schedules should shift based on weather patterns — live predominantly in practice, and less so in documents.

Luke visualized a “screen time” heuristic: the fewer minutes per day operators in a vertical spend on a screen, the harder it is for a general-purpose AI to find trainable data. A teacher on MagicSchool might use software 30-60 minutes a day, but the value of the product is powering the 6 hours of non-screen work around it. Jobber’s AI appointment-confirmation bot may seem like a commodity — but the company’s foothold in home services might yield an edge over platforms that don’t understand the specific, predictable moments and reasons trades clients cancel.

Some of the most defensible Vertical AI platforms will access difficult-to-aggregate data — or better yet, create it de novo, as a function of their product. Parafin is leaning into their edge with Paraformer, a proprietary transformer model for time series analysis of SMB cash flows. It powers AI agents for fraud detection, lien checking, and entity verification. As we noted in Payments as a Shadow System of Record, such industry financial layers can become stickier than traditional systems of record because of their access to proprietary information across customers.

Sidebar: The Vertical AI Utility

Parafin is embedded fintech. You might also call it Vertical Infrastructure. What Sahill has built also reminds us of an emerging category we’ve been thinking a lot about lately at Euclid: the Vertical AI Utility. It’s infrastructure that scales with consumption, works whether the platform calls it directly, embeds it, or connects it via an agentic layer. It doesn’t compete with core application layer — it makes such platforms more valuable by doing something the platform can’t (or doesn’t want to) do alone.

Utilities aren’t glamorous, but they can be difficult to displace once they’re wired in. As the owner of a function few deem their core competency, they often have the breathing room — and data access, across many customers — to compound advantage.

As the likes of Stripe have demonstrated in broader fintech, that may give you a ubiquitous foot-in-the-door to offer adjacent infra as well. We believe that as Vertical AI markets grow, so will the need for such utilities. Many will be relatively industry-agnostic — or perhaps more customer-size-oriented, as is Parafin. Others, perhaps in spaces like manufacturing, healthcare, or finserv — will likely cater on a sector basis. It’s one to watch.

The Takeaway for Vertical Founders

Channel-first can be a cold-start problem from hell. You need a platform to trust you before you have proof; you’re building for someone else’s customer; the integration is deeper; and distribution & feedback runs through an intermediary.

But the compounding advantages are enormous. You inherit the platform’s distribution, trust, and — perhaps most importantly — data. These layers of stickiness have never been more important than now, in the era of Vertical AI. As Sahill put it, “if your platform is predominantly just UI and database and that’s all you got, yeah, it’s probably gonna be a tough road.”

Increasingly, defensibility comes from things you and you alone have earned the right to learn.

See you next week.

The artifact is a room, which is awkward for a museum: 90 Power servers stacked across ten racks. IBM named the cluster Watson, after its first CEO.

In February 2011, IBM put Watson on Jeopardy!, where it defeated Ken Jennings and Brad Rutter, two of the show’s best players. Jennings took the loss well. “I, for one, welcome our new computer overlords,” he wrote on his Final Jeopardy screen.

Watson, it turned out, had peaked at trivia.

The Job

After beating Ken Jennings, Watson got a promotion: oncology.

IBM split the medical push between two famous cancer centers. Memorial Sloan Kettering would teach Watson for Oncology to recommend treatments. MD Anderson hired IBM to build a related system called the Oncology Expert Advisor. The assignment sounded simple: ingest the patient information, review the literature, and return a ranked list of options supported by evidence.

Oncology was less cooperative. A treatment decision can turn on tumor type, stage, mutations, prior therapies, other conditions, incomplete tests, available drugs, new evidence, and what the patient is willing to endure. Two reasonable oncologists can see the same case and choose different paths. In contrast, Jeopardy! gives you a complete clue and one correct answer. Patients are less considerate.

The Clever Bit

Much of Watson for Oncology’s later training came from hypothetical cases written by MSK doctors, with the recommended answers attached. It became less an oncology oracle than an extremely expensive impression of one Manhattan cancer center.

That imitation traveled badly. Internal IBM documents recorded multiple “unsafe and incorrect” test recommendations. Doctors outside the US found that some recommended drugs weren’t available in their countries, or that the advice conflicted with local protocols. In one case, Watson suggested a drug with a severe bleeding warning for a patient who was already bleeding. MSK caught it before a real patient received the advice. On Jeopardy!, a wrong answer costs you money. Healthcare has a less forgiving scoring system.

At MD Anderson, the failure was less dramatic but more expensive. Its Oncology Expert Advisor struggled to ingest patient data, integrate with the hospital’s electronic record system, and produce useful recommendations. MD Anderson spent $62.1 million on its state-of-the-art oncology AI. Five years later, after only testing it on old cases, the hospital pulled the plug. It never treated a single patient.

The invoices, at least, were in excellent health.

Why It’s in the Museum

Watson belongs here because answering a question correctly and doing the job are not the same thing.

Many studies reported concordance – how often Watson agreed with doctors. But agreement wasn’t the target outcome. It showed no clinical usefulness, and most of the time during training, doctors would close Watson and keep working. A second opinion that is useful only when it agrees with the first is mostly a very expensive nod.

By the end of 2020, IBM shelved Watson for Oncology. Two years later, it sold much of Watson Health’s data business for parts.

Kushim’s clay tablet survived millennia. Watson just barely survived a decade.

The Catalog So Far

Exhibit 01 — The Beer Ledger of Uruk
Exhibit 02 — The AI that Peaked at Trivia

Next up: the machine that ruined Mark Twain’s life.

Rick Zullo is the Founder & Managing Partner of Equal Ventures, a seed-stage fund with over $200M in committed capital investing across energy, insurance, retail, and supply chain. Rick has been investing in vertical software, marketplaces, and services for 14 years. One of his earliest exits was Vettery, a staffing platform acquired by Adecco, and his portfolio includes companies like Equal Parts, Threeflow, and Odyssey.

Rick brings a Munger-inspired value-investing lens to venture, focused more on competitive advantage and long-term free cash flow than most in the asset class. In this episode, we dig into one of his core convictions: in a market flooded with AI-driven growth, much of the growth managers are are seeing is beta, not alpha. Caveat emptor.

Rick has also build a defined practice around services, well before “AI-native services” was a category. He’ll share his POV on why it’s never been easier to start a services company… and never been harder to build one that lasts.

Watch Full Episode Here

And don’t forget to subscribe on YouTube to get new episodes every week!

Today’s Episode

Venture capital is strange business these days. AI-native startups are growing faster than software ever did. Services — historically anathema — are in vogue. Physical AI and defense tech are absorbing billions, despite hardware being a VC graveyard and defense LPA-restricted for decades. Valuations (and hence revenue multiples) at seed have detached from any rational framework. And the narrative has shifted so dramatically that some LPs are questioning their exposure to software and embracing “deep tech” (although the vast majority of that category is still software, making that dichotomy deeply confusing).

Naturally, the instigator of these shake-ups is AI. Rick Zullo believes many investor assumptions in this moment are misplaced. His argument: the current boom is producing the illusion of product-market fit at scale. When a technology wave lifts every company in a category simultaneously, you can’t mistake it for idiosyncratic company success. Conflating beta with alpha1 can turn out very costly.

Beta-Driven Product Market Fit

Rick has a term for what happens when a technology catalyst makes everything in a sector work at once: beta-driven product market fit. “Because of a technology or market catalyst, everything in a single category is working,” he explains. “That long-term means a lot of these companies are going to struggle as they become more well-capitalized, compete against each other, and invite more competition.”

He feels the AI-Native Services wave is a perfect example. When an LLM makes it trivially easy to automate insurance claims processing, and five companies all go from zero to $5M in six months, the growth looks like signal. But it might not be the signal founders are looking for. When the technology wave lifts your boat, it lifts competitors and incumbents too.

The fundamental distinction between software and services still matters. Historically, SaaS entered greenfield markets: the product was new, adoption required friction, and success in overcoming that friction yielded defensibility. Services companies operate in brownfield territory. “You’re largely trying to steal share from other competitors in the industry,” Rick argues. “That’s a very different game.” The incumbents have distribution, cost of capital advantages, and they’re implementing AI as aggressively as anyone.

The insurance TPA market is his clearest case study. Equal has studied it extensively and found what you’d expect in a mature services industry: tremendous multihoming behavior, low switching costs, and a cost curve that every player is riding down together. “Are you going to assume that because you have early growth, competitors aren’t going to respond?” In most services categories, they already are.

In a world where buyers are trying to figure out how to use AI, it makes sense to be more consultative. Forward-deployed everything is all the rage, for good reason. When used for the right reasons — to find a unique, defensible, compounding application of AI — it’s powerful. When the footing is win business and hope the product will come, it’s dangerous.

Summarized: SaaS revenue was hard to get, so strong growth meant product-market fit. If AINS revenue is easier to get, it’s likely easier to lose as well. So VCs touting their zero to $5M ARR as proof of market leadership and justification of valuation might be in for a rude awakening.

What if there was a partner that was built for vertical platforms, that could grow your revenue, and bring huge value to your customers? Enter Parafin — the benefactor behind Verticals. They are the leading embedded lending platform, working with top logos like DoorDash, Gusto, Amazon, and TikTok Shop. From working capital to Buy Now Pay Later, they have your customers’ backs. And their own AI-powered underwriting & analysis has the hard part covered. They’re smart, supportive, and growing like a weed. Get set up with Parafin today →

AINS & The Revenue Multiple Trap

If beta-driven PMF is the disease, revenue multiple investing is the symptom. Rick is blunt: “Revenue multiple investing in services has never worked.” The reason is mechanical. Services companies carry cost structures that software companies don’t, even fully automated ones. The question isn’t what your gross margin looks like in year one.

The long-term question is: what is your net margin? …At early stage, it’s really hard to tell what net margin is unless you understand the moat structure of that industry.

He sees investors underwriting AI services companies at 500x to 1,000x revenue, in some cases calling GMV “ARR.” Hidden costs compound the distortion: companies loading up on risk assets and keeping the cost of capital below the line, out of the gross margin calculation. In insurance, opex optimization barely moves the needle. What matters is loss ratio, and most AI-enabled carriers have no idea what their loss ratio will be until their book matures. “A lot of them are trying to outrun their book with growth,” Rick says. “We saw that play through in 2021.”

WeWork, for example, had strong revenue growth, positive early margins, and a compelling narrative. Benchmarked against any legacy REIT on unit economic ratios, it was terrible. Rick sees the same trap in AI services: companies that look exceptional in isolation and mediocre when compared to the incumbents they’re supposedly disrupting.

As we’ve argued many times before, the “death of software” narrative misses so much nuance that it’s more or less worthless. Rick believes the same effect is at play with the blind trust many VCs are placing behind AINS. Like in all categories, there are of course excellent businesses. But if the market broadly prices services revenue as if it carries the same margin structure and defensibility as SaaS ARR — when much of the time, it doesn’t — that is a potentially fatal mistake.

Venture’s Blackstone Problem

So why is capital flooding into categories with these competitive dynamics? Rick’s answer sits upstream. It’s an incentive problem that starts with VC fund economics.

“Big funds are starting to look a lot more like Blackstone,” he observes. “They’re looking a lot less like Benchmark.” When you manage $10B at 3-and-30, you need companies that consume hundreds of millions in capital. Physical AI, deep tech, defense, nuclear, autonomous vehicles: these are plays that require the deployment of billions. “It’s not [just] convenient,” Rick says of this alignment. “It’s actually extremely strategic.”

Taken to an extreme, the result is a market that systematically favors capex-intensive plays and penalizes capital efficiency. Veeva Systems, which raised $9M before its IPO and has been compounding cash ever since, “would not exist in this market,” Rick argues. Not because the opportunity isn’t there, but because no large fund can write a meaningful check into a company that doesn’t need one. As we explored in The Fundability Trap, the market increasingly rewards fundability over investability. Rick extends that critique to capital structure itself: funds that need to deploy billions will find theses that require billions, whether or not those theses create durable value.

Physical AI is the latest expression. Rick acknowledges that cornered physical resources (cameras in warehouses, sensors on shipping routes) can create genuine moats. He likes how Steel Atlas put it in their “data gusher” thesis: limited providers, expensive-to-replicate infrastructure, sensors that throw off valuable data, software that compounds on top. But the broader category gives him pause. Most physical AI companies pitching AI-optimized mining, energy services, or manufacturing face incumbents with lower cost of capital, deeper distribution, and decades of vendor relationships. A truly defensible, proprietary hardware wedge is one thing. Off-the-shelf hardware plus compute is another. GPU capex, unlike railroads or transmission lines, has a useful life of two to three years. The analogy to durable infrastructure is questionable.

At Euclid, we’ve discussed hardware as a moat in Vertical AI, and why unique datasets have become more important than ever to the AI application layer. We do find it interesting, however, that deep tech proponents point to data generation as a key source of defensibility for the theme. Even the Data Gusher piece above lists Palantir as the canonical success story of a “data factory” — a Vertical AI business that is 100% software, eschewing both proprietary hardware and infrastructure by design. Deep tech or physical AI may be a compelling way to deploy capital at high velocities, but its unit economics and data advantages largely remain to be seen.

The Quiet Compounder

Rick’s answer is his portfolio. Between a third and half of Equal’s Fund I companies are free cash flow positive while growing at rates above 100%. These aren’t companies in the venture zeitgeist. Most aren’t raising $100M rounds or getting profiled in The Information. They’re vertical businesses that found defensible positions in their markets and are compounding quietly.

Equal’s internal filter centers on one question: “Can this company hit $100M of gross margin within the next five to seven years?” Very much in the vein of the Bling Capital underwriting framework we explored with Kyle Lui a few weeks back. The GM filter alone eliminates most AI services companies raising at astronomical multiples, because it forces you to confront the margin structure rather than assume it away.

And it all comes back to fund economics. Rick’s “dragon number,” the exit value needed to return the fund at average ownership, is roughly $570M. A $1B outcome still returns 2-3x the fund. That is a fundamentally different game than a megafund that needs $100B outcomes to move the needle. Founders Fund, by comparison, has invested in 600 companies, and six really matter to them, at average entry points well above $1B on a dollar-cost-average basis.

There are many ways to draw up a 10x fund in venture capital: 4 vs. 40 investments a year, $5M vs. $500M fund size, 20% vs. 2% ownership. They can all work. But the math has to hold together — and that math determines what the managers have to prioritize, and therefore, founder alignment. We laid out all the numbers in our essay, The Quiet Death of Founder-VC Alignment.

Hope is not a strategy. Yet VC fund strategies that assume $100B exits or capturing double-digit percentages of all venture-backed exits abound. Rick sees a landscape in which most AI services companies have zero chance of monopolizing their sector, and most physical AI plays face incumbents with structural advantages. The contrarian position might be the most old-fashioned one: invest in models that have actually been proven to make money.

The Takeaway for Vertical Founders

If you’re building in Vertical AI and everything seems to be working — revenue is growing, customers are signing, VCs are circling — first of all, congrats. But second, take a beat. Beta-driven product market fit means the rising tide is lifting your boat alongside everyone else’s. The question, to paraphrase the Buffet quote we all know and love, is what happens when the tide goes out.

Most high velocity-startups prioritize early commercial success, and more than ever, we feel that’s the right move. Vertical AI adoption is still nascent, economy wide. You’ve got to get in, grow, and figure it out. But along the way, the founders who build durable companies in this cycle will work just as hard to answer the long-term questions. Here are Rick’s top three:

1. What is your path to not simply a market-leading position, but a monopoly?

2. What does your net margin look like at scale, not just gross margin?

3. What’s defensible when every incumbent in your sector is also implementing AI?

As we’ve explored at The Verticalist many times before, workflow depth and data gravity are earned. Early market pull is critical, and the right path to product-market fit. But don’t mistake it for a moat. If you’re not leveraging traction to build something structurally defensible, you’re running a race that gets harder every quarter. Assume you have more competition than is visible, because beyond startups, incumbents will take their shots too, even if they’re slower to materialize.

On the capital side, we also agree with Rick on this: pick aligned investors. Beyond personal simpatico, a fund’s economics should match your company’s trajectory, stage, and vision. If you’re building a capital-efficient Vertical AI compounder, the partner who needs you to raise $200M and exit for $20B is not your partner. The one who will celebrate a $500M exit at 15% ownership is.

Leave a comment

Alex Oppenheimer is Founder and General Partner at Verissimo Ventures, a pre-seed and seed fund investing across the US, Israel, and Europe. Before launching Verissimo in 2020, Alex spent years at NEA, where he became a specialist in SaaS mechanics. He’s written 80+ checks since, backing vertical software, infrastructure, and fintech companies at the earliest stages. Tune in for a sharp debate on whether the vibe coding wave is actually a threat to Vertical SaaS, or just to the cheap end of it — and where the new lines of build vs. buy are being drawn.

Watch the Full Episode Here

Today’s Episode

Software is dead. We’ve all heard it. AI has made it trivially easy to spin up a tool — a scheduling widget, a personal CRM, a dashboard — and the natural conclusion is that nobody needs to buy software anymore.

While the conclusion is wrong, it’s not for the reasons most people assume. What AI can replace and what’s worth investing in are two very distinct categories. The cheap end of the market (discrete, small-scope tools) and the complex end (truly enterprise-scalable platforms) are night-and-day. They have wildly different build times, maintenance costs, and ultimately, economics. Many have assumed that all of software should take a hit — in reality, enterprise solutions may have gotten harder to build in the AI era, with rising expectations of interoperability and customization. While the simple solutions, designed for prosumers to execute on a small, repeatable task, have more-or-less lost pricing power altogether.

Alex has been investing in SaaS nerly his entire career. He joined NEA in 2013 when calling yourself “the SaaS guy” was meaningfully differentiating. He’s since written 80+ checks and founded Verissimo, investing at pre-seed and seed across three continents. His vantage point as a first-check software investor with deep finance roots, an engineering degree from Stanford, and a penchant for analysis aren’t foreign to the venture world — but his grounding in those fundamentals even as AI takes the world by storm, re-writing conventions, is somewhat refreshing.

The Cheap End Is Exposed

Vibe coding may be a tongue-in-cheek term — but its impacts are very real, both tangible and psychological. People who never thought of themselves as developers are now shipping functional (if not scalable or enterprise-ready) software. the claim isn’t that software development is completely democratized, but that the mindset of entrepreneurial people — builders with ideas but lacking coding expertise or budget — has shifted. even if they don’t build all their solutions, their expectations for ownership, flexibility, and scope are wholly new. Halfway through the episode, Alex pulls up a photo of his barber rebuilding his entire booking and scheduling system in ChatGPT in between haircuts.

A guy who cuts hair is shipping software. Most barbers or mechanics or teachers aren’t — and probably never will — spend their time on technology. But 1% of them will. and over time, we imagine that will change the psychology of industry. When people see that it is possible for fellow domain experts to mold the software tools of their trade (if not own them), that shifts expectations.

“It works for me” is a long jump from “it works for thousands of users”, however. Edge cases, compliance, integration depth, multi-tenant architecture: the requirements compound at every order of magnitude of software scale. Anyone who’s shipped enterprise software knows this. Anyone who hasn’t dramatically underestimates it, because their only reference point is themselves —an inherently, you’re more likely to forgive your own product for being clunky.

I would like to see somebody rebuild a very complex, enterprise-ready solution, vibe code it themselves as a non-developer, and then scale it up to be a billion-dollar business. Just not gonna happen.

Around small tools (which can be vibe-coded, cheaper, even nearly free) but also around the systems of record in their space (which should be more interoperable, flexible, and customizable).

There is an 800-pound gorilla lurking behind the vibe-coding wave: economies of scale. Salesforce is able to spread its infrastructure cost across tens of thousands of users. Alex painted the picture of what is probably happening to most of those barber-developers who first try to scale their product out to many more users: “I churned my Salesforce subscription and now it costs me $80K a year — I had to hire three developers instead of just paying Salesforce $25K a year and being done with it.”

Beyond maintenance and product lift of solving edge cases for every new customer, you’ve got actual hard infra: your costs on Vercel, Supabase, etc. look free at first, then really expensive on a unit economics basis. Paying $1000 a month to support 10 users paying $50 a month isn’t sustainable — it works well if you scale to 100 users quickly, and it doesn’t if you grow linearly. Some builders will simply nail it and blow past the diseconomy of scale moment. Many more will tire out and keep a few select tools, pressuring their core system of record — whose price is getting commoditized down anyways — to enable integrations or open APIs.

A quick word from our sponsor, Parafin — they power embedded lending, cards, and insurance for the likes of DoorDash, Jobber, and TikTok, having extended $35B+ in offers to small businesses. Purpose-built for vertical platforms serving SMBs, no infrastructure required.
Explore a custom program for your platform today →

Big vs. Small, Not Vertical vs. Horizontal

The line on build vs. buy isn’t vertical vs. horizontal — it’s big vs. small.

This holds regardless of whether the software serves healthcare, logistics, or construction. A lightweight scheduling tool for a single barbershop? Buildable. A revenue cycle management platform for a hospital network? Not buildable. And while the operating enterprises themselves are more deep-pocketed, they face the same hurdle of economy as vibe-coders. Is your core competency building and maintaining technology solutions — and if not, do you want to commit to making it one? If the answer is no, you are ultimately going to continue to buy.

If it’s a small, very tangible solution where you can write up a summary of what it’s going to do in one page, that’s probably something you can build. But if it’s not your core competency, and it’s important and it’s complicated — that probably makes sense to pay someone else to do.

The reason is the oldest one in economics: specialization of labor. Does it make sense for an HVAC business owner to spend half their time on what is effectively software development? No. It makes more sense to pay someone who spends 100% of their time on that solution and can amortize the cost across hundreds of customers — especially now that traditional, broad-based SaaS is under significant pricing pressure. If a system of record refuses to lower their price? Someone will fill that gap and offer a cheaper solution, but that someone will still be an organization that is committed to scaling as a primary focus.

Attio is a great example. They are a 7-year old horizontal CRM that has raised over $100M in funding to date from GV, 01 Advisors, Redpoint, and Point Nine. They began as a vertically specific solution focused on the venture capital industry, but went horizontal due to TAM concerns. In doing so, they cut prices down to about a third of comparable solutions and stopped catering solely to VC. But with perhaps the most open APIs on the market, they are arguably more successful selling into VCs today than they were then they were horizontal. Openness and price won out.

The Second TAM Expansion

Years ago, Alex created a widely-shared presentation called “SaaS 101” comparing Vertical SaaS to horizontal SaaS. The framework was clean: vertical players had smaller TAMs but lower CACs. They grew more slowly but more efficiently into what everyone assumed were niche markets.

As Alex puts it, however, what were considered “niche” markets at the time turned out to be anything but. Software’s steady march into the economy meant TAMs expanded under investors’ feet while they were still anchored to old sizing models.

Now, it’s happening again. AI is creating a second TAM expansion for vertical markets. AI taps into and replaces services budgets, not just software budgets. When a Vertical AI product can do the work of a consultant, an analyst, or a back-office team, the addressable market expands beyond prior software spend to include the total cost of labor and services.

Some verticals still sit at single-digit SaaS penetration. A few may skip the SaaS generation entirely and go straight to AI. The retrospective pattern matching that led VCs to dismiss vertical markets a decade ago — “too small,” “too hard to sell into”, “not enough historical wins” — is the same pattern-matching causing them to underestimate Vertical AI today. In a past episode of ours (Why TAM is the Wrong Question) along the same lines, Nate Baker suggested the question most are inherently asking is: “is the initial wedge TAM big enough?” But here’s the one you should be asking: “can you durably keep expanding what you do for the customer?” AI is widening that answer — and even time it does, another market perceived to be too small, crosses the threshold of meeting venture-scale standards tomorrow.

Survive to Iterate

Alex tells every pre-seed founder they should have one mantra above the rest:

We survive to iterate.

These four words run directly counter to what most large funds will tell early-stage companies. “When the big funds come in at really early stages, their advice mostly centers around scaling,” Alex said. “Hiring and being aggressive around sales, without really understanding that the product’s not baked yet. And if you’re truly an early-stage investor, the market’s probably not ready yet.”

You have to be there when the market is ready. That means being there before it is. Which means it won’t look interesting right now.

This connects to a broader dynamic both of us observed: the current fundraising market is bifurcated. The average founder pitch has gotten better (more resources exist, polish is higher) but the median hasn’t moved much. Alex put it well: “the average has gone up, but the median has gone down in terms of quality.” You can raise $10M on three slides at an $80M valuation. Or you can raise at a multi-billion dollar valuation on real traction. But the middle, where most real companies live grinding from early product to product-market fit, is a valley of death.

The founders who survive that valley in verticals are the ones with earned insight. They didn’t graduate and decide to be founders. They spent years in an industry, identified a specific problem, and had the patience to iterate through years when the market was still forming. That type of judgment usually can’t be replaced by technical fluency alone.

The Takeaway for Vertical Founders

The build-vs-buy question will be asked again every time tools improve. And every time, the same mistake will be repeated: people will confuse what they can build for themselves with what can be built at scale for an industry.

If you’re building something simple enough to summarize on one page, you’re competing with every operator who can now ship their own version in a weekend — technical or not.

If you’re building something complex enough to require deep domain knowledge, multi-party workflows, edge case handling, solid scalable architecture, and years of data accumulation — vibe coding just isn’t a comp (yet). That said, they journey will never be the same. Low-end tools in your space are losing pricing power. Baseline systems of record are faced with unprecedented pressure to lower pricing. Expectations around interoperability and customization are rising, fast. The value of truly proprietary data — especially those that have near-term, defensible impact on outcomes — is an order of magnitude more important. And the market opportunity in vertical you’re serving isexpand again, as AI reaches beyond software.

The new dividing line of build vs. buy is big vs. small. Build accordingly.

See you next week.

Naturally, we’re starting with the first vertical killer app: beer rationing.

A brewery manager five thousand years ago pressed symbols into wet clay. One is a pointed jar filled with wavy lines: beer. Another shows a bowl tipped toward a human head—the sign to eat, marking this as a ration account. These are among the earliest writings found anywhere in the world.

It was pressed around 3000 BC in Mesopotamia, modern-day Iraq. The tablet records beer issued to workers as daily rations—a thick, starchy brew, essentially liquid bread.

The Job

A Copper Age village of just a few dozen people had no need for formal records. By the late fourth millennium BC, the first city-states emerged in southern Mesopotamia, some with populations approaching 100,000, defined social hierarchies, and specialized divisions of labor. Every individual had specific obligations and entitlements. Questions arose: who owed what to whom? What was stored, and what disbursements were required?

And this wasn’t one lonely beer tablet. Another recovered tablet tracked 37 months of barley deliveries to an official named Kushim. One set of tablets recorded barley delivered to the brewery; another recorded beer leaving as rations. The first vertical system of record arose from the need to manage this inter-organizational ledger.

Kushim may have been a person, a title, or an office. If it were a person, Kushim is a candidate for the earliest surviving personal name. History’s first recorded name survived five millennia by diligently signing off on inventory statements.

In full disclosure, Kushim’s work had flaws. Early arithmetic was error-prone: it had just been invented! The first name recorded in history also committed the first accounting mistake. The profession has been living it down ever since.

The Clever Bit

The first system relied on tokens, but the Sumerians recognized that their complex operations required a centralized ledger. Unfortunately, the general-purpose alphabet would not arrive for millennia. So the vertical schema shipped first: they recorded objects—beer, barley, workers, offices—and quantities.

The scribes of Uruk initially labored under an accounting regime that would give any modern auditor a migraine: not one number system but roughly a dozen, each tied to its own class of goods. The same wedge in the clay, in other words, meant one thing for barley and another for sheep — measuring jars rather than any universal quantity. Historians have identified nine types of beer tracked by Sumerian scribes of this era so far, mostly through signs depicting various types of beer jugs.

The upshot is that the first ledger was counting things, not numbers. The abstract integer being a luxury they had not yet thought to invent. But over the course of about a thousand years — and plenty of beer — the Sumerians would evolve their humble ledger into a mathematical powerhouse: the sexagesimal (or Base 60) place-value system.

Base 60 was perfect for beer accounting — over our decimal, or Base 10 system — because sixty divides so many ways that ugly fractions come out clean. It’s great for quotidian counting systems and it’s not as unintuitive as it seems. You have Kushim’s ledger, after all, to thank for there being 60 seconds in a minute.

Why It’s in the Museum

Calling this a vertical system of record is doing a little work, but the job is unmistakable: raw materials in, finished product out, and the people, obligations, and records connecting them.

In the end, the ledger outlasted the beer. Keeping it required the development of one of the earliest traditions of writing and bookkeeping. Their notation system developed into cuneiform, which combined word signs with syllabic signs and remained in use for more than 3,000 years. Some 1,300 years after the beer tablet, the Code of Hammurabi was carved in cuneiform on a seven-foot rock slab in nearby Babylon. Kushim walked so Harvey could run.

Another important innovation: tracking rations was essentially the precursor to payroll. It also served as a useful form of labor control. Workers were paid rations based on output. Systems of record have had politics since version one.

Every Vertical AI pitch ends the same way: and then we become the system of record. Fair enough. Just remember, the competition is sometimes really old, and some of the first users were paid in beer.

The Catalog So Far

You’ve missed nothing. The museum just opened.

Next fortnight, we jump 5,000 years forward: the AI that won Jeopardy! and promised to cure cancer. It was later sold for parts.

Thanks for reading The Verticalist. If you're building a system of record with a better retention curve than clay, we'd love to hear from you.

Jake Saper is a General Partner at Emergence Capital, where he’s spent 12 years backing companies at the frontier of enterprise software, including early bets on Zoom, Veeva, Ironclad, and Physical Intelligence. In 2024, Jake coined the term AI Native Services in his essay The Death of Deloitte, arguing that AI would fundamentally reshape professional services.

Emergence was also an early bettor on open-source AI infrastructure, leading investments in Together AI, which just raised $800M at an $8.3B valuation, and in OpenAI’s $4B deployment spinout, DeployCo.

Tune in to hear our in-depth discussion with Jake on why the open-weight surge is unbundling the model labs and the coming reckoning over where value actually accrues in Vertical AI.

Watch the Full Episode Here

Today’s Episode

Last week, Chinese model lab Moonshot AI released Kimi K3 — a 2.8T parameter open-weight large language model. Kimi K3 is the first open-weight model to crack the frontier to this extent — scoring first in a popular front-end test (ahead of both Fable 5 and GPT-5.6 Sol) and second only to Fable 5 on another measuring long-horizon knowledge work. And while not “free,” as open-source1 might imply, K3 achieves the above at roughly half the cost per task (and that’s on a relatively favorable Opus 4.8 comparison). K3’s full weights go public on July 27.

K3 is just the latest in a cascade of open-weight models. It is estimated that open models are used by 80% of startups and represent 25-50% of volume across major platforms like OpenRouter and Vercel. Together AI’s annual bookings crossed $1.15B as customers seek to cut ballooning inference costs. The gap between frontier and open-weight performance — once measured in years — has compressed to 4–6 months, and appears to be shrinking.

Jake and Emergence called this shift early, backing Together AI in 2023 on the conviction that enterprises would want the flexibility, security, and cost advantages of open-source models. In July 2026, the question is no longer whether open weights will be in demand, or matter — it’s what happens when the model becomes less of a moat.

The Tokenmaxxing Hangover

Tokenmaxxing describes a phase of eager, first-wave LLM adoption in which enterprises encouraged pure token usage volume from its employees, presumably to accelerate adoption, value, or just generally prove they were AI-pilled and braced for the future. Some of the results are unsurprising: Uber blew through its annual AI budget in a quarter, Amazon shut down its leaderboard because employees were setting up agents in pointless loops just to climb the ranks; founders bragged about monthly token spends in the millions.

Jake uses a fitting term for the aftermath, which is no doubt a catalyst for the open-weight surge: the tokenmaxxing hangover.

There was an era three months ago when there was a lot of pride that companies would take in how much their token spend was going up as an indicator of — turns out — a very poor indicator of value created by AI. The hangover treatment has been to more heavily scrutinize what those tokens have been spent on and to find effectively cheaper tokens to get the same or better outputs.

Companies that were routing everything through frontier APIs at premium prices are discovering that open-source models — used judiciously, in the right use cases — can deliver comparable quality at a fraction of the cost. This is all pushing the ecosystem toward a multi-model paradigm: frontier closed-source models for the highest-intensity tasks (a small portion of total runs) and significantly cheaper open-weight models for more quotidian tasks (the majority of workloads).

Three forces are accelerating this more diverse future of model use:

1. Open-weight model quality has caught up. Kimi K3 is months behind the leading closed-source labs, not years. And that trend appears to have accelerated. Few deny that foreign open-weight models — less exposed to US IP law — are distilled from frontier models, likely putting a natural limit on how small the gap can get for holistic performance. But the capability difference simply isn’t as noticeable as it once was, and in certain domains, it’s arguably non-existent.

2. The US federal government has shown a willingness to intercede. Both to gate powerful frontier models (as they did with Mythos) and, potentially, to crack down on Chinese open-source models almost certainly distilled from US models.2 Regulations incentivize companies to further control their AI infrastructure.

3. Enterprises are beginning to seek their own AI moats. For example — as we’ve discussed at length here at The Verticalist — capturing, protecting, and embedding proprietary data into their stacks. Beyond this, there’s the possibility of edging into the model layer: fine-tuning an LLM with open weights to their unique purpose. That makes open source not just cheaper, but strategically essential for anyone trying to build a defensible data asset.

A quick word from our partner on the Verticals podcast, Parafin. They just launched Spend, a business credit card purpose-built for SMBs that vertical platforms can deploy in weeks, adding to lending and insurance products already powering DoorDash, Gusto, and TikTok Shop. Named to the 2026 Forbes Fintech 50, with a fresh credit facility led by Goldman Sachs, Parafin is built for founders like you. Explore a custom program for your platform today →

The Labs Respond

The frontier labs seem to be doing their best to couch panic over the open weight news in grave national security concerns. A tweet from Dean Ball, OpenAI’s “head of strategic futures” is making the rounds branding open weights as “inherently decelerationist,” likely to result in “AI communism” that he characterizes as a “dystopian hellscape.” The incentive for frontier labs to support government intervention to suppress direct competition requires no explanation. The deeper response, however, lies in their approach of new layers beyond the model.

OpenAI hired Jason Boehmig, the co-founder and former CEO of Ironclad, to lead its new legal vertical in June, a direct move into application-layer territory. Harvey and Legora have been among its largest customers. Both OpenAI and Anthropic are hiring vertical-specific leaders across industries. And the app layer hasn’t been the labs’ only expansion salvo. OpenAI’s DeployCo raised $4B backed by Emergence, TPG, SoftBank, and Goldman Sachs to embed forward-deployed engineers inside enterprises. AI implementation services are perhaps an even more peripheral adjacency to their model-layer competencies that would create TAM breathing room as infrastructure is squeezed.

While their call for open-weight regulation is louder, the layer expansion feels even more telling about how seriously the labs are taking the potential erosion of their core moats. How successful these new initiatives will be is a separate question we will be writing more on soon. But as we wrote about Claude for Science last week (and OpenAI’s legal product is not dissimilar), the vision is nascent — today, a core model plus a set of connectors. As we discuss with Jake, that’s not what made vertical software companies defensible. While Vertical AI undoubtedly moves faster than its SaaS forebears did, it’s not just a question of product development — those with the strongest ostensible moats (e.g. Veeva, Fair Isaac, Epic) have gone deep on unique workflows and parsing of first-party vertical data for years.

The labs’ dilemma is that their core business is under commodity pressure from open weights. Therefore, moving up the stack to applications and / or services looks like a natural escape hatch — and one they may be uniquely positioned for, if only due to their vast balance sheets. But that means competing with their own API customers in a very existential way. It’s a self-reinforcing loop: the labs go vertical, which makes their customers warier, which accelerates the shift to open source, which puts more pressure on the labs’ core margins, which pushes them further into vertical plays... and so on.

As they look to their own long-term defensibility, even enterprise operating businesses are catching on. As Jake put it, if you’re a pharma company and you give a model lab all your proprietary molecular research, you’re handing your most valuable IP to an entity that’s now hiring vertical leaders and building competing products. “I think for good reason, people are increasingly wary. And that’s another reason why open source is being utilized — so people can keep their data within their house.”

As we covered in The Dispatcher Problem, the big question is who captures value as the marginal cost of intelligence falls toward zero. The recent step-changes in open-weight capabilities and the shrinking frontier gap make that question more urgent. When the model is no longer a meaningful differentiator, value migrates to whoever controls the workflow, the data, and the customer relationship. Perhaps the value also flows to the third parties facilitating all these value-chain shifts (e.g. integrators and other AI services).

We’ve said it before and we’ll say it again — we believe the immutable primitives of app-layer defensibility are workflow depth and data gravity. With more margin to go around, cheaper tokens only make the Vertical AI opportunity more compelling. And in a world of model-layer commoditization, deriving defensibility from the app layer has never been more important. The model labs themselves seem to agree.

Controlling Your Own Destiny

The best Vertical AI companies aren’t waiting for the labs to sort themselves out. They’re already taking steps to cement their app-layer moats — and even to take incrementally more control over their AI infrastructure.

Harvey is a visible example. The legal AI company is training open-source models to encode law firm workflows while simultaneously running a multi-model stack across Anthropic, Google, and OpenAI. The goal is to ensure that no single provider has leverage over the business. As Jake put it, “You’re seeing a lot of the leading agentic software companies focus an increasing amount of their spend on fine-tuning their own open-source models versus relying on the closed-source labs.”

What makes this possible is the convergence we’ve been describing: open-source models are good enough for the majority of workloads, and fine-tuning them with proprietary data is feasible only when the weights are open (unless you want to fund ground-up LLM training yourself). Making proprietary evals3 more valuable than they’ve ever been.

Jake made a related point worth noting: the companies that own the full outcome — AI-Native Services, or AINS4 — may be in a unique position to gather that eval data. A fund administrator like Hanover Park (a past guest on Verticals) sees the complete lifecycle: the AI output, the human review, the client feedback, the audit result. A software vendor selling into the same fund admin sees a truncated slice.5 That’s a separate argument from the open-weight thesis, but it points in the same direction: value is migrating from model access to the proprietary data and workflows built on top of it.

The Takeaway for Vertical Founders

The open-weight unbundling is a control story.

A year ago, Vertical AI companies were largely dependent on one or two frontier labs for model access, fine-tuning, and the roadmap. That dependency is unwinding. Open-weight models today are good enough for many tasks. The cost advantage can run as high as 10x. And the frontier labs may be accelerating the process by making dependency more dangerous, as they seek to compete with their own API customers.

From what we’ve learned so far, the founders navigating the unbundling well are doing three things:

1. Going Multi-Model: running multi-LLM architectures that route by complexity and cost and function. Knowing what works, for what, and why is an edge.

2. Exploring Fine-Tuning: supercharging open-weight models with proprietary data, turning customer feedback into a compounding asset beyond simple reinforcement learning. The app layer strikes back.

3. Depthmaxxing: Maximizing channel checks, customer data depth / breadth, champion relationships, physical proximity (via forward-deployed X), ICP granularity, and general user obsession. Businesses are both greedy and fearful around AI today. They want partners who are on their side and will help them take advantage — but only businesses that go all-in on understanding them will earn their trust (and thus the right) to solve their big AI problems. Founders securing airtight footholds and relationships are finding they can offer more (and more upmarket) than they expected.

As the cost of intelligence falls and the weights open up, the bulk of LLM use recedes into infrastructure. What remains is the vertical knowledge, the proprietary data, and the willingness to stand behind the outcome. As we’ve argued in many past essays, what makes a Vertical AI business durable hasn’t changed — open weights just make the ultimate prize dearer and clearer.

See you next week.

The argument is as follows: AI is commoditizing an increasing portion of software, leaving proprietary data, context, real-world feedback, trust, and distribution as the remaining scarce assets. As a result, vertical startups are incentivized to own the entire value chain. This ambition now extends beyond traditional technology and business services to include AI-enabled home services, construction firms, senior living, and even hospital systems.

NFX concisely summarizes this view in “The AI Wrapper Is Dead”:

The wrapper generation is dead. The survivors and the next generation of startups will think vertically. ‘How do I own every piece of the value chain?’

No doubt, thin wrappers are vulnerable. However, “own every piece” feels like an oversimplification that describes several very different strategies. A company might expand to multi-product, own physical assets, combine hardware and software, operate as a service provider, or vertically integrate into another part of the market. These are radically different strategies with radically different economics.

Our view is that AI expands the feasible boundary of venture-backed startups without determining the optimal one. AI companies should focus on capturing segments of the value chain where data, workflow position, and high-value outcomes accumulate. Sometimes that requires vertical integration, but it often does not.

The form of vertical integration, moreover, matters as much as the extent of it. A company may own the entire stack yet still gather undifferentiated data, depend on the same models as everyone else, and operate with higher friction and worse economics than focused competitors.

Integration maximalism ignores a long history of dominant vertical application-layer platforms that achieved powerful scale and defensibility without expanding into every possible industry revenue stream. Even in the current era of AI, businesses will thrive — and may even be better off — by focusing on controlling the specific portion of work-to-be-done where learning can accumulate.

In Vertical AI, as with SaaS before it, the objective is simple: own the control point. Naturally, control points are industry-dependent and shift over time as technology differentiates some layers and commoditizes others. But the full-integrationist narrative suggests that LLMs have turned every vertical surface into a control point. This is its fatal flaw. In any B2B landscape, there are always a limited number of consequential loci in the customer’s workflow that can serve as the launching point for systems of records.1

Clay Christensen’s Other Theory

Clay Christensen is best known for The Innovator’s Dilemma. He had another, less famous theory of vertical integration and modularity. Christensen first described it in “Disruption, Disintegration, and the Dissipation of Differentiability.” In a 2006 interview with Bloomberg, he summarized the argument this way:

During the early stages of an industry, when the functionality and reliability of a product isn’t yet adequate to meet customer’s needs, a proprietary solution is almost always the right solution — because it allows you to knit all the pieces together in an optimized way.

But once the technology matures and becomes good enough, industry standards emerge. That leads to the standardization of interfaces, which lets companies specialize on pieces of the overall system, and the product becomes modular. At that point, the competitive advantage of the early leader dissipates, and the ability to make money migrates to whoever controls the performance-defining subsystem.

Put another way, when a market is young, it is by nature experimental:

1. Buyers are figuring out whether they should buy, build, or wait

2. Sellers are fragmented and inconsistent, without baked-in economies of scale

In this phase, assuming buyers determine there is value in the new technology, more proprietary approaches look attractive. From the buyer's perspective: given that the current sellers are small and expensive, perhaps we should just control our destiny and grab value ownership in this hot new space. It follows that vertical integration is more interesting for sellers — get consultative, solve the whole problem, own more. In this moment, technology buddies becoming operating businesses even makes sense. Why shouldn’t sellers become what buyers were, owning the fruits of their in-house tech.

As the category matures and standards emerge, however, modular offerings improve. Customers become less willing to pay extra for the integrated solution and begin to value interoperability. At that point, the early integrator is exposed. Low-cost providers assemble products from components, undercutting the integrator’s stack on specialization and price. Open source solutions emerge. And tech businesses who became operating businesses can no longer maintain internal technology autarky — if they want the best tech, they need to become tech buyers themselves. In Christensen’s theory, the long-term market profits are not captured by the integrator, but by whomever owns the “performance-defining subsystem.”

AI makes this theory newly relevant by expanding what a startup can plausibly do in-house. More of the stack can be automated, bundled, or absorbed. That makes integration more tempting, but it does not change where profits ultimately migrate.

In A Guide to Disrupting Incumbents, we shared how this applies to startups disrupting incumbents by commoditizing complements, undercutting pricing power, and using the surplus they create to earn distribution and trust. In that piece, we discussed the canonical example of commoditization with Microsoft and the PC market:

In the early 1980s, software wasn’t even considered a market separate from hardware. When PCs started taking off, IBM was so eager to be a first-mover that it launched its 5150 model with an open architecture: using off-the-shelf hardware, publishing the technical specs, and encouraging expansion cards.³ The PC landscape, including IBM and its fast-followers, would quickly descend into a knife fight, kicking off a race to the bottom on pricing. Meanwhile, Microsoft’s DOS—offering the preferred⁴ OS, a critical complement to PCs—skyrocketed.

Integration can win early and still lose the market later. The graveyards of once-prominent technology stalwarts bear this out. Vertical integration may be structurally necessary or serve as a temporary bridge to market entry before ecosystem maturation. It might also be a company-killing error. Commoditization will eventually affect the integrator’s stack. Vertical AI founders should focus on owning components with lasting defensibility and profit potential.

NVIDIA is the modern counterexample to integration maximalism. The company tightly integrated GPU architecture, CUDA, networking, and the software ecosystem around accelerated compute. But NVIDIA left semiconductor fabrication to companies such as TSMC. It was integrated, with co-design strengthening its control point and enabling access to the capital-intensive layer, where a specialist had better economics. That position helped make NVIDIA one of the most valuable companies in the world.

The Minimum Sufficient Stack

Integration creates a durable advantage only when it secures something scarce. The argument for capturing more of the value chain centers on defensibility. For example: a product that senses the world, captures first-party data, influences an action, observes the result, and improves. If the company owns the entire loop, competitors cannot access the same context.

However, vertical integration does not guarantee defensibility. A company may own the full stack yet still collect only commodity data. Long-term value depends on the quality and structure of the signal, which must be proprietary, dense, and difficult to replicate. A vertically integrated startup may own sensors that produce undifferentiated data. In such cases, integration increases capital and operational burdens without enhancing defensibility or long-term prospects. The company is committed to eventually competiting with startups taking advantage of falling input costs to undercut. Having made the argument for the necessity of integration, management is often forced to wage a losing battle that saps the strength of its core product.

The traditional lesson from vertical software remains relevant. Leading vertical software companies seldom win by capturing the entire value chain — they win by capturing the workflow control point. And the concept of “owning everything” extends as readily across operating lines as it does technological ones. Toast owns the POS and transaction layer, not the restaurant. Veeva owns critical commercial and regulatory workflows in life sciences, not the drugmaker.

The strength of these platforms lies in their ability to understand and monetize the workflow without inheriting the industry’s full cost structure. Vertical AI raises the bar because software can now do more of the work directly. But the strategic question is the same:

Durable AI-native companies succeed by building towards securing control points. In our earlier work on successful Vertical AI playbooks, we recommended that many AI-native companies should begin with an authoring layer — a starting point for a valuable internal workflow that generates revenue and immediate ROI. This was a product strategy argument.

Our view on vertical integration builds on that: at the earliest stage, Vertical startups need to own the minimum sufficient stack to capture the control point: the part of the workflow where that loop compounds.

There are many ways a company can secure a control point, dependent on industry and state of technology: workflow embedding, licensing, integrations, contractual rights, proprietary hardware, and more. What matters is capturing dense and repeatable signals. The rule is simple:

Previously, the key question was whether to build or buy. For Vertical AI, the strategic question is own versus access: which parts require direct control, and where is reliable access sufficient?

The shift from on-premise software to SaaS has already demonstrated this lesson. Most SaaS companies rent compute and storage. A SaaS product that offered no unique value beyond generic infrastructure was vulnerable, but the answer was not to buy a data center. The answer was to capture the layer with data and workflow gravity and ultimately outsource the rest to cloud infrastructure providers.

Vertical integration, in itself, confers no specific market advantage. Sensors, models, integrations, services, and hardware matter when they enable access to the control point. Past that basic threshold, the risk is that ownership adds friction without durability.

Control Points in the Wild

There are three independent variables to describe the components of vertical integration. Access means you can use an asset without owning it. Ownership means the asset belongs to your company. Control means the control point itself.

Four examples show how they combine in practice:

OpenEvidence: access without ownership

OpenEvidence sees the clinical question at the point of care, retrieves the relevant evidence, and presents an answer to the physician.

Commercial agreements with medical journal publishers secured access to licensed content. The control point is the clinical question at the point of care. It does not need to own the journals, the hospital, or the physician workflow end-to-end. It needs trusted access to evidence and repeated presence at the clinical question.

Its defensibility, therefore, depends on the durability of its content access, physician distribution, trust, and the value of aggregated demand, not yet on a fully closed clinical learning loop. In this case, licensed access, without ownership, may be sufficient for long-term durability — because ultimately it isn’t necessarily the content that provides differentiation. It is their superior understanding of physicians’ questions and answers in situ, and the resulting ability to provide the best physician experience.

Halter: ownership for control

Halter captures cattle data, directs movement through virtual fencing, and gives the rancher an application for managing their herd.

The collar is required because it provides access to the control point. Without the hardware, Halter would not have access to the signal, action, or outcome.

However, the hardware itself is not the moat. It is mission-critical today because it secures the observation-and-action loop and the rancher workflow where learning compounds. Over time, modular hardware components should be allowed to commoditize, provided this expands access to first-party data without weakening control point ownership.

Abridge: control without ownership

Abridge captures the clinical conversation, creates documentation, and connects the note to downstream charting, coding, and revenue workflows.

While the sensor is the microphone, Abridge’s control point is the authoring layer: the clinical note starts in their software, flows into the health system through deep integrations, and serves as the base layer for critical downstream work. That position captures clinician edits, approvals, coding changes, and revenue outcomes.

Abridge’s learning loop does not require any additional parts of the healthcare value chain beyond their control point and its downstream workflows. Even as clinical notetaking commoditizes, what compounds is that downstream position, not the transcription itself. In the medical scribe world, the question remains whether the underlying technology supports long-term defensibility of this control point, especially in the face of highly dominant and motivated incumbents — but the control point that matters remains clear.

The AI-Native Law Firm: ownership without control

Now consider a fully integrated AI-Native law firm: client relationships, lawyers, and delivery all in-house.

While owning the entire value chain may appear defensible, in practice, it often results in collecting undifferentiated data that is easily accessible to frontier model companies, traditional law firms using tools like Harvey, and other AI-Native law firms. We discussed this dynamic in our essay on AI-Native services:

Model improvements only make a service defensible to the extent that no competing service adopts the same LLM advances. That’s not a bet we’d take.

The strongest moats in software have always been those that are not easily replicated — proprietary data loops, workflow integration, distribution, network effects — not the infrastructure they’re built on, regardless of whether you charge for the tool or the output.

When Integration Is Actually Necessary

While cautioning against the reflexive urge to own the entire value chain, vertical integration can often be structurally necessary to win a market.

Integration can secure supply, reduce latency, improve quality control, establish a single accountable operator, or prevent a critical supplier from capturing product economics. These are real advantages, but they are mechanisms rather than moats.

Integration is most defensible when one of three conditions holds.

1. The category is not stable enough for modularity

If the handoffs between the model, workflow, hardware, and surrounding environment change constantly, separate vendors may not be able to coordinate fast enough.

In these markets, integration is necessary for rapid iteration. However, this supports integrating across unstable interfaces, not owning every adjacent business.

2. Performance is below the buyer’s threshold

A multi-vendor system may be too immature or brittle, and the buyer may opt for a single party to guarantee reliability. Integration may be necessary to establish trust. RLHF (Reinforcement Learning from Human Feedback) vendors likely fall into this category.

Even in these cases, companies should distinguish between layers required for current deployments and those likely to be defensible profit centers in the future.

3. Durable access cannot be secured without ownership

A supplier may refuse to provide the data. The sensor, literal or figurative, may not exist. A software vendor may make integrations nearly impossible or severely restrict the actions a product can take.

If the only reliable way to secure access is to own the relevant asset, ownership may be required.

These conditions are real but more limited than the integrationist narrative suggests. They also evolve rapidly, especially in the LLM era. Markets mature quickly, fierce competition drives interoperability, and buyers are quickly becoming more attuned to their needs. Modular entrants will inevitably target individual layers of the integrator’s stack, often after the integrator has established the category.

Modular competitors do not always win. But the advantage created by ownership is often time-limited unless the integrator continues to own the performance bottleneck, the control point, or both.

We are in a classic Christensen’s Young Market moment today. Buyers are struggling with understanding what to buy or build — and what is even worth adopting in the first place. To that end, it can make sense to be more experimental, to be more consultative, and to take on more than you otherwise might have to solve enterprise problems. And the reward may indeed be owning more than a traditional wedge, faster.

We believe, however, the prescription for Vertical AI founders is not to blindly commit to owning everything from the get-go. Toeing this line requires strong management discipline and taste, but the ones who do it well — testing everything while retaining focus — will benefit greatly. There are a number of strategies founders can use to maintain balance, manage a wider footprint, and test expansion.

Commoditize Complements, Don’t Adopt Their Economics

In A Guide to Disrupting Incumbents, we wrote about eating complements and nuking pricing power. A startup can weaken incumbents by making adjacent products cheaper, free, embedded, or otherwise less able to capture margin.

But commoditizing a complement does not require ownership, nor does it require adopting its economics or business model. Often, the better strategy is to make the complement abundant, standardized, and cheap so that value shifts back to the layer you control.

Startups can undercut on price or promote interoperability standards. The objective is to remove the bottleneck without permanently inheriting the complement’s cost structure.

If a healthcare AI company identifies services as the bottleneck, it may initially require human intervention. However, the goal should be to convert repeated service work into access to a control point, rather than simply becoming a lower-cost services provider.

The primary advantage of software is leverage. AI changes the boundaries of what startups can do, but it does not negate the attractiveness of software and asset-light business models. If anything, leverage matters more as baseline intelligence gets cheaper. Fast followers, perhaps even the frontier models themselves, will modularize your offerings.

Before founders ask what part of the stack they should own, they should ask a more important question:

If a complement hinders adoption, integrate only enough to reduce its pricing power, coordination costs, or reliability risks. If a layer does not strengthen the control point, consider renting, licensing, or exposing it through a modular interface.

Commoditize complements, but avoid adopting their cost structures. The specialization of function and labor is as close to a fundamental economic truth as we have — great founders pick their battles against this principle wisely, and recognize they are unlikely to win a protracted war.

The Control Point in 1841

After the Panic of 1837 exposed the lack of reliable commercial credit information, the abolitionist merchant Lewis Tappan founded the Mercantile Agency. The agency built the first B2B credit network: a network of paid correspondents across the country, mostly lawyers, bankers, and other local observers, who reported on merchants’ solvency, reputation, and creditworthiness. A young Abraham Lincoln was reportedly among the early network correspondents.

The merchants were the vertical integrators of their era: they bore the capital and operating risk. They held the inventory, warehouses, supply chain, and customer relationships. The Mercantile Agency owned something else: a valuable information layer critical to recurring, high-value business credit decisions.

The correspondent network gave the agency the data. That data was then compiled into useful decision-support systems for credit. Each new piece of information, including payments and delinquencies, improved the system, enabling an “asset-light” merchant to capture the control point. The Mercantile Agency eventually became Dun & Bradstreet, one of the most enduring business-data companies in the United States.

Christensen’s Law of Conservation of Attractive Profits applies to the Mercantile Agency as well as to Vertical AI: profit does not necessarily accrue to the company that owns the most layers. It migrates toward the company that controls the performance bottleneck.

The Mercantile Agency needed thousands of correspondents and decades of patience to capture its control point: the merchant credit ledger. They owned neither the inventory nor the warehouses, but rather the clearinghouse of business trade.

Vertical AI founders should look for the modern equivalent of that ledger. Own enough of the stack to make it accurate, durable, and compounding. Serve a domain so well, displacement becomes not only difficult but uninteresting.

If Tappan or Christensen could give you one piece of advice today, it might this: bundle and unbundle with care.

Thanks for reading The Verticalist. Euclid partners with Vertical AI founders at inception. If someone in your network is building in the space, we would love to meet them. LinkedIn is the best place to reach us.

Zach Rosen is the Co-Founder and CEO of Brellium, the AI compliance platform that audits every patient visit across ambulatory healthcare. Brellium serves over 250K providers across all 50 states, auditing more than 1M clinical records per month against payer, regulatory, and OIG standards.

The company has raised >$30M from First Round Capital, Left Lane, Kearney Jackson, and Menlo Ventures. A second-time founder, Zach struggled with a personal misdiagnosis that drove him to build a safety net for patient care.

Watch the full episode here

Tune in to hear why: the real compliance moat isn’t the public rulebook; how Brellium is building intelligence no EHR or model lab can replicate; and why compliance is a dead wedge in some AI markets… and a killer in others.

Today’s Episode

Compliance is having an AI moment. Norm AI just raised $120M at a $1.2B valuation led by Khosla Ventures, turning regulations into AI agents in legal and finance. But it isn’t the first time compliance has been in vogue. Veeva built a $40B+ business on compliance in life sciences, Avalara did it in tax, and Workiva in financial reporting. Boring on the surface, billions underneath.

But healthcare compliance is a different beast. The rules aren’t deterministic, the enforcement is inconsistent, and the stakes are orders of magnitude more personal. Zach Rosen has built Brellium to tackle a problem space where all three collide. This episode taps his experience to get at questions that matter for any founder eyeing compliance as a wedge: when does it become a moat, when is it just a feature, and how has AI changed the game?

Non-Determinism is Your Best Friend

The conventional wisdom on compliance as a Vertical AI wedge is straightforward: regulations are public, the data isn’t proprietary, the moat is thin. If you can Google the requirements, so can your competitor and so can the incumbent EHR.

Zach argues this framing fundamentally misunderstands what compliance means in ambulatory healthcare. The public bulletin is the starting point, not the finish line.

Take Aetna’s ABA requirements in Texas. You can find the 12 items on the payer bulletin: sessions must be individualized, care must be medically necessary, and so on. Simple enough on paper. But what does “individualized” actually mean? What does “medically necessary” look like against a specific patient chart? Two providers reviewing the same documentation will come up with different answers. And Aetna’s auditor may have a third definition entirely.

There are kind of fractals of knowledge. The closer you zoom in... it’s not as cut and dry. If it’s 45 words of medically necessary care that’s also individualized to patient, then all of a sudden you start to understand — wait, there are a bunch of edge cases here that I maybe didn’t account for.

This gap between exposure and enforcement is one we’ve explored in the past. And it’s the core of Brellium’s defensibility strategy. The regulations may be public. The enforcement and remediation patterns — written, tribal, and word-of-mouth — are not.

About a year ago, one of Brellium’s customers was placed in prepaid review and had funds clawed back because a payer in Georgia required backdated co-signatures from two separate providers on all documentation. That requirement wasn’t on any public bulletin until weeks after the enforcement action. The provider lost a substantial sum essentially out of the blue.

In the AI era, doing the work is as important as identifying it. That’s why it’s critical that Brellium doesn’t stop there. They intake that enforcement event, codify it, and push it to every other customer billing the same payer in the same state — the next provider doesn’t get blindsided. Multiply that across 250,000 providers, all 50 states, and dozens of payers, and you get a compounding intelligence layer that no system of record can replicate because EHRs don’t see how audits actually play out. These are decision traces only a platform that sees and does the work can capture.

Parafin is the partner that makes Verticals possible. They just launched the Spend Card, a business credit card that platforms can deploy to their SMBs without becoming a bank. Fresh off a Goldman Sachs credit facility and the 2026 Forbes Fintech 50, Parafin has extended over $35B in offers to small businesses across platforms like DoorDash, Gusto, and TikTok Shop.

Explore a custom Parafin program

Compliance Products Are Secretly Training Products

Brellium sells compliance. That’s the “why to buy.” But the actual first-order value delivered is workforce training.

A real-time compliance flag can tell a provider to swap out one diagnosis code for another. That protects the documentation. It doesn’t change how the provider treats the next patient.

At the surface, we’re a compliance product and that is why most of our customers choose to partner with us. But actually if you dive deeper, the value that we provide — that’s actually first order, not the second order effect — is that we’re a training product.

Brellium’s weekly reviews track provider performance over time: what they did well, what they need to improve, why, and how. Providers who improve get placed into higher performance buckets that qualify for bonuses and promotion tracks. Providers who repeat mistakes get reinforced constructive feedback. It’s a closed-loop system that changes behavior, not just documentation.

This is a pattern we’re seeing across Vertical AI more broadly. The products building lasting defensibility aren’t the ones automating a task; they’re the ones making the human better at their job over time. Rilla does this for field sales in the trades. Brellium does it for clinical compliance. The product starts as a tool and becomes, over time, indistinguishable from the org’s training infrastructure. That creates switching costs no feature comparison can overcome.

Two Paths to Win in Vertical AI

Zach views the basic option set for nascent Vertical AI products as being an evolution of the Vertical SaaS playbook:

• Path A — Start with a wedge product, get customers to pay for it, then bolt on adjacent products to become the system of record / action. This was the HubSpot playbook, the Salesforce playbook, the playbook 90% of vertical software companies ran. Start with one thing, make it excellent, then add a dozen more over five years.

• Path B — Start with a wedge so good the system of record / action can’t possibly compete with you, because the wedge is powered by data that is proprietary (or that the incumbent can’t feasibly access). Ideally, build a network effect around that wedge. Gong ran this play for six years as the “best call intelligence tool,” plugged into HubSpot and Salesforce as a pure point solution, before pivoting into a platform.

The AI era has made these two paths diverge further. Path A is an order of magnitude easier now; AI makes building adjacent products and expanding a platform faster than ever. But Path B, building a genuine network effect, is marginally harder, because the incentive structure shifts away from perfecting the individual customer journey and toward aggregating data across many.

If you are obsessed with compliance in a highly regulated industry... you have two core options. If compliance is simple and deterministic, you should probably take Path A — use it as your wedge and then build adjacent products. Path two, which is where we’ve leaned, is when compliance is non-deterministic. And that network effect is actually passed through taking word of mouth information and or information that is not publicly available.

AI-native Path A plays might be starting with a different kind of platform wedge, or even a service. For Path B Vertical AI businesses, the type of data that is defensible may be changing in the AI era, but the core strategy mostly rhymes.

Brellium has bet on Path B. The enforcement intelligence compounds. A provider switching to a competitor loses access to enforcement data accumulated across Brellium’s entire network. The two-sided nature of their model — solving a costly and dilatory issue for both payors and providers, translating between them without need for communication — create an additional network effect.

The Provider-Payer Flywheel

Brellium’s go-to-market exploits one of healthcare’s unusual structural features: the provider and payer sit on opposite sides of the same compliance problem.

Providers want to avoid funds being clawed back. Payers want to enforce compliance because every dollar out of their bank account goes into a provider’s bank account. Both sides spend heavily arbitrating what payers call fraud, waste, and abuse, essentially non-compliant documentation and care.

Brellium started with providers: mid-market, lower burden of proof, faster to stack proof points. Once enough providers in a given payer-state combination were on the platform, the data told a clear story. Brellium-approved providers were measurably more compliant than the market average.

That became the pitch to payers. Once a payer endorses Brellium, it makes the next provider conversation more straightforward. Not easier, Zach is quick to clarify (“none of this is easy — you’re selling next gen products to a highly regulated market that is generally not incredibly excited about technology”), but more straightforward.

The flywheel accelerates because each new provider contributes enforcement data, each payer relationship validates the product, and both sides compound the network effect. It’s a version of the two-sided vertical platform dynamic, grounded in compliance intelligence rather than transactions.

Healthcare’s notoriously closed EHR ecosystems actually help here. Unlike most Vertical SaaS categories where the rational incentive is to integrate and play nice, healthcare EMRs are, in Zach’s words, “almost entirely closed ecosystems.” Brellium’s GTM playbook: partner with the #2, #3, and long-tail EHRs first (who are incentivized to fill gaps), build the customer base on the dominant platforms organically, then formalize distribution partnerships up the chain once the proof points are undeniable.

The Takeaway for Vertical Founders

Compliance is one of the oldest wedges in vertical software. Veeva, Avalara, Workiva: the playbook is proven. But in the AI era, the game is bifurcating:

In verticals where compliance is deterministic, with clear rules, consistent enforcement, and public standards, AI makes the wedge easier to build and easier to copy. Use it to get in the door, build fast, and expand into the platform. Path A, in Zach’s framework.

In verticals where compliance is non-deterministic, with subjective standards, private enforcement, and word-of-mouth rule interpretation, AI enables something different entirely. The product becomes an intelligence layer that aggregates how rules are actually enforced across an entire market. This Path B offers a shot at building something genuinely defensible. In Zach’s eyes, it also requires more: a stronger leadership team, a more dynamic product org, and a tolerance for the slower burn of network-effect compounding.

The compliance products that win the AI era won’t be the ones that can read the rulebook. Translating deterministic rulesets and workflows from analog to digital was a SaaS-era play that’s now little more than a foot in the door (if you’re lucky enough to have found an urgent, blue-ocean use case). The AI-native winners will be startups that solve for the hairiest compliance issues — which they can do only because they’ve seen when and why they’re broken, in real time, over and over again.

See you next week.

Adam Harris is the Co-Founder and CEO of Cloudbeds, the hospitality management platform operating across 157 markets with $20B+ in annual transactions and 550,000 daily logins. Harris built Cloudbeds over nearly 14 years from a fragmented stack of nine disparate systems into a unified system of record for hoteliers worldwide, and now, a system of action. A UC Berkeley-trained economist who started on Wall Street before crossing to the building side, he recently launched Signals, a conversational AI layer built on Cloudbeds’ proprietary hospitality data model — more than a copilot, it’s built on 14 years of hospitality pricing data and is evolving into a true system of action. Tune in to hear why he thinks the AI gold rush is producing more pirates than pioneers, what it’s taken for a an incumbent to go all-in on Vertical AI, and why 80% of mature market leaders will be too slow to survive.

Watch Full Episode Here

Today’s Episode

Every podcast and pitch deck in 2026 tells the same story: AI startups will disrupt slow-moving incumbents. The system of record is dead. The future belongs to whoever ships the fastest agent.

Adam Harris has heard it all. He runs a platform that will check in 57M guests in the next six months, processes $20B+ in transactions, and sits on 4B data points across 157 markets. Last month, Cloudbeds displaced 117 competing systems — Harris had never heard of most of them. His take on the disruption narrative is blunt: the treasure is the dataset, and right now, everyone else is just plundering for it.

The Pirate Problem

Adam doesn’t mince words about what he’s seeing in hospitality AI. “I call it pirate season,” he says. “Everyone is looking for the gold and they’re just plundering wherever they can take it.”

What makes someone a pirate is the shortcut: taking the treasure without doing the work to earn it. That takes two forms:

1. The visible one is the credential pirate, the operator who entered the industry six months ago and is now a conference regular and, in Harris’s words, one of the “big voices on LinkedIn.”

2. The substantive one is the product pirate — the startup shipping a Claude-powered revenue tool or a chatbot that hallucinates cancellation policies. Wrap a frontier model and call it a moat.

Unfortunately, for would-be pirates, access to true vertical data gold isn’t easy. Anyone can spend $20/month on an Anthropic subscription and spin up quick solves. What they can’t access is 14 years of booking patterns, rate decisions, guest preferences, and operational data across 20,000+ properties — not, at least, without winning access through customers, partnering with platforms that have been around the block, or creating a flywheel of their own. As we’ve argued before, the moat in Vertical SaaS has always been the compounding dataset. Incumbents are getting increasingly sharp-elbowed around their hard-won assets.

“If ChatGPT wanted to optimize for hotel transactions, they should come talk to me,” Harris says, “because I do it every single day.” When we asked him if he’s giving full data access to any AI-native startups, Adam responded with a resounding “hell no.”

A quick word from our sponsor on the Verticals podcast, Parafin. Fresh off the 2026 Forbes Fintech 50 and a new credit facility led by Goldman Sachs, they’ve now extended over $35B in offers to small businesses through platforms like DoorDash, Gusto, and TikTok Shop. Purpose-built embedded capital for vertical platforms. Explore a custom program for yours today.

From Record-Keeper to Action-Taker

SaaS is metamorphosing. Harris frames it as a regime change: the value of a system of record is no longer the record itself, but what you do with it. “Salesforce kind of feels like a big dumb database to me,” he says. “I don’t want to log into Salesforce. I just want the data and the action point.”

This tracks with what we found in our analysis of the SaaS repricing: the public markets didn’t punish all software equally. They punished software that was a passive store of data with no path to becoming a driver of outcomes. The winners had consumption-based models, AI-augmented workflows, or both.

Cloudbeds’ answer is Signals, their proprietary AI layer. The stat that matters: 80% of hoteliers see the right pricing move but act too late. Not because they lack awareness, but because they’re running a 24/7/365 operation where humans physically cannot process every signal at the speed the market demands. Signals runs millions of permutations per hour across Cloudbeds’ 4B data points and furnishes the action ahead of time.

The distinction Harris draws is between probabilistic and deterministic outputs. “I cannot have a probabilistic output for whether a rate is $100 or not. It has to be $100 if it’s $100.” A thank-you note can tolerate hallucination. A pricing decision cannot. That’s why Cloudbeds built deterministic harnesses around their AI (confidence scores, source attribution, audit trails) before connecting any of it to customer-facing workflows.

The Two-Year Blueprint

The most counterintuitive thing Harris said on the show: Cloudbeds spent two years rebuilding its data architecture before shipping any AI products. In the time it’s taking some startups to go from zero to tens-of-millions of ARR, Adam has been methodically laying the groundwork for his transition to an AI system of action.

He compared the process to building a house. “Permitting and planning takes most of the time. The actual building of the foundation and walls—it comes along pretty quick.” Most AI startups skip the permitting. They start with the construction and then go back to fix what they should have planned from the start.

Harris places the agentic AI cycle at “phase one of five,” roughly where self-driving cars were before anyone trusted a Waymo. “They’re super cool,” he says of agents. “I love swarming them up. But there’s a lot of garbage out at times.” He’s bullish on the long arc but thinks the industry is confusing current model capability with production readiness.

Interestingly, the jury is still out for him on whether an AI-first Cloudbeds will command more or less revenue. He ran parallel experiments: traditional SaaS pricing versus performance-based pricing where they take a cut of the revenue lift they generate. The outcome: both models hit approximately the same number. The value was equivalent regardless of how they charged for it. That’s a signal the product is genuinely driving outcomes, not just billing for the promise of them.

The Takeaway for Vertical Founders

If you’re building against an entrenched incumbent with a real data corpus and an active AI strategy, your window is narrower than you think. Harris displaced 117 competitors last month. He’d never even heard of most of them. The graveyard of “AI-powered” point solutions is getting crowded fast.

But not every incumbent is Cloudbeds. Harris himself pointed at the PE-owned, margin-optimized platforms that are “running the same exact business they were pre-ChatGPT.” 80% of PE money is sitting on the sidelines right now. They maximized profit but neglected the data layer. Those are the ones where the data asset exists but nobody’s investing in it. That’s where the real opportunity for founders lives. Not in outrunning incumbents who are already moving, but in replacing the ones who haven’t picked their heads out of the sand.

And for incumbents? Harris’s advice is unglamorous but firm: get your data right first. Build the harnesses. Don’t ship AI slop just to say you have AI. The companies that resist the urge to protect their hoard and instead aggressively invest in the vault, won’t just survive in the era of Vertical AI — they’ll thrive. Incentives as they are, however, neither we nor Adam expect that share to be more than ~20% of legacy market leaders. See you next week.

Zain Jaffer is the founder and CEO of Blazel, an AI-powered marketing platform that helps companies build authentic presence on LinkedIn and generate leads. Before Blazel, Zain built Vungle into the fastest-growing company in mobile ads, scaling it from zero to >$400M in ARR before Blackstone acquired it for $780M — after which it went public at a $4B valuation. After a few years on the VC side, he jumped back into the founder seat with a $10M pre-seed for Blazel. Tune in to hear why he’s betting that Vertical AI companies can (and should) claw back control from the model labs by building or adapting their own capabilities. He makes the case that the economics of renting intelligence may not hold up as long as founders assume.

Watch the Full Episode Here

Today’s Episode

Every Vertical AI founder faces a version of the same question: how much of your intelligence stack should you own? The default answer — build on Anthropic or OpenAI, move fast, ship product — has been the right one for most early-stage companies. The models are excellent, API credits are abundant, early gross margins can be forgiven, and nobody needs to reinvent inference infrastructure at Day Zero.

But a growing number of founders are hitting a wall. Costs escalate unpredictably. Latency from multi-agent architectures compounds. The frontier labs continue their saber-rattling about moving into the application layer as every API call you make teaches them more about your workflow. Meanwhile, open source seem to catch up faster every cycle. Zain lays out why he decided to build Blazel’s own model capabilities, pivot away from an AI-Native Services model, and what his decisions reveal about where value will actually accrue in Vertical AI.

The $10M Trap

Zain’s journey toward model independence started with a pivot. Blazel began as an AI native agency — one of the first, before Sequoia, a16z, and YC started championing the category. The agency grew fast. Replacing existing marketing budgets was a frictionless sale: buyers already have a purchasing workflow, there’s no new vendor approval process to navigate, and you don’t even need a UI. Zain’s take on agency-style AINS: “It’s very easy to get to $10M ARR. I’m not kidding.”

The problem is what happens at $10M. You’ve hired a team of humans. Any AI you introduce threatens their jobs, and they become a barrier to the very automation that should be your competitive advantage. More agencies flood in because the barrier to entry is so low. The allure of hiring more humans before the LLM cost structures warrant them — to maintain growth and raise more capital — is hard to deny. And if you go to your VCs and propose cutting revenue to pivot toward a platform model, you face a high likelihood of resistance, even disillusionment.

We’ve discussed this dynamic in the SaaS-to-AI transition: incumbents are generally terrified to cannibalize their own revenue, and it gets harder with scale. Zain saw it forming and made the call to pivot before the agency got too big. “It’s very hard to pivot when you’re 10M ARR.” Zain also emphasized the difficulty of running a people and technology business simultaneously — at Vungle, he ultimately felt their decision to split focus between brand advertising and app installs was a major misstep. Competitor AppLovin kept singular focus on one ICP and one problem. Vungle exited at $4B. AppLovin is worth $150–250B.

The agency, in Zain’s telling, wasn’t the business — it was the product roadmap. Humans doing the work revealed exactly which workflows to automate and where domain-specific AI could outperform general-purpose models.

A quick word from our partner and sponsor on the Verticals show, Parafin. They power embedded lending across Amazon, DoorDash, Walmart, and others — with >$35B in financing extended to date. Purpose-built for vertical SaaS & AI platforms serving SMBs.

Explore a custom program for your platform today →

When the API Bill Hits

Blazel’s content pipeline initially ran on Anthropic’s Sonnet. Multiple agents would generate content, other agents would edit out AI slop, then the output fed back as labeling data to train evals. It worked. Until the bills came.

“How did I spend $70K last month on Claude?” Zain recalls. “I wasn’t monitoring my tokens because it’s free credits.” The cloud credits that seed-stage companies get from AWS, GCP, or Azure mask the real unit economics. At API pricing — roughly 20x what you’d pay through consumer products — the math gets ugly fast for any company processing high volumes of domain-specific content.

Cost was one forcing function. Latency was another. Every agent call is a round-trip to the model provider. Chain four or five agents together — generate, edit, check for slop, evaluate, reformat — and the latency compounds. For a product that needs to feel responsive to time-strapped CMOs, that kills the experience.

But the existential forcing function is competitive. Frontier labs are building applications — or at least threatening to. Just by existing on their rails, you’re helping them accelerate your obsolescence.

The model providers themselves aren’t content sitting in the model layer. They want to go into the application layer… You are basically training the model provider. — Zain

Fine-Tuning Your Way Out

So Blazel did something that sounds, on its face, absurd: a pre-seed startup started building its own model capabilities.

The approach wasn’t training de novo models from scratch. The team fine-tuned open-source models using reinforcement learning and pre-training techniques, then built a harness of prompts and tools around the result. The labeling data came directly from the agency’s human editors: months of expert corrections that encoded what “good” content actually looks like in specific domains. This is why the agency phase mattered — it generated the proprietary training data that no frontier lab could easily replicate.

We got to the point where our own AI model was producing better content than what our team could produce. And we hired the best content people you can imagine. — Zain

A domain-specific model, trained on domain-specific evals by domain experts, beat a frontier model at a specific task — not because it’s a better model overall, but because it’s been shaped by data the frontier labs don’t have. Your content editors’ judgment, your customers’ voice patterns, your vertical’s quality standards. These artifacts can’t be scraped from the open internet.

We made this argument in Dude, Where’s My Moat? — workflow and data are the immutable primitives of software defensibility. That principle applies equally to the model layer. The companies with the deepest workflow integrations, that win them access to the most relevant proprietary data, will build the most defensible model capabilities. They do it not by outspending the labs or going toe-to-toe with them on compute, but by outlearning them on narrow, high-value tasks.

To work toward model independence, evals are key. “Labeling is a fundamental layer when it comes to AI and that’s where humans are absolutely important. For the LLM to be a good judge, it needs to have some intuition, and intuition comes from hiring humans.” They’re the mechanism through which domain expertise gets translated into both moats, whether at the application or AI infrastructure layers.

A Trillion Parameters Under Your Desk

While Zain’s bet seems quite contrarian at the moment, it’s increasingly plausible as we consider where compute is heading. Nvidia’s DGX Station, announced at GTC 2026, puts 20 petaFLOPS of AI performance and 748GB of coherent memory in a deskside form factor — enough to run a trillion-parameter model locally.

If running large models locally becomes cheap enough, the economic argument for API intelligence dependency weakens substantially. Companies can fine-tune and serve open-source models on their own infrastructure, keeping proprietary data in-house and eliminating per-token costs at the margin. The labs’ moat narrows to whatever capabilities can’t be replicated with open-source architectures, custom training data, and increasingly affordable compute.

Zain frames it as an existential risk for the frontier labs: “A model itself could actually become a commodity. I think that’s a threat to OpenAI and to Anthropic, which is why they’re trying to go” app-side. If the model layer commoditizes, value accrues to whomever owns the workflow and the data — and the remainders will be either have to make their bones on supply ownership, become Dispatchers to the cloud provider oligopoly, or get competed to death. The irony is that the frontier labs, continuously making noise about their path into the application layer, seem to agree. Even SpaceX — a business that has basically no B2B apps — claimed in their S-1 that the enterprise application layer represents >90% of their TAM.

The Takeaway for Vertical Founders

None of this means you should rush to fine-tune an open-source model at pre-seed or seed stage. The API-first playbook is still the fastest way to product-market fit. Use the credits. Ship the product. Learn what your customers actually need.

But be intentional about what you’re learning. Every human edit, every customer correction, every eval your team runs — that’s labeling data. It’s the raw material for the domain-specific model you may need to build sooner than you planned. Zain’s team didn’t set out to build their own model. They set out to build a marketing agency, and the data they generated while doing it became the foundation for something far more defensible.

Founders who want to best position themselves should consider treating API dependency as a strategic phase, not a permanent architecture — at least to maintain optionality and set the right product culture. Build on the frontier labs today, but capture the data and develop the evals that would let you peel off high-volume, domain-specific workloads when / if the economics demand it. The compute to run your own models is getting cheaper by the quarter. It’s too soon to imagine every startup with it’s own LLM. But the question is not whether Vertical AI companies will own more of their model layer — it’s when, and in which form factor.

Zain has lived the cycle. He built a company to a $780M exit by disrupting the mobile ad incumbents. He watched from the VC seat as SaaS companies failed to adapt to platform shifts. Now he’s back, betting that a startup with domain-specific data, proprietary evals, and fine-tuned models can build something the trillion-dollar labs can’t match in his vertical. The labs have scale. Vertical founders have specificity, earned insight, and (in the best cases) data frontier model providers can’t scrape or buy. In a world where compute is commoditizing fast, the model landscape of the future — at least in Vertical AI — might not be as homogenous as recent history has trained us to believe.

See you next week.

Adeel Khan is the Founder and CEO of MagicSchool, the leading generative AI platform for K-12 education. It grew from zero to 7m users — and >$10M in ARR — in less than a year. Adeel is a true domain-expert founder: he spent his career as a teacher, principal, and school founder. After launching MagicSchool’s MVP in May 2023, the company hit 1M users in just five months — all organic — and today partners with districts representing one in five children in America. And Adeel did it with a sales team that is 85-90% former teachers. MagicSchool has raised $65M across seed, Series A (led by Bain Capital Ventures), and a $45M Series B led by Valor Equity Partners. Tune in to hear how Adeel turned a career in classrooms into the fastest-growing Vertical AI in education ever — and why his hyper-vertical GTM strategy allowed him to turn an AI wrapper into a category-defining platform.

Watch the Full Episode Here

Today’s Episode

In VC circles, K-12 education has a reputation as a startup graveyard. ACVs are small. Buying cycles are fickle and cyclical, following the school calendar. Decision-makers are risk-averse by necessity: incentives for innovation are limited and the downside of a bad technology bet is a room full of nine-year-olds staring at a broken screen. The sector has produced vanishingly few large software outcomes in the modern era, and the even EdTech VCs are often skeptical of the category.

MagicSchool AI — after launching a ChatGPT wrapper in 2023 in K-12 — went from zero to tens of millions of ARR, 7M teachers, and district partnerships covering one in five American children, all in less than three years.

The obvious question is how. The answer, in large part boils down to two things. The first is some combination of timing and luck, which frankly, all hyper-growth startups need. The less obvious — and more instructive — answer is that the founder leaned heavily into his vertical domain expertise. MagicSchool was built by teachers; and from inception to today, it’s sold by teachers. 85-90% of Adeel’s GTM team have backgrounds as educators — teachers sell it, implement it, and show up when a superintendent needs to understand what AI means for her tens of thousands of students. That hyper-vertical GTM was critical to MagicSchools breakout success.

Leaning into Your Domain Expertise

We’ve written extensively about how founder backgrounds correlate with Vertical SaaS success. The data is unambiguous: domain expert founders capture 84% of exit dollars and raise more, from higher-signal investors. Adeel Khan is a case in point — though he’d be the first to tell you it was far from obvious in early days.

When MagicSchool was raising its seed in 2023, the deck was stacked against it in every conceivable direction. Application-layer AI companies were, in Adeel’s words, “at their absolute bottom.” EdTech wasn’t anyone’s darling. And then there was the founder himself: a former school principal with no technical co-founder, no software background, no Stanford pedigree. As we explored in The Fundability Trap, VCs pattern-match on consensus profiles — and Adeel matched none of them.

Even the ed tech VCs... they were just like, we don’t have the pattern of a teacher founder or a principal founder. Whether or not they said that out loud, it was kind of a thing that I saw. — Adeel Khan, MagicSchool

Ironically, the very thing that made investors nervous — a non-technical founder who’d spent his career in schools — is what made the product work. Adeel’s first customers in Colorado weren’t buying software. They were buying him. His reputation, his understanding of their problems, his ability to walk into a CIO’s office and speak the language of someone who’d been burned by the last five EdTech vendors. Adeel isn’t the post-pedigree founder — he’s the strong-pedigree founders, just one without the typical pedigree VCs favor.

Parafin is the partner and sponsor that makes the Verticals podcast possible. They are a breakout in their own right. With credit backed by Goldman Sachs, Parafin extends embedded lending across partners like Amazon, DoorDash, and Walmart. If you’re a vertical platform looking to offer capital products to your SMBs (without becoming a bank), Parafin has over $25B in experience. They also recently made the Forbes Fintech 50 list! See if a custom Parafin program could make sense for you.

Vertical GTM: Hire the Expert, Not the Seller

This is where MagicSchool diverges from the standard “domain expert founds company” narrative. Most vertical founders leverage their expertise at the top — in product vision, in early sales calls, in pitch decks. Adeel built his entire organization around domain fluency.

Their solutions architects — the closest analog to a forward-deployed engineer — are former district technology implementers, pulled directly from the schools they now serve. When a solutions architect walks into a superintendent’s office, they’re not demoing features. They’re describing problems they’ve personally lived. While acknowledging the trade-off, Adeel framed the logic this way: a former teacher can learn how MagicSchool’s prompt architecture works; a career SaaS seller can’t learn 15 years of navigating school board politics in an onboarding sprint.

The companies that invest in domain-fluent GTM teams compound trust faster than those that bolt on traditional sales orgs. As we’ve explored across our founder-led sales analysis, that difference shows up in pipeline velocity, win rates, and retention, especially early. It’s not the solution for every market. In some — especially those with faster-cycle, more trasnactional sales — pure sales athletes can work best. That’s what worked at Owner, a platform in the restaurant space, as CRO Kyle Norton shared in a prior episode.

In K-12, the buying decision is as much psychological and emotional as logical. MagicSchool’s competitors — and there are dozens, from legacy SIS providers to curriculum companies to horizontal AI copycats — have launched products that “even look like ours,” as Adeel puts it. Features can be copied. An organization that speaks the buyer’s language and knows how to grow into AI with them cannot.

Monetization: Just Get Them to Sign

MagicSchool’s monetization story is a case study in creative pricing for risk-averse markets. The product was entirely free until January 2024 — eight months after launch, with a million users already on the platform. Adeel’s reasoning was disarmingly simple: “The biggest companies in the world that I know were free.”

When monetization came, the team didn’t hire enterprise sellers or build a traditional pipeline. They invented a pilot structure designed to remove every possible objection: $1,000 for a single school, $2,000 for an entire district — regardless of size. Miami-Dade, with 350K students? $2,000.

The catch was in the contract. Each pilot agreement included the full renewal price, timed to the district’s natural buying cycle in July. The spring semester pilot was a zero-risk trial with a committed upsell baked in. “The goal was sign a contract with us. If you sign a contract with us, we’ll figure the rest out later.”

The open question — for MagicSchool and for Bain Capital Ventures, which led the Series A before renewals came due — was whether districts would convert at full price. Adeel’s COO predicted 50%. Adeel predicted nearly 100%. BCV landed somewhere in between.

Adeel was right. Near-universal renewal at full ticket price. The pricing strategy worked because of three things that don’t generalize easily: a million organic users already in the buildings providing bottom-up pressure, a product teachers genuinely loved (94% were enthusiastic fans, per internal NPS), and a team that could walk into a renewal meeting and speak the buyer’s language from lived experience. Strip out any one of those, and instead of a hyper-growth Vertical AI story, MagicSchool might’ve been proof-positive of the poor retention rates of most AI wrappers.

The GPT Wrapper That Won

In 2023, “GPT wrapper” was the most derogatory label in venture capital. Adeel concedes the point freely: “In the most derogatory way, we were certainly a GPT wrapper. And so was everyone else.”

A prevailing thesis at the time was that horizontal platforms — ChatGPT, Claude, Gemini — would eventually subsume every application-layer product. Users would learn to prompt; platforms would make it easier to customize the UI, harness, and skills. Killing the wrappers in the process. Even Adeel, at the time, thought it was a real possibility. And as it turns out, it wasn’t altogether wrong — plenty did die.

Three years later, however, the opposite has played out for MagicSchool. Getting real value from horizontal AI tools now requires connectors, projects, context ecosystems, custom skills, etc. And while some users can string these together in their free time, teachers have one or two hours a day at their computers and none of the expertise to execute. They’re not building context architectures for fun. And as MagicSchool grew, it found ways to become less vibe-codable.

The structural insight — one we’ve explored in Software Is Dead — Long Live Software — is that as the underlying models get more powerful, the complexity of extracting domain-specific value increases, not decreases. A horizontal platform doesn’t know yesterday’s quiz scores, the district’s curriculum standards, which students are on IEPs, or how their teacher-users are likely to think or decision. MagicSchool does (or at least gets closest). Adeel is adamant, the moat isn’t the model. It’s the context layer on top of it — and the domain experts who built it know exactly what context matters.

Adeel shared a great example: Jasper — front of mind for every AI founder in 2023 — never materialized in education. Why, given it’s clear overlap around the content creation value props MagicSchool et all centered around? Jasper’s users were marketers and writers who lived on computers all day, cycling between tools as better ones emerged. Teachers aren’t tool-hoppers. They adopt what works and stay put. The “GPT wrapper” wasn’t a liability. It was the fastest path to becoming the trusted, contextually aware platform that no horizontal product could replicate. And like in most great Vertical AI businesses, there were vertically specific dynamics key to this story that wouldn’t held up on other markets. Adeel saw them early.

The Takeaway for Vertical Founders

MagicSchool’s trajectory compresses about a decade of normal vertical software company-building into three years. Explosive PLG, creative monetization, near-perfect conversion, the transition from order-taking to strategic sales, a highly vertically trained GTM team, and the emergence of legacy competitors who copy everything except the one thing that matters — teacher’s love for the product.

The lesson isn’t “hire domain experts” — that’s obvious enough to be useless, and as mentioned, won’t extrapolate to every vertical. The lesson is that, when compounded across an organization, domain expertise is infinitely more valuable. One domain expert founder gets you early customers and a good product. An entire GTM team that’s lived the buyer’s daily reality gets you a trust advantage that copycats can’t reverse-engineer. The competitor can match your feature set. They can’t match the fact that your solutions architect used to run IT deployments in the buyer’s own district. And in MagicSchool’s case, it’d be very difficult to catch up on district lock-in and brand. First-mover advantage is provably unreliable — but it had real value here when paired with excellent continuing execution.

Adeel put it well when describing how schools choose technology partners: “So much of the work that we’re doing with Magic School isn’t just about the app or the product. It’s also about the people and the partnership.” This episode is a great reminder that in Vertical AI, the org chart can be a true, tangible moat.

See you next week.

We believe that judgment will prove to be the real prize in Vertical AI, separating great products from great companies. Agents can already follow rules, pull the right data, and generate useful documents. What they can’t do is handle exceptions: the case that doesn’t fit the template, where a human has to intervene based on market, customer, and counterparty-specific context. Those calls don’t live in an ERP, CRM, WMS, etc. They pile up in the exception queue, and they vanish as fast as humans clear them. Longer term, enterprise value in Vertical AI accrues to whoever captures the reasoning behind that human judgment.

We’re already seeing this pattern in the early vertical AI breakouts. Tennr owns referrals. EvenUp owns demand packages. Abridge owns clinical documentation. On the surface, the products look different. Underneath, they all do some version of the same thing: they sit inside a repeated, high-consequence workflow that leaves behind a reusable decision trace with every resolution. This is the exception queue, and we believe it’s the most valuable asset for each Vertical AI startup to capture.

Why Wedges Work

In evaluating wedges, we often think of a two-part test: on one axis, exception density; on the other, the cost of failure (or delay). The former, as you might expect, considers the frequency of workflow exceptions requiring remediation. The latter considers what happens when exceptions aren’t caught swiftly or handled properly.

A workflow with high exception density but low consequences may drive usage — and even near-term revenue growth — but usage is not a proxy for urgency or durability. It becomes a productivity tool, and productivity tools are fragile in a world where the frontier of intelligence is rapidly commoditized. A workflow with high consequence but low frequency can matter, but it often behaves more like a consulting project than software.

When the cost of failure or delay is high, the buyer has urgency. “Cost” can show up as unrealized revenue, idle labor, margin leakage, compliance exposure, or cash flow getting stuck. That is why delay matters as much as failure. A prior authorization that technically clears but arrives too late has still failed the workflow. A recoverable denial that is not appealed within the window still results in leaked revenue. A freight dispute that closed after the invoice ages out can still damage cash conversion.

When the cost of failure or delay is high, the buyer has urgency. “Cost” can show up as unrealized revenue, idle labor, margin leakage, compliance exposure, or cash flow getting stuck. That is why delay matters as much as failure. A prior authorization that technically clears but arrives too late has still failed the workflow. A recoverable denial that is not appealed within the window still results in leaked revenue. A freight dispute that closed after the invoice ages out can still damage cash conversion.

What the test points you toward is the exception queue: a steady stream of non-standard cases that humans already triage, escalate, document, resolve, and audit. Take prior authorization. In Medicare Advantage alone, a recent KFF study found nearly 53 million prior authorization requests submitted in 2024, and 4.1 million were denied. Only 11.5% of those denials were appealed, and 80.7% of the appeals won. The constraint? Physician capacity. A recent AMA survey shows that physicians complete an average of 39 prior authorizations per week, which consumes roughly 13 hours. The system is designed to drain provider resources and capacity. Of the denials anyone bothered to fight, more than four in five were reversed.

Logistics has the same patterns. A shipment can technically arrive after a missed appointment or a disputed accessorial charge, but the delay still burns money. For 3PLs, for example, up to 35% of invoices are disputed. Errors, missed charges, and improper billing not only hurt profitability and increase average DSOs, but they also lead to customer churn.

Another useful indicator: glue functions. Care coordination, claims processing, freight audit, permit coordination, and procurement exceptions often live across spreadsheets, email threads, phone calls, portals, and people’s heads. The context is fragmented across multiple systems and people, and the resolution depends on judgment.

That’s what a great wedge looks like: judgment calls someone is already paid to clear, high cost to the constraint or delays, and a recurring queue.

Where the Best Enter the Queue

The best vertical AI companies don’t all look like “exception queue” companies on the surface. Some start at the queue itself: referrals, denials, permits, and billing disputes. Others start one step upstream, at the authoring event that feeds the queue. But the question that determines long-term value is the same: does the product sit where decision traces are a natural byproduct of usage?

We see two viable entry points.

1. At the exception itself.
   Tennr starts where the referral is already broken. EvenUp starts where the claim is already disputed. The exception happens often, failure or delay is expensive, and resolution requires context. Every resolved case produces a decision trace, the thing the customer trusts, forwards, reviews, audits, and acts on. It is also what lets the product learn. When resolution *is* the product, the trace comes free with every case cleared.

2. At the authoring event upstream.
   Abridge is a strong example. The clinical encounter is not always an “exception” in the operational sense, but it is always a high-context moment that must be documented. The exception queue emerges downstream when that clinical detail flows into medical coding, risk adjustment, and revenue protection. The decision trace comes initially from what’s documented, then deepens as the documentation moves through the workflow. Once you own the context, you begin to see the exceptions downstream.

Both paths work because both end up in the same place: sitting inside a high-consequence workflow, generating decision traces with every resolution. The difference is the entry point. The risk of starting at the queue is that you’re in the messiest, most integration-heavy part of the workflow. The risk of starting upstream is that you might generate context without ever reaching the queue, where it compounds. Abridge works because clinical documentation *necessarily* flows into coding and billing. An authoring tool that produces a useful document with no downstream workflow destination is just a productivity tool.

Broader platforms, Harvey, Legora, or TrunkTools, face the same question at each workflow they touch. They can be very large without looking like a single exception-queue wedge. But the defensibility question remains: does usage compound into proprietary workflow memory? Does the product learn the firm’s playbook, the client’s preferences, or category-specific nuance? Or does it remain a better interface to general-purpose intelligence? These companies need to identify queue depth *somewhere* within their breadth.

Owning the Decision Trace

As many investors in both private and public markets worry about the long-term health of the software industry, we believe the exception queue offers a compelling case for the defensibility of Vertical AI (and some Vertical Software incumbents). By executing work agentically rather than serving as a system of record, Vertical AI businesses aggregate proprietary knowledge of industry workflows. This knowledge, accumulated from customers using the platform over time, is arguably more important than simple information access: (1) it doesn’t exist in any structured format today (it mainly lives in domain experts’ heads) and (2) it is truly first-party, generated by product usage itself, mitigating any challenges around IP or ownership.

Aman Gour, CEO of FurtherAI, a startup automating insurance workflows, shares an instructive observation about the defensibility of the exception queue:

Over time, every escalation becomes a signal, every exception is feedback and every human correction shows where the runbook was incomplete. Over time, the workflow stops being a script and starts becoming the [insurance] carrier’s operating memory. This is the part the labs will find hard to reach. They will keep shipping better models and better general agents, and they should. But they do not sit inside a carrier’s production workflows long enough to learn why one account was escalated, why one risk was declined, or why an underwriter overrode the appetite guide and was right to do so.

That understanding only comes from running the workflow, in production, many thousands of times. The workflow you ship on day one is not the moat. The loop that production usage creates over time is.

Every resolved exception leaves behind a decision trace: which policy was applied, which precedent was relevant, who approved the override, what the customer accepted, how long it took, and what occurred afterward. Collect enough of these traces across cases and years, and you begin to build proprietary workflow memory, a growing record of how high-stakes decisions are made within a specific vertical and customer. Over time, workflow loops improve the product and move it closer to a true System of Intelligence.

The sequence looks like this:

The authored object is the bridge. It is the container for workflow-specific judgment. It records what evidence mattered, which rule applied, what exception was granted, what changed, and whether the resolution worked. In our Emerging Playbooks in Vertical AI essay, we described the ideal wedge as a product that creates what a human would otherwise have to create — documentation, a referral packet, a demand letter, an appeal — and hands it off to another system, as a human would. The exception queue sharpens that idea. The best wedge not only creates the authored object but also captures the decision trace—the reasoning behind it. Each trace adds to the company’s workflow memory, which is what makes the next case easier.

Two workflow loops stack from there. An across-customer loop, where patterns grow as you encounter more variants of the same problem within a vertical. And a within-customer loop, where the product learns the reasoning behind a specific customer’s decisions — the unspoken exceptions, the firm’s own decision logic that emerges only through direct interaction with the system.

As we argued in our essay on Vertical AI moats, the software vendor earns additional customer trust over time, enabling multi-product expansion. The core products continue to accrue defensibility through workflow loops, while each new product benefits from growing platform lock-in. The expansion path should follow the workflow and the decision trace.

Startup examples make this more tangible. Tennr can build referral-specific workflow memory: which providers send incomplete packets, which specialties create the most back-and-forth, which payers slow down which paths, and which missing fields predict leakage. EvenUp can build firm and case-specific workflow memory: which records change settlement posture, which injury patterns matter, which insurers respond to which arguments, and how long different case types take to resolve. For Abridge, the workflow memory surrounding the clinical notes can include specialty-specific documentation patterns, billing and coding support, care-team handoffs, and the institutional memory of how clinical cases are handled.

This is the missing middle between the authored object and the System of Intelligence, and where the moat begins to deepen. Not because the company “has data” in the abstract, but because the company builds a growing record of how high-stakes decisions are resolved within the workflow. A product that only reads data may still deliver insight. A product that resolves exceptions and captures decision traces can serve as the foundation for a System of Intelligence.

The Exception Queue Playbook

The queue is your wedge. Once you find it, the rest of the company-building question becomes more concrete. What does delay cost? What work product resolves the case? What reasoning should survive the resolution? And what does every cleared case teach the system that makes the next one easier?

The goal is not just to clear the queue faster. The goal is to turn each resolution into durable workflow memory: what evidence mattered, which rule applied, who accepted the answer, and what happened afterward. That memory is what produces the workflow loops for long-term defensibility. It is what lets the product improve, expand, and eventually become the System of Intelligence.

Here is the playbook as we see it:

1. Find the queue. The queue is usually visible before the software exists. It shows up as inboxes, spreadsheets, faxes, tickets, review comments, phone calls, aging reports, unanswered requests, and entire teams whose job is to chase the same class of exception every day. If no one owns the exception today, selling may be difficult. If an overworked team owns it with email, tribal knowledge, and manual judgment, pay attention.

2. Price the exception cost. The cost of failure or delay is how much the customer loses when the exception is handled incorrectly, late, or not handled at all. High cost is often the cleanest signal of urgency. Delay can be outright failure for certain queues. A prior auth approved too late can still fail the patient. An RFI answered too late can result in losing the bid.

3. Choose your entry point. You can start at the exception itself, where resolution is the product, or at the authoring event upstream, where you capture the context the queue needs—both work. What matters is that decision traces are a natural byproduct of usage, not a reporting afterthought.

4. Identify the authored object. The product should create something the customer already recognizes as work product: a referral packet, appeal, demand package, RFI response, permit submission, audit workpaper, coding recommendation, freight dispute file, or clinical note. This is where adoption starts, because the product does the work a human would otherwise have to do.

5. Capture the decision trace. Every resolution should preserve the reasoning: what evidence mattered, which policy applied, what precedent was relevant, who approved it, what changed, how long it took, and what happened afterward. This is cheap to design early and painful to retrofit later. If the product clears the queue but loses the judgment, it has automated labor without building memory.

6. Build the workflow loops. Every new case should make the next case easier. Not because the company has more raw data, but because the workflow memory becomes denser with useful relationships: payer behavior, reviewer preferences, jurisdiction quirks, customer policies, approval patterns, escalation paths, and outcomes. The question is whether each resolution adds to what the system knows.

7. Expand from the queue. The next product should be adjacent to the workflow memory already earned. Tennr can move from referrals into adjacent patient-access workflows. EvenUp can move from demand packages across the claim lifecycle. Abridge can move from clinical notes into coding and revenue cycle management. The expansion path follows the workflow and the decision trace.

Of course, no queue is the same, and false positives abound. High-volume, low-consequence workflows create usage without urgency. High-consequence, low-frequency workflows create consulting projects. Read-only workflows can produce insight, but often struggle to create compounding workflow memory. Broader platforms can grow large, but they still need queue depth somewhere within their breadth.

The best wedges sit where repeated resolution creates proprietary workflow memory. The best Vertical AI businesses will power a high-consequence workflow long enough to institutionalize that memory and turn it into a product that’s harder to leave every time it’s used. In the vertical application layer, selling intelligence is an ephemeral game — the real winners will own judgment.

Thanks for reading The Verticalist!

Euclid is an inception-stage VC built for Vertical AI founders. If anyone in your network is considering building in Vertical AI, we’d love to help. Just drop us a line via the comments below or on LinkedIn.

Dan Teran and James Gettinger are the co-founders of Gutter Capital, a New York-based fund that just closed its $75M Fund III focused on pre-seed investing. Previously, Dan founded Managed by Q — a vertical SaaS platform and B2B marketplace for office services — that he sold to WeWork for >$200M in 2019. James is a computer scientist turned professional gambler. Together, they made 100+ angel investments, launched Gutter, and most recently Elbow Grease: an accelerator that invests $300K into startups and spend 10 weeks working alongside them in Gutter’s NYC office. Tune in to hear our discussion on the changing profile of Vertical AI founders — and learn the three counter-intuitive founder signals Dan & James are leaning into.

Watch Full Episode Here

Today’s Episode

Most VCs will tell you they’re “founder-first” investors. In practice, that usually means they’re pedigree-first — screening for model lab resumes, repeat founders, and known technical programs as proxies for competence. The logic is understandable: at inception (indeed, often at pre-seed), there’s little else to evaluate. But Gutter Capital’s Dan Teran and James Gettinger argue the default is not just lazy — it’s empirically wrong. The biggest venture outcomes in history were built by founders who would have failed today’s pattern-matching. And AI is widening the distance between who looks fundable and who actually ships.

We’ve written about the fundability trap before — the recursive loop that sees investors pattern-match on the founder archetype most likely to raise the next round from downstream investors, making that archetype more common, reinforcing the pattern. But inherently, this behavior is retrospective. In a world changing this quickly, the founder profile best suited to win is changing too. Gutter’s entire model is a rejection of this tautological thinking in founder profiling. What emerged from our conversation are three signals Dan and James have found actually predict founder performance from Day Zero — and none of them are resume-standard.

Evidence Density Over Narrative Polish

Dan revealed that when Gutter backtested all of their Elbow Grease applications — using AI to analyze what actually predicted which founders advanced — the single strongest signal was what they call “evidence density.” Not narrative quality. Not market size. Not pedigree. Evidence density: the amount of hard, specific proof per word in a pitch.

How much hard evidence — per word in a pitch — is there was the biggest predictor of if we were actually going to pass them on to an interview.

Most founders spend their prep time perfecting the story — the narrative arc, the vision slide, the TAM chart. Gutter is saying the story matters less than the stack of receipts behind it. Specific customer names, metrics, quotes: “We reduced this from X to Y for Bob. And you can call Bob and here’s his phone number. And by the way, Bob has a cousin Jim, and Jim converted too.”

Gutter’s pro tip for founders takes it a step further: before an investor even asks for references, proactively send them three recorded customer conversations. A pitch to a prospect. A QBR with a design partner. A particularly insightful customer discovery call. The point is that it’s live, unscripted, and features a real person the investor could theoretically call. Most investors won’t. But knowing they could changes the calculus entirely.

Their experience aligns with ours here at Euclid: the strongest founders aren’t always the smoothest presenters — they’re the ones who demonstrate an uncanny aptitude for turning over every stone in their focus area. They can answer any question about their customer’s workflow in granular, unrehearsed detail. They’ve put in the work to understand the key objections and the right first ICP. Even if they aren’t from the industry themselves, they can talk the talk.

Evidence density is really a proxy for customer obsession. And as we’ve argued over and over again in our past work — and heard frequently from top vertical founders — customer obsession unlocks moats. It enables deep workflow integration, uncovers distribution advantages, powers referrals, and informs future extensibility of product and TAM. Nothing deflates interest in a deal like signs that a founder is less than completely obsessed with their vertical and customer.

A quick word from our partner on the Verticals podcast, Parafin. Recently named to the Forbes Fintech 50, Parafin has extended more than $25B in financing to their customers’ customers, powering embedded capital for platforms like DoorDash, Gusto, and Kajabi. Purpose-built for vertical platforms serving SMBs — lending, cards, and insurance without building the infrastructure yourself. Explore a custom program for your platform today.

Bad References, Good Founders

While evidence density was the most important pattern-match we dug into with Dan and James, there were several others worth noting for founders. One involved references — the cornerstone of inception-stage VC diligence.

Many treat reference calls as pass / fail. Get on the phone, ask banal questions, and if the answer is lukewarm, the deal is dead. Gutter takes a more interesting position — and it starts with asking the right question. Dan told the story of an Elbow Grease founder whose past references weren’t exceptional. But when they double-clicked with the CEO of the founder’s prior company, that “bad reference” was completely reframed:

Is the question ‘would I hire him back into the job in which he worked for me,’ or ‘would I bet on him as a founder?’ Because those are two completely different questions.

Dan speaks from experience. When he started Managed by Q at 24, he shared, his references weren’t particularly strong. He’d been a paralegal. James jokes that Dan offered him, a friend at the time, up as a reference because there just weren’t that many to choose from. Yet Homebrew led Dan’s first round pre-product because they didn’t mistake an absence of evidence for evidence of absence, as the saying goes. James — a former pro poker player turned angel — also understands the value of unconventional backgrounds better than anyone.

The broader point: at inception stage, you’re not hiring someone into a job. You’re betting on them in a role they may not have ever had. A reference that says “difficult to manage” might be the signal, not the noise. Not thriving in a large, slow organization, managerial intransigence, a desire to make the big decisions — these are all supportive of CEO potential. The question isn’t whether a founder was a good employee. It’s whether there’s evidence they will be a relentless, resourceful builder. Those are different skill sets, and conflating them is a fundamental investor mistake.

The Five-Kid Vibe Coder

There’s a prevailing narrative that AI has made the 19-year-old technical founder the apex predator — native to the tools, no legacy mental models, infinite energy. Dan pushes back: “When everyone’s telling you that the best founders are the teens, we also want to keep an open mind that that might not be true.”

Dan pointed to one of the standout founders from Elbow Grease’s first cohort — the founder of a company called Punch. Five kids, sixth on the way, based in Atlanta. His wife gave him permission to come to New York for 10 weeks. He was in the office every night and every weekend because, as Dan puts it, “his attitude was, I know this is my shot.” He ships product faster than anyone in the Gutter portfolio.

We explored the age question in No Country for Old Founders, and AI is accelerating the trend. While YC has elevated the youth profile, the numbers suggest that experience and expertise matter greatly. At Euclid, we generally find a balance between youth and experience to be the most common sweet spot. Gutter believes that AI has opened up opportunities for mid-career operators, even those outside the tech sphere or straight from industry. Domain experts who spent a decade without engineers or butting heads with them — who understood their industry’s workflows cold but couldn’t build the solution themselves — now have the tools to start independently. They’re not burning time gleaning their customer’s problems from user interviews because they lived them.

For Dan and James, it’s not that every person in every field will become a well suited startup founder. It’s that a rarefied slice of domain experts who had the tenacity and vision, but lacked a natural catalyst are now unshackled. Gutter often sees them as better positioned than any CS grad because they already have the customer relationships, the industry context, and the specific insight that makes a Vertical AI product right rather than just functional. As Dan put it, “People who have chafed at engineers their entire career now don’t really need to deal with them that much to build something that actually might be really great.”

The Takeaway for Vertical Founders

The founder profile that wins in Vertical AI is shifting, and most of the ecosystem hasn’t updated its priors. At Euclid, we still believe presentation, references, and youthful velocity are important. But we’ve always felt it’s more important for founders to be authentic — and to lean into what makes them a unique founder — than to mold themselves to trends on Hacker News or whatever platform funds choose to market around that year.

If you’re raising at inception stage: stop optimizing your pitch for narrative polish and start optimizing for evidence density. Every slide, every answer, every follow-up email should be stacked with specific names, specific numbers, and specific proof that you understand the problem better than anyone. Send customer call recordings before they’re asked for. Make the evidence so dense that investors don’t need to trust your story — they can verify it.

And don’t hide from your unconventional background — lead with it. Truly first-principles funds are increasingly embracing the “post-pedigree” founder as AI changes everything. Euclid and Gutter are aligned on one thing: domain and customer obsession are paramount. If you spent 15 years in the trades and just learned to vibe code — yes, you’re still going to have to justify why you are best suited to do everything a startup founder has to do to build a big business, and that’s no joke. But it doesn’t have to be a weakness to paper over either. With proper self-awareness, it’s a strength to build around. Most investors, like most humans, are allergic to inauthenticity. When someone leans into their reality, perceptions-be-damned — that is a rare and powerful trait.

If you’re a VC who outsources their diligence to consensus perceptions of pedigree, you might be missing out on some of the biggest contrarian winners of this AI era. At minimum, take note: a professional gambler and an exited founder just raised $75M to bet against you. And they aren’t alone.

Nikhil Basu Trivedi is the Co-Founder and General Partner of Footwork, a Seed and Series A firm with ~$400M in AUM across two funds. A 16-year venture investor and two-time Midas Seed List honoree, Nikhil has built a track record around a specific type of company — the ones that refuse to check a single box. He was an early backer of Canva, Lattice, and Frame.io, and Footwork’s current portfolio includes a strong Vertical AI lineup: Elicit (AI for scientific research), Fuse (AI-native loan origination for credit unions), and Confido (financial operations for CPG brands).

Tune in to hear why Nikhil believes some of the best Vertical AI businesses will blur the line between consumer and enterprise — and much more.

Watch Full Episode Here

Today’s Episode

Software venture capitalists often self-select into one of two groups: consumer and enterprise. Since firms and GPs often prefer one or the other, there’s an incentive for founders to do the same. But some of the most valuable companies of the last decade — Canva, Slack, Zoom — tapped into both user types, even if they began in one. In the AI era, we’re seeing that crossover more often and earlier. Especially in Vertical AI strategies where the end user is both the buyer’s customer and a consumer or prosumer in their own right.

Nikhil Basu Trivedi has spent 16 years betting on companies that straddle this divide. In his view, B2B vs. B2C can be a bit of a false dichotomy — and that divided perception can hold both investors and founders back from seeing the biggest long-term opportunities. In this episode of Verticals, Nikhil joins us to explore why the consumer-enterprise line is blurring, what that means for expansion and defensibility, and how to identify verticals where AI can build businesses that are category-defining across both consumer and enterprise use cases.

A quick word from our sponsor on the Verticals podcast, Parafin. Recently named to the 2026 Forbes Fintech 50, they’ve extended over $25B in financing to nearly 50K businesses on platforms like Amazon, DoorDash, and Walmart. Embedded financing programs, custom-built for vertical founders and operators like you. Click here to learn more today.

Consumer vs. Enterprise is a VC Construct

One of Nikhil’s core convictions at Footwork is that some of biggest companies in the world serve both consumers and enterprises. While he does believe in early focus, expecting founders to “pick a lane” and stick to it can be short-sighted. About three-quarters of Footwork’s 25 investments are companies that “self-identify as both consumer and enterprise in their DNA.”

The poster child is Canva. When Nikhil led the seed round in 2014, it was a consumer design tool. Today it does over $4B in ARR, with more than $500M coming from enterprise contracts (B2B seats of 25+ growing at 100% year-over-year). Nobody planned for that S-1 line item in 2014. It happened because the team built an end-user experience so good that it pulled itself into the enterprise organically.

Nikhil sees the same dynamic playing out at the model layer. Anthropic’s success versus competitive labs is widely attributed to their focus on enterprise coding agents — but that focus led to uptake in the enterprise and with consumers. Even if their total consumer market share (in the low-mid single digits) is less of a focus, the ICP of higher-end prosumers is increasingly Claude-centric. Meanwhile, OpenAI’s heavier consumer tilt, in Nikhil’s view, “has hurt them [in the enterprise],” forcing them to play Codex catch-up.

The takeaway isn’t that you need to do both on day one. Footwork — like Euclid — agrees that maniacal early focus is critical. But a CEO’s job is to continually balance two priorities: near-term execution with long-term vision. If you’re not building the team, product, and roadmap to eventually serve your most ambitious market opportunity — which very well may span enterprises and consumers — you risk capping your outcome.

AI Makes the Crossover Easier Than Ever

What’s changed is the product surface area that AI unlocks. Pre-AI B2B SaaS was reliant on UI for data input / output — a form to fill out, a dashboard to check, a record to update. The value proposition was worker acceleration: move things through the funnel, inform an analysis, serve as a system of record. LLMs have enabled many of today’s AI-native products to do the actual work — interact with users in the course of their day-to-day activities, through natural language, voice or even without active user input altogether.

Three companies from Footwork’s portfolio illustrate the pattern:

Elicit started as a self-serve AI research tool — type a question, get a systematic review of published literature grounded in 125M+ papers. Academics and researchers adopted it bottom-up. Now it’s spread into all of the top 20 pharma companies with enterprise-wide licenses. Pharma is a $1.2T revenue vertical that has produced remarkably few scaled software companies — Veeva being the notable exception. Elicit is finding purchase through the prosumer door.

It’s a strategy reminiscent of Open Evidence, which took a similar path in clinical medicine. The self-serve AI search & discovery tool for physicians may have leveraged hospital systems for early distribution, but the only way for them to achieve their incredible uptake (the majority of US doctors) was via a consumer-esque, PLG clinician-first adoption model. The company recently raised $250M at a $12B valuation. On the episode, Nikhil, Nic, and Luke all agreed one Vertical AI high-flyer that has a strong case for meriting its rich valuation — and a big part of that is precisely because of its dual enterprise-prosumer presence.

Fuse, another Footwork portco, is a different flavor of hybrid Vertical AI. It’s an AI-native loan origination system for credit unions that raised $25M in its Series A led by Footwork. In this case, the consumer element is inherent to the ICP of credit unions. Core to the value proposition is that the borrowers applying for a loan — their customer’s customer —gets a fundamentally better experience thanks to the AI-native features. Fuse gives them a shot at matching neobanks and other more digitally-inspired borrowing options, without rebuilding from scratch. AI’s form factor lends naturally to reimagining the B2B2C experience.

In short, AI has collapsed the distance between the B2B buyer and the end consumer. It’s also increased the bandwidth of user data capture, while reducing its friction — something inherently valuable to enterprises. When your product does the work instead of just recording it, the consumer-facing experience isn’t a “nice to have” — it’s the product.

The Toast Warning

If the upside of crossing over is a bigger TAM and stronger distribution, the downside of not crossing over is vulnerability. Nikhil pointed to Toast as the cautionary tale — a Vertical SaaS darling that was late to capturing the consumer side of the restaurant experience.

“I think one of the big issues for Toast,” Nikhil shared, “is having not captured the consumer side and therefore being vulnerable. They’ve not built a brand with consumers. I think much more about ordering using DoorDash than I think about Toast.”

This maps to a broader pattern we’ve written about in Navigating Vertical B2B2C: in verticals where the end consumer transacts through the platform, owning only the business-side relationship leaves the door open for someone who captures both. Embedded payments, embedded lending, consumer-facing ordering — these aren’t just revenue lines. Ultimately, businesses that serve consumers want to be where their customers are, and capturing a more highly fragmented consumer user base can be a much more difficult strategic moat to unseat.

Tune into the conversation with Nikhil Basu Trivedi, Co-Founder & GP at Footwork VC here — and subscribe on YouTube for new episodes of the Verticals pod every week!

Watch Full Episode on YouTube

Brand: The Moat VCs Keep Ignoring

We also discussed how brand may be the most underrated defensive asset in Vertical AI. As we’ve discussed plenty here at The Verticalist, proprietary data, workflow loops, and network effects are the pillars of defensibility that seem to be growing only stronger post-LLMs. Nikhil’s argument is that brand — the thing people trust, the thing that occupies mental real estate — is another reason some incumbents will survive the AI transition. Even if they’re slow on agentic features.

Inverting this point: Footwork actively looks for verticals where the dominant players have terrible NPS — massive companies that nobody actually likes. His examples:

• LabCorp and Quest Diagnostics: both $20B+ market cap companies in blood testing. “Does anyone say they love Quest or LabCorp? You feel terrible going into it.” Footwork has a stealth investment going after this exact gap.

• The Farmer’s Dog: Nikhil’s prior investment, now a scaled brand in fresh pet food, built in a category where legacy brands had eroded consumer trust.

• CPG: This is one where it’s less poor NPS because of a dominant player with no incentive to change. It’s that the landscape is fragmented, with dominant players that aren’t even universally well known. Footwork backed Confido.

The implication is that, in a world where AI makes it possible to rebuild category experiences from scratch, low-NPS verticals with entrenched incumbents aren’t just underserved markets — they’re brand vacuums that can perhaps be filled faster than in past eras, thanks to the potential for faster switching cycles that AI is enabling.

Mapping the Model Lab Kill Zone

Not every Vertical AI company is equally defensible. We closed the episode with a rapid-fire exercise: are the most highly valued Vertical AI companies justified, frothy, or too early to tell? Here’s where the crew landed:

Harvey (legal) — ~$200M in ARR, $11B valuation. The 3/3 consensus was frothy. The legal vertical was one of the first to adopt AI at scale and the ACVs are large. But the product sits squarely in what LLMs already do well — document analysis, summarization, research. Competitive intensity is only increasing, from both new dedicated Vertical AI players like Legora, and most worryingly, from the model labs.

Basis (accounting) — est. ~$100M in revenue, $1.15B valuation. Our take: 1 frothy, 2 too early too tell. As Luke put it: “Selling AI into white-collar industries where somebody’s sitting in front of a computer and they’re definitely going to try it right next to Claude and ChatGPT is tougher.”

Abridge (healthcare) — $100-200M in revenue, $5.3B valuation. Our take: 1 frothy, 1 justified, 1 too early too tell. Luke finds health system incentives and the competitive threat from Epic — which we’ve covered in past episodes — too concerning. Nic was a bit more bullish, with respect to the likely data advantage the company has managed to amass so far. Interesting, this is one no one seems to think the LLM “kill zone” will expand to.

Which teams will build enduring businesses outside the kill zone? Footwork and Euclid share a similar view: deep domain knowledge built over years helps founders understand why a vertical needs an AI solution, and the industry-specific workflows and data that will make their solutions defensible over time. Vertical AI values builders who understood the customer — and perhaps the customer’s customer too — long before they pitched their first design partnership.

The Takeaway for Vertical Founders

The consumer-enterprise divide was always more of a fundraising or VC-matching construct than a company-building truth. AI is blurring the line. When your product does the actual work, the end user isn’t abstracted away behind a dashboard a consumer would never use. Nor is building unique BI features for every ICP a crippling consideration. Users are more a fundamental part of the product surface and perhaps also the long-term data or revenue opportunity.

That doesn’t mean every Vertical AI company needs a consumer angle on day one. It means founders should be “open and aware,” as Nikhil put it, that the path to a category-defining outcome may well run through both sides of the market. Build a team that can execute on multiple surfaces. Think about the consumer experience even if your first sale is enterprise. And pick verticals where the incumbent brand is weak, the market is large, and where the model companies can’t serve the pain point by lightly rebranding their core LLMs. Perhaps the convergence of the consumer-enterprise divide is just another expression of something we see as increasingly obvious: the market size for the application layer of the future is bigger than you think.

See you next week.

Kyle Lui is a General Partner at Bling Capital, where he leads early-stage investments in product-driven founders across SaaS, consumer, digital health, and fintech. Before Bling, Kyle was a Partner at DCM and two-time founder (his startup ChoicePass was acquired by Salesforce). Bling has done something rare in VC: open-sourced its entire DD playbook. This week on Verticals, Kyle joins us to explore how his diligence framework is evolving in the AI era — and how founders and investors should update their priors.

Watch the Full Episode Here

Today’s Episode

Most VCs have spent the majority of their career in an environment where SaaS was so ubiquitous that strong gross margins were implicit, market sizing centered around a clear IT budget, and “hair on fire” problems outside of what software could do were irrelevant. In the AI era, every one of those shortcuts breaks down. AI margins range from negative to 90%. The TAM for an AI company might capture 5-10x as much as SaaS did. And the standard retention expectations and benchmarks may not apply.

Bling has published a seven-part diligence series that walks through exactly how they stress-test these questions — from hair-on-fire problem definition to customer segmentation, market sizing, and financial modeling. The shared thread is that diligence isn’t a test to pass — it’s an alignment exercise between founder and investor on a believable plan to reach venture scale. Kyle walked us through how each piece of that framework applies to AI-native companies, and where the conventional playbook may be falling apart.

The Hair-on-Fire Test Gets Harder

Bling starts every diligence process with a deceptively simple exercise: describe the hair-on-fire problem you’re solving, then show — step by step — how the current workflow works and how your product replaces it.

In Vertical SaaS, this was often straightforward: manual process, expensive, error-prone — software automates it. In Vertical AI, the exercise is more revealing because the “10x better” claim is harder to pin down. An AI immigration law firm compressing 40 hours of attorney work into four is a clear 10x. But a voice agent that handles 80% of customer calls and escalates the rest — is that 10x better, or is it delivering more work to the customer when escalations go sideways?

Kyle's succinct advice: if you can’t document the current behavior and the future behavior of customers in concrete steps, with clear ROI math — even if those steps are likely to evolve down the road — you don’t yet understand your own product’s value.

ICP Segmentation is Increasingly Important

Bling’s customer segmentation exercise requires founders to define segments that are mutually exclusive and collectively exhaustive — then attach average customer value, margin, and a go-to-market strategy to each. The most common mistake they see is reliance on top-down TAMs. Especially when you’re offering a net-new value proposition, a bottoms-up breakdown of ICP is critical.

Segmentation is more important in the era of AI. A roofing company with a $30K ACV wouldn’t touch voice agents two years ago — every lead was potentially worth $30K, and the ROI for humans vs. 80%-resolution AI was clear. Today, with a greatly improved voice stack, that same company might adopt at 95% resolution. But let’s examine an adjacent segment: residential HVAC. With its $5K average tickets, different math is at play. With every lead less valuable, it’s more of a volume game — meaning the bottleneck is less conversion and more lead processing. That means, they probably were open to adopting Voice AI sooner.

So if you’re sizing “home services voice AI” as a homogenous market, you’re hiding the fact that your actual addressable segment today might be 10% of that. Bling’s framework forces you to show which segments you can win now, which you’ll expand into, and what has to change for each.

Getting Venture-Scale on Gross Profit

Kyle runs a specific exercise with every company Bling backs: how does this business get to $100M and $500M in gross profit? Not revenue — gross profit. The framework is a fill-in-the-blank sentence: “To get to $100M in gross profit per year, we need X customers paying us $Y per year, where X represents K% of the market.” If you can’t complete that sentence with credible inputs — in Kyle’s mind — you don’t have a venture-scale business.

In the AI era, Bling has added a third tier: $1B in gross profit. Their rationale is that multiples have compressed, competition has intensified, and the power law is steeper than it was two years ago. Founders sizing their TAM around top-line revenue or GMV instead of gross profit is more dangerous than ever because AI inference has created a wildly variable COGS situation from startup to startup.

The gross profit lens also reveals which markets are bigger than SaaS economics suggested. Bling portfolio company Alma, an AI-native immigration law firm, captures the margin delta when AI compresses 40 hours of attorney work into four — without proportionally lowering the price. Miniva, in high-volume manufacturing, replaces quality control hires that factories can’t find. Part Bay is building a vertical AI marketplace for used auto parts — a category most VCs would call too niche, until you size it in gross profit against the labor and inefficiency it replaces. A vertical play that looked too small at $80M in software TAM might now be venture-scale.

The D+ Financial Model

The final piece of the playbook might be the most counterintuitive. Bling’s financial modeling exercise explicitly calls for a “D+” model — a skeletal framework that captures 80% of the signal in 30–60 minutes. It covers four key questions:

• Who are you hiring, when?

• When does revenue start?

• How does burn scale?

• How long does the money last?

While the model is simple, there are some nuanced considerations, especially when it comes to the burn question. On some boards, Kyle is seeing companies now combine token spend and headcount into a single R&D line item — a shift unthinkable two years ago. If your negative margins come from compute costs on a falling curve, downstream investors will tolerate it. But if you’re burning because you’ve hired an army to deliver the service, Kyle’s question is pointed: “What’s your actual advantage against the non-AI-native companies?” As we explored in Deus Ex CapEx, the uncertainty around the steady-state of AI margins is more pervasive than ever.

The Takeaway for Vertical Founders

Bling’s playbook offers some great tactical advice for thinking about how you position your business going into a Seed process. The fundamentals remain as important as ever: solving a specific problem with a quantifiable ROI, rationally segmented customers with scalable unit economics, a credible venture-scale gross profit expansion story, and a financial model that shows an understanding of the tech-economics underlying your business. But for AI-native Vertical AI founders, the SaaS-era shortcuts for evaluating those fundamentals — top-down TAMs based on IT budgets, assumed gross margin opportunity, constraints on solvable “hair on fire” problems — must all be thoughtfully reconsidered.

See you next week.

Intrigued by the game as a possible model to explain price fluctuations in equity markets, British economist John Maynard Keynes discussed the optimal strategy in his 1936 book, The General Theory of Employment, Interest and Money. Keynes observed that the winning strategy was not to choose the face you found most attractive, but the face you believed everyone else would choose. And because everyone else could deploy the same strategy, to stay ahead, a person must think one step beyond everyone else—guessing others’ preferences, their guesses, what others think of those guesses, and so on. With each layer of analysis, the consideration drifts further from the original question: who is the most beautiful?

That recursion is now known as the Keynesian beauty contest. Keynes saw the same loop in markets: success in speculation comes less from picking the soundest asset than from guessing which one the crowd will bid up next, and when.

Venture capital now runs on the same logic, and the hyper-concentration of capital means ‘beauty’ is defined by an increasingly small number of players. This is the most concentrated venture capital market in history: roughly 75% of all venture dollars were raised by five firms in the first quarter of this year. In 2025, nearly 50% of VC dollars went into rounds greater than $500M — up from roughly 10% a decade ago.1

We’ve written about the structural side of this story before: the asset management theory of venture, the planned-economy dynamics, and the slow conversion of early-stage VC into an assembly-line market that prioritizes fundability over investability. What we wanted to expand on is the cognitive consequence — what consensus capital does to how investors think. In Vertical AI, that shift is especially dangerous because the thing being copied is not an idea, but a mistaken abstraction of a successful archetype.

When billions of dollars need to be deployed by a handful of decision-makers across a few hundred decisions a year, the firm's operating system has to rely on transferable templates. You cannot underwrite every company from first principles at that volume. So the operating system changes. You only invest in a narrow style of companies. You pattern-match on what worked before, outsource conviction to whoever bid before or after you, and compress every new opportunity into the shape of an old one. Pattern-matching isn’t a personality flaw of lazy investors. It’s the inevitable mechanism of scaled capital.

Pattern Matching, Predictably

Inevitable, however, isn’t the same as logically justified. Consensus capital didn’t invent pattern-matching — investors have always relied on it in some shape or form. What’s new is how narrow the band of acceptable patterns has become, and how punishing today’s venture market can be for those who diverge.

In a 2022 study of more than 16,000 startups and $9B of committed capital, Chicago Booth’s Diag Davenport found that roughly half of bad venture outcomes were predictable using only information available at the time of investment. The good-outcome model leaned on product characteristics. The bad-outcome model leaned heavily on the founder’s background. In a surprising twist, a model built solely on founder education data proved to be the single strongest predictor of which investments would underperform.

What’s striking is that this study isn’t telling venture anything it didn’t already know. It’s confirming one of its oldest adages, as Andy Rachleff described plainly: “When a great team meets a lousy market, the market wins. When a lousy team meets a great market, the market wins. When a great team meets a great market, something special happens.” In a recent essay, Dan Gray from Odin succinctly summarizes why this pattern occurs:

Somewhere along the way the industry convinced itself that the ability to raise capital was itself a desirable founder trait and this logic became recursive. Investors began pattern-matching on the founder archetype most likely to raise the next round, making that archetype easier to fund, reinforcing the pattern. —Dan Gray, Odin

That is the Keynesian’ beauty contest applied to hundreds of billions of dollars of venture capital. Ben Graham famously described the market as “a voting machine in the short run, a weighing machine in the long run.” Venture has gotten very good at voting. It has gotten worse at weighing.

Vertical AI Mad Libs

Walk into almost any demo day over the past year, and you’ll hear it:

“We’re [voice / agentic / services] AI for [vertical].”

This is not an indictment of founders but of the funding ecosystem. Founders use these formulations because incubators, accelerators, and investors want legible startups. And pattern matching by analogy is the easiest way to get there. But this is pattern matching at the worst possible level of abstraction. It’s not even pattern matching on markets or products. It’s pattern matching on the application of a model capability across markets, treating verticals as interchangeable substrates that a known template can be poured into.

The error compounds as the templates compound. Pattern matching on the technology template ("voice AI works”), the founder template ("ex-Stripe, technical, YC, young”), and the category template ("this worked in legal, so it will in another vertical”). This looks, from the outside, like a lot of analysis. It is not. None of those templates contains any actual constructive work on the market or the problem. The founding team’s insight into the market's shape and unique characteristics is the variable that determines whether the technology can achieve product-market fit.

Vertical markets are not fungible startup substrates. Legal looks nothing like supply chain, which looks nothing like construction. In fact, even most construction tech does not look like other construction tech. Construction is not one market, but the amalgamation of hundreds of markets. The technology may travel; the market dynamics never do. Document review and generation may be a multi-billion dollar problem in legal, but structurally irrelevant in construction. Pattern matching across verticals, or even sub-verticals, is a category error, treating the variable that matters most as if it were a constant.

In our essay on Vertical AI diffusion, we discussed how early Vertical AI breakouts are largely workflow enhancements: taking an existing task and using an LLM to make it cheaper or faster. Superficially, this is what templates copy when they replicate documentation tools like Harvey, Abridge, and EvenUp, or voice agents like Assort and Avoca.

Even in the largest Vertical AI categories like legal, medical, and real estate, combined spending is only in the hundreds of millions. Coding (and to a lesser extent, legal) broke out because their market structures made AI adoption unusually favorable. Part of that is capability — LLMs are strong coding agents. But part of it is function. Software development is a rare category where the buyer is often the user, the output is digital, and the result is at least hypothetically measurable. Switching costs are also unusually low: developers can adopt and abandon coding agents much faster than most vertical workers can replace core workflow systems. Legal looks similar on the surface, but for different reasons: the corpus is large, the output is reviewable, and human judgment remains in the loop.

Anthropic’s own analysis speaks to this vertical heterogeneity. Its observed-exposure measure — which combines theoretical LLM capability with actual Claude usage in work settings — finds that LLMs could theoretically handle 88% of legal tasks versus observed exposure of 15%. That is the largest gap of any vertical on the chart. The model is not the main bottleneck. The diffusion of AI across idiosyncratic workflows across vertical markets is the constraint, and everything between theoretical capability (which will continue to increase as technology matures) and observed usage is where future Vertical AI applications will thrive.

The frictions that define most real vertical workflows differ across markets. The pattern matchers look at existing vertical winners and try to template that deployment model across new sectors. But the real opportunity in vertical markets is not to copy what worked in one vertical into another in the hope of striking gold. It is to build new, vertical-specific Systems of Intelligences that can shrink the usage gap and create new categories. While the pattern matchers race to apply yesterday’s templates to new markets, the problem finders will reshape those markets with earned insights.

Problem Finding is the Job

A strong framing of what we call a vertical founder’s “earned insight” comes from a 2022 paper by Mattia Bianchi and Roberto Verganti. They argue that entrepreneurship has been systematically misunderstood as an exercise in problem-solving, when in fact it is primarily a problem-finding endeavor. The founder’s most important act, in their analysis, is identifying a problem worth solving. Everything else — the product, the go-to-market, the wedge — follows from the quality of that framing.

Dan Gray’s “centaur” concept gets at the same point: “rather than the tired question of jockey or horse, the investor’s job is to recognize the centaur.” The founder and the idea cannot be evaluated separately, because the idea is the material expression of how the founder believes the future of a market can be shaped with technology. Templated pattern matching starts with a known solution and searches for new markets to apply it. Problem finders work in the opposite direction: they start with the market, identify the problem others have missed, and only then decide where technology can attack it.

Our analysis of Vertical AI founders attempted to quantify the backgrounds of successful founders in the AI era. Seventy-one percent of $15M+ Vertical AI rounds in 2025 went to teams with prior domain experience. In other words, earned insight (i.e., successful problem-finding) into a vertical market was the strongest indicator of success.

Per the anti-expert argument, if AI truly democratizes domain knowledge — enabling technical founders to iterate and learn industries in weeks rather than years — then the vertical-founder advantage should shrink in vertical AI relative to vertical SaaS. AI-native, low-experience founders should be pulling ahead on attracting dollars and VCs, as the outsider disadvantage compresses.

Our analysis, however, shows the opposite. In Vertical AI, >70% of leading 2025-funded startups have founders with prior experience in or around their industry — and those founders raised deals 2.25x larger, with greater overall capital velocities, and higher VC scores. Every metric points in the same direction: the more AI-native the company, the more domain expertise matters, not less. — The Verticalist

Let us pre-empt the obvious objection: aren’t we now claiming that founder background is a strong predictor of success? Isn’t that just pattern matching?

No. The distinction is between pedigree and proximity. Pattern matching evaluates the founder against a narrow set of archetypes: school, prior startups, prior funding, network, and fundability. Problem finding evaluates the founder’s relationship to a specific market problem: what they know, what they have seen, and what they understand that outsiders miss.

Founder + Problem + Market

Underwriting early-stage Vertical AI should not be defined by the founder, the idea, or the market in isolation. It should be founder-plus-problem-plus-market, evaluated as a combined entity. Finding the problem within a market is closely intertwined with the structural shape of the opportunity and the founder’s frame for it. At Euclid, this allows us to evaluate opportunities from as early as inception-stage as founder+market+problem. As we discussed in our essay on PMF, we don’t mind investing pre-product and pre-revenue, but we do care about underwriting “identified market demand.”

There’s a through-line here to a piece we wrote early last year on market sizing in vertical markets, where we argued that the standard VC approach to TAM is backward-looking: it determines which categories are venture-scale based on past outcome sizes. We think early-stage vertical investors must train themselves to do the opposite: a forward-looking analysis that asks what a vertical will look like after AI reshapes it.

There’s the infamous example that many VC firms initially passed on Shopify because they perceived the market as too small. At the time, there were roughly 50K online stores, and most concluded that the market was too small for venture-scale returns. Of course, the missing critical insight was that by lowering the friction of launching an online store, Shopify created a much larger category in the process. Backward-looking TAM is another lazy pattern-matching exercise. Great teams with earned insights can create large, new categories. As we highlighted in our original market sizing essay, it’s not the market size at entry that matters; it’s the market at exit.

Evaluating the problem and market requires answers no template can provide. What is the urgency of this problem for buyers, and what gets unlocked when AI can be applied? Is the shape and scale of the problem widely found within the segment? Who are the natural buyers, and what do they already spend on the outcome the product produces, not the current tech spend? What’s the path from wedge to multi-product TAM expansion— which adjacent workflows are reachable, and where are the data & workflow loops that strengthen defensibility?

None of those questions can be answered solely by analogy to another vertical. There might be similarities, for certain, but templated analysis is doomed to fail, especially in new and rapidly expanding markets across Vertical AI. They require sitting with the market AND the problem alongside operators who will shape it. Nabeel Hyatt at Spark describes how he evaluates founders and distinguishes real executors from polished pitchers by examining what comes out of their hands: the product itself and the human behind it.

At our stage, the combination of market and problem largely stands in for product, but the premise still holds. We look for people who understand the day-to-day constraints of a vertical so deeply they can take the work apart — and who have enough technical instinct to rebuild it with AI. No matter the pace of LLM capabilities, we continue to believe problem-finding founders will remain the rate-limiting factor in industry transformation.

First Principles & Vertical Investing

From inside a large fund with a partnership that needs to see legible decisions, narrow pattern-matching is an inevitable outcome. The incentives are different but equally alluring downstream, as Rob Go from NextView illustrates:

As capital is concentrating, it's moving into a very narrow set of styles of company, and the companies that work within this narrow band raise more and more and more capital at higher and higher prices. And so, if you're a seed fund and your goal, you know, is to sustain yourself, not necessarily to like generate a great return like from an ultimate exit, but it's like, "Let me get a markup really fast and maybe multiple markups." It behooves you to then play that game. — Rob Go, NextView Ventures

There’s no denying the dopamine hit of quick markups. Combined with the desire to show strong performance metrics to a fund’s own investors, it’s an exceptionally strong incentive. As Howard Marks famously said, “you can’t eat IRR” — but you can raise your next fund (or accelerate your career) if you’re seen to be backing the consensus names.

Except the premise underneath that game is shakier than it looks. For all the talk of a closed circle of elite funds, the field of firms backing venture’s largest winners has widened substantially. In the 1990s, just 25 firms had at least one of the top ten deals in a given year, and five of them each captured more than 8% of all top deals over the decade. From 2005 to 2015, during a boom decade for new firm launches, 70 firms landed top-ten deals, and no single firm captured more than 8%. If you expect this dispersion to accelerate, as we do, then the era of consensus capital — where a handful of funds compress the industry into a single style of investing — is an inefficient planned economy. And planned economies misallocate capital by design.

The Imitation Game Will Have Few Winners

Set aside whether consensus even contains the winners, the game doesn’t pay even when it works.2 This is the Keynes beauty contest as played by today’s VC: the bet you’re really making is on the bidders, not the businesses — and the analysis drifts further and further from the original question: is this a good investment? And unlike a stock you can dump when sentiment turns, there’s no off-ramp when the music stops. There’s no crying in the casino. (Unless you convinced everyone your bespoke crypto thesis was actually about AI all along — in which case, congratulations.)

There’s a name for the posture this rewards. A few years ago, Glowforge’s Michael Natkin called out one of the tech industry’s favorite mantras — “strong opinions, loosely held” — as a glorification of toxic overconfidence. The original idea was sound: state your case hard, then update the instant the data turns. But the loosely held half rarely survives contact with practice, least of all in venture, where the feedback cycle runs long enough to bury the evidence. Deals get marked up, new funds get raised, and the conviction hardens.

We want to be clear about what we’re not saying. We’re not claiming that optimizing for problem finding and holistically evaluating team/problem/market will make you universally right. It doesn’t. We’re wrong more than we’re right. That’s the nature of the job: the base rate of being wrong at first check is undeniably high, and no framework fixes that.

But the humbling nature of the venture is no excuse to offload conviction or set aside rigor. Perhaps more importantly, seeking out problem finders doesn’t raise your hit rate so much as change what you’re holding when you do hit. Pattern matching crowds you into the same shrinking set of bets as everyone else, where being right pays little because conviction is borrowed, markets are crowded, and the price reflects the scale of consensus. Problem finding puts you in markets nobody else has underwritten, where being right — on the rare occasion you are — actually pays for all the times you weren’t. Meanwhile, shallow pattern matching is the operating system of consensus, the shortcut that lets concentrated capital deploy at scale without doing the market work underneath.

There is, of course, a certain irony here. The same instincts that once made a VC’s name — non-consensus, first-principles investing — now risk being recast as vectors of adverse selection. But the ones who hold their conviction — founders and investors alike — may be the ones who eventually look prescient. It makes the problem-finder’s patience cheap to acquire and expensive to hold, but in the long run, the market always wins.

In VC, playing the imitation game — backing startups predicated on the opinions of investors, allocators, or the market generally — is a tempting but fundamentally flawed strategy. What sets the greatest apart is their ability to find and commit to problems worth solving, regardless of what others may think.

Thanks for reading The Verticalist!

Euclid is an inception-stage VC built for Vertical AI founders. If anyone in your network is considering building in Vertical AI, we’d love to help. Just drop us a line via the comments below or on LinkedIn.

2

Data shows there’s very little evidence that king-making is a real phenomenon or a viable strategy for long-term startup success.

Ben Evans - AI Eats the World

The Odin Times

Best Served Cold

Read more

3 months ago · 24 likes · 1 comment · Dan Gray

The Odin Times

Mythical Beasts

Read more

4 months ago · 30 likes · 1 comment · Dan Gray

Mike Droesch is a Partner at Bessemer Venture Partners, where he’s spent nearly nine years investing across vertical SaaS, B2B marketplaces, supply chain, and — for the last two — leading the firm’s voice AI roadmap. He backed VAPI, which just closed its Series B at a $500M valuation. BVP is also in Abridge in healthcare, Rilla in home services, and Axiamatic in enterprise transformation. Before venture, Mike was an engineer and management consultant. He’s one of the most methodical vertical-specific roadmap builders in the business — tune in to learn what he’s seeing in Vertical Voice AI.

Check out our latest episode of Verticals with Mike and continue below for our breakdown of the landscape and the thesis that stuck with us most: voice as the new process mining, and the fuel for future Vertical AI defensibility.

Watch the Full Episode Here

A quick word from our sponsor on the Verticals podcast, Parafin. They’ve surpassed $100M in revs, extending >$25B in financing to their customers’ customers. They power embedded capital for platforms like DoorDash and Jobber, and are purpose-built for vertical platforms. Click here to explore a custom program for your platform today.

Today’s Episode

Voice AI has produced some of the fastest-growing vertical startups over the last two years. It’s also addressing some of the biggest open questions around application-layer defensibility. The infrastructure to make a voice agent that sounds human and can call tools is increasingly commoditized. So where does the moat come from? And for founders building voice-first products — what separates a feature from a company?

Mike has a view on this, shaped by two years of running Bessemer’s voice AI roadmap and many more backing companies across the stack — from VAPI at the platform layer to Abridge and Rilla at the application layer. The conversation covered the full landscape: build vs. buy decisions in voice infrastructure, why regulated industries adopted fastest (counterintuitively), where voice is overdone vs. underdone by vertical, and AI-native monetization models.

The most trenchant insight— and the one most relevant to how we think about defensibility at Euclid — is Mike’s thesis on voice as a fundamentally new data layer. Your system of record captures end states. Voice captures everything upstream: how decisions actually get made, how processes actually work, what sentiment and context look like in real time. That data has never existed in structured form before. Perhaps it’s never even been written down. And in a world where AI can actually make sense of it, the companies that capture it first will build something their competitors can’t replicate.

The Voice Stack: Build vs. Buy

When Mike started the voice roadmap two years ago, the impetus wasn’t platforms — it was the fact that vertical voice agents in home services, restaurants, and healthcare were starting to see incredible traction. But when his team dug in, they found something surprising: standing up a production-grade voice infrastructure stack required a 10-plus person engineering team full-time. That’s what made the VAPI thesis so powerful.

If voice isn’t your core product — you’re a ServiceTitan extending into a voice customer service agent, or a bank building an internal training tool — you’re going to reach for a developer platform rather than staff an entire engineering team around infrastructure that isn’t your differentiator. À la Twilio. But if voice is your core product,’ you probably want to control every detail of the stack. Mike sees plenty of voice startups doing exactly that.

Where VAPI has found particular resonance is regulated industries: healthcare, insurance, financial services. These are verticals that need discrete conversational eval gates — validate the customer’s birthday before releasing health information, for instance — with strict governance around every leg of the interaction.

The Trust Gap

The technology bottlenecks in voice have largely been solved. Latency is down to ~500ms end-to-end turns — snappy enough to feel close to a natural conversation. LLMs support tool calling well. The bottleneck now is trust.

Mike shared a telling anecdote: two years ago, a roofing company with a $30K average deal size wouldn’t touch voice agents. Every lead was potentially worth $30,000 — they were going to staff a human every time. Today, that same company would adopt. What changed wasn’t the underlying technology. It was the eval infrastructure. As eval suites get more reliable, businesses extend more autonomy to agents. More autonomy means more workflows handled end-to-end.

There is a simple implication for founders: don’t build a voice agent company unless you can own the customer’s entire journey — not just have a delightful conversation and hand it off to a human. If you can’t fully resolve the query, you might actually be delivering your customer more work. Scheduling is a good example of something low-stakes enough that businesses will fully trust an agent. The question is whether you’re in a use case where you can push into higher-dollar, higher-stakes territory. If the answer is no, or if you can’t articulate how you get there, the step-function scalability that makes voice agents venture-interesting might not be there.

Voice as the New Process Mining

Here’s the thesis that resonated most with how we think about the future of vertical defensibility.

Your system of record — your CRM, your ERP, your EHR — captures the end state. The deal closed. The claim resolved. The migration shipped. It tells you what happened. As Mike put it, it doesn’t capture any of the discussion about how you got there. Voice does.

Consider what that upstream context actually contains. In a sales call, it’s not just what the rep said — it’s how they said it. How much time they spent listening versus talking. How well they read the prospect’s sentiment. In Rilla‘s case, that data is the foundation of a coaching product for home services reps that couldn’t exist without voice. But the insight goes way beyond sales.

Mike pointed to Axiamatic, which raised $54M from Greylock and Bessemer to apply this to enterprise transformations. When a company runs an SAP migration, consultants run dozens of parallel workshops where stakeholders describe how business processes actually work — not how the spec says they should. Traditionally, one person was taking notes. Now every session is transcribed and fed to agents that build a migration plan no single human could assemble, because no single human could sit in on all those workshops simultaneously.

Then there’s Qualitate, an AI-native expert network that uses voice agents to conduct primary research at scale. Mike shared an example of a corporate buyer running an M&A process that commissioned 200 expert network calls over a single weekend. By Monday morning, perfectly compiled, structured intelligence — a process that would have taken an analyst team weeks. Qualitate has now conducted over 350,000 minutes of AI-moderated expert discussions, and every conversation trains the next one.

This maps directly to how we think about moats. In “Dude, Where’s My Moat?” we argued that usage and data loops are the most critical moats at the growth stage. Voice dramatically widens the aperture of what those loops can capture. The keyboard was always a bottleneck — low bandwidth, high friction, lossy. Voice is high bandwidth, low friction, and captures signal that was previously trapped in someone’s head. We explored some of these dynamics in our piece on voice-first playbooks in vertical AI, and Mike’s portfolio is proving the thesis out.

Use Cases That Didn’t Exist Before

What excites Mike most isn’t the customer service replacement stuff — he views that as table stakes at this point. It’s the use cases that flat-out were not possible before voice AI. Recruiting interviews at scale. Sales reps logging mandatory training hours against an AI opponent before ever touching a real customer. Mixed-modality support where an agent watches your cursor navigate an app while talking you through it. Construction inspections where a roofer talks to an app and takes photos simultaneously, stitching together real-world data that was never captured digitally.

And then there’s the ambient layer. Every meeting recorded by Granola. Mike mentioned a CEO who wears an always-on wrist recorder and feeds his entire day into Claude. We’re moving from voice as a channel to voice as a continuous intelligence layer — and the companies that figure out how to capture and compound on that context will own something no system of record ever could.

The Takeaway for Vertical Founders

The potential of Voice AI defensibility is increasingly important: infrastructure is table stakes, distribution (developer community, organic flywheel) is the near-term differentiator, and the long-term moat is the compounding loop of conversations → better evals → more agent autonomy → more conversations. True data gravity.

For founders building voice-first Vertical AI products, the implication is direct. If your product touches a workflow where conversations happen — sales, consulting, intake, support, training, field work — the voice data layer isn’t a feature. It’s the foundation of your long-term defensibility. The system of record captured the scoreboard. Voice captures the game. Or perhaps another analogy is better: in an era where AI can watch the film, the company with the best footage wins.

Subscribe to Verticals to get new episodes every week, available wherever you watch or listen.

Youtube • Spotify • Apple • Amazon

Public market performance, of course, is an imperfect teacher for founders and VCs. After all, public SaaS downturns can have both bad (the application layer is dying) and good (the next generation is taking over fast) implications for startups. Especially in moments of uncertainty, markets can be sentiment-driven — a Buffett voting machine that’s at best distracting to those concerned with 10-20-year cycles.

Yet certain conclusions have been popularized in the wake of the “SaaSpocalypse” that, absorbed without proper context, lead to the wrong vision for the future of the software. This essay is our attempt to make sense of the lessons we can draw from public SaaS movements over the preceding months: dispelling some that unfairly discount the potential for AI-native software, and highlighting others we think are portents for the future of the Vertical AI landscape.

1. Vertical Software Has Stopped Growing

Tom Tunguz’s article from a few months back was a great, concise summary of the recent public market movements. Ultimately, it laid the blame for poor vertical SaaS performance on fundamentals. Despite ostensible value in their moats, he said, they were repriced more aggressively because they are slowing down:

You wouldn’t have to indict vertical SaaS as a business model to believe that story. On average, public vertical SaaS companies are nearly 10 years older than horizontal peers. A thick cohort of legacy, pre-internet vertical players — Dye & Durham (1874), FICO (1956), Agilysys (1963), Tyler (1966) — pulls the 75th percentile vertical age to 42 years old versus 27 for horizontals, which might explain slower growth across the bucket.

There’s just one small problem with the “slower growers” explanation: it’s not true. Our review of 130 public SaaS stocks yielded effectively no difference between median LTM growth: horizontals grew 14.7%, verticals 14.1%. Across our SaaS comp set, fundamentals had almost no correlation with price performance. Revenue growth correlated with returns at r = 0.07; EBITDA margins at -0.03. Within Vertical SaaS, the bottom 15 performers had higher median margins and higher median growth than the top 15.

2. Vertical Software is Less AI-Moated

In his piece, Tom also summed up the popular view well: “Vertical software has fallen 43% this year. DevTools, just 21%. The gap between them… tells you what markets actually believe about AI.” While facially accurate, this takeaway misses a few things.

It implies that the market concluded that vertical software itself — by nature of being industry-specific — was vulnerable to AI. Many have run with that conclusion, surmising that vertical SaaS is easier to replicate with LLMs.

Those conclusions fly in the face of two realities, however:

1. The performance differential had more to do with a handful of horizontal SaaS companies that were “graded” by the market as being not just defensible but poised to benefit from the rise of AI. The following tickers got the “AI picks-and-shovels” treatment, appreciating by >50%: Bandwidth (telecom APIs), Datadog (observability), MongoDB (database), Twilio (comms APIs), Fastly and Akamai (CDN), JFrog (software supply chain), and Innodata (AI training data).

2. Vertical SaaS still trades at a premium to horizontal peers, even accounting for fundamentals. The recent repricing reflects a significant rethinking of that “narrative premium.” In other words, vertical software traded much higher to begin with, so this is perhaps just a case of the bigger they are, the harder they fall. From our perspective, this largely reflects a devaluing of moats that are not immediately visible in the P&L (more on this to come).

None of this suggests that Vertical SaaS is less defensible from AI. It suggests that perhaps the traditional valuation bonuses they fetched are less rich—and that only the most legible, near-term AI tailwinds are being rewarded (more on that below).

3. The Market Has Correctly Repriced Terminal Value

Looking across the whole data set, consumption-based access & pricing models were the clearest driver of software performance over the last several months. We came to this conclusion by blind-scoring all 130 software stocks on the six most important a priori considerations1. Strong correlations should suggest how the market is “grading” AI exposure. Whether a company’s existing product and monetization will thrive in a world where agents are doing more and more work is a very reasonable measure of immediate AI benefit. It’s telling, however, that less tangible moats — workflow depth, proprietary data, regulatory & domain complexity — were not rewarded.

We can even simplify the moat that mattered to one question: Are you agentic infrastructure?

Let’s put that question to the test by comparing Bandwidth (a horizontal that appreciated 280% despite an R402 of 6) and Doximity (a vertical that drew down 65%). If you aren’t familiar with Bandwidth, they’re a Twilio competitor — a CPaaS (Communications Platform as a Service) that sells APIs for voice and SMS used by RingCentral, Zoom, etc. In an AI world, every voice agent that makes a phone call using BAND generates usage-based revenue. As Tom put it, “the data infrastructure tailwind is structural. More AI means more queries, more embeddings, more vector operations.” It’s less immediately clear how Doximity — LinkedIn for physicians, more or less — will benefit from agentic work under the current business model. After all, pharma and hospitals pay subscriptions to market to Doximity’s audience.

Much is lost in this simplistic view of AI moats, however. Doximity, for example, has strong network effects, with 80%+ of doctors on the platform and many hospitals requiring them to sign up. They have data gravity, aggregating proprietary clinical datasets (e.g., PeerCheck, Pathway Medical), which have immediate value to current customers. They have a relatively AI-ready talent, with a 380-person R&D team already layering AI tools (Scribe, DoxGPT) to drive new revenue from health systems. They have workflow loops that deeply integrate with telehealth, faxing, and clinical documentation. And finally, they arguably have regulatory moats from operating all this in a HIPAA-regulated environment.

It’s further evidence that the SaaSpocalypse offered reprieve only to those with immediate revenue opportunity — the potential to benefit from AI in any manner more complicated than picks-and-shovels tomorrow was flatly ignored. Not exactly a deep consideration of terminal value, as one might expect from public markets mostly concerned with next quarter’s earnings.

As Ben Thompson observed in his Stratechery interview: “the destruction and the value creation are not simultaneous in time.” Public markets are pricing the destruction phase. They are pricing the obvious acceleration. They have ignored the longer-term value creation.

4. All Vertical Moats Are Eroding Equally

When we classify the 57 public vertical SaaS players in our dataset by source of defensibility — using the moat taxonomy from Dude, Where’s My Moat? — three groups fall out.

1. Proprietary data moats. Verisk, FICO, Cadence, Veeva, CCC (20 total). These sit on data that cannot be reconstructed or easily replicated. A year ago, they traded at 220% above what a horizontal with the same growth and margins would command. Today: 72%. 18 of 20 are still trading above their horizontal equivalents.

2. Regulatory barriers without compounding data. Tyler Technologies, ADP, Constellation, nCino, Q2 (16 total). Legal and procedural barriers to entry, rather than data-driven ones. Their premium went from 120% to 15%.

3. The vertical halo. ServiceTitan, Par Technology, Toast, Lightspeed, MNTN (15 total). They traded at a 41% premium a year ago on the vertical SaaS narrative — own the vertical, high NRR, extensibility. Today, they trade at a 40% discount to the horizontal equivalent.

Clear and present vertical data moats are still rewarded in the market with elevated multiples (even after accounting for fundamentals). Along with pretty much everything outside of immediate AI revenue tailwinds, however, the recent repricing assigned zero value to even the strongest cases of data gravity.

In an essay three months ago, we shared a simple litmus test on the defensibility of Vertical SaaS: Is the data proprietary? Is there regulatory lock-in? Is the software embedded in the transaction? Two or three yeses and you’re probably fine. The market’s repricing roughly tracks this — proprietary data moats still command a 72% premium. But it’s assigning nearly zero credit for the second and third.

5. The Application Layer Is Dying

The market is pricing the impacts of widespread AI adoption in our economy, including both the decreased cost of software development and the resulting absorption of work by agents. It isn’t yet pricing in the rise of a new generation of AI-native software. Accordingly, pipes that serve AI, generating near-term revenue uplift, get insulated — while software across the board sees erosion, with staying power in question. It certainly isn’t pricing to an application-layer AI steady state, in which data and workflows will be more valuable than ever.

As we shared in our initial essay on the SaaSpocalypse earlier this year, we are still quite early in the destruction phase, despite the market panic.

Some vertical software companies will die, though we think the durability of those names is far higher than that of their horizontal peers. Some of the next generation of vertical AI companies will be built on the rubble, but most of them will be built on green fields. The market, by definition, cannot expand if it’s just ripping and replacing legacy vendors.
— The Verticalist

LLMs have demonstrated the value of scaling for training data and the ability of agents to become more capable through reinforcement learning. To expand beyond language-centric use cases, however, AI will need domain-specific data and decision context. It exists nowhere on the public internet, isn’t for sale, and in many cases, it lives only in the minds of professionals. Vertical platforms have always been best positioned to capture this data. Only now is it becoming so valuable — even if the market hasn’t yet priced it in.

The bears assume AI will shrink the market for vertical software. They're wrong — it will dramatically expand it. AI needs domain-specific data and decision context to move beyond language-centric use cases, and that data exists nowhere on the public internet. It lives in the workflows, transactions, and institutional knowledge that vertical platforms are uniquely positioned to capture. Some incumbents with strong moats and the willingness to cannibalize their own products in favor of an AI angle will thrive. But the biggest winners will be a new generation of AI-native companies — built not just on the rubble of legacy software, but also on greenfield use cases, budgets, and verticals the public market hasn’t even begun to consider.

This Week on Verticals

• A reverse interview with Vignesh & Brendon at Sierra Ventures

• The Claude 14 survivability test

• How consensus capital destroys alpha

• Why Vertical AI is fertile ground for contrarianism

• VC hopes & fears in the era of AI

Brought to you by Parafin — the embedded finance platform behind Amazon, DoorDash, Gusto, WorldPay and many more. Learn more & support the show here!

Watch Full Episode Here

There’s a question we ask nearly every founder we meet these days: it’s 2030-something and Claude 14 just dropped. It ingests any document, integrates with any system, spins up any basic analysis or BI. Why does your startup — as a vertical founder building in the application layer — still exist? Wildly theoretical, of course, but often telling. Founders skinning LLMs and reselling tokens tend to point to hazy future “data moats”… playing right into VC fears that they are building against capabilities that Anthropic, OpenAI, and Google are racing to commoditize.

This defensibility question featured centrally in this week’s conversation on Verticals with Vignesh Ravikumar and Brendon Schmidt of Sierra Ventures— a firm that’s been investing since 1982, with over $2B deployed to date. Vignesh cut to the core of the problem for VCs: with so much demand and so much capital chasing AI, how do investors tease out what’s real? The answer, uncomfortably, is that most of what’s being funded right now isn’t informed by a comprehensive thesis on longevity, but rather a need for speed. A desire to have “AI growth rates” in the portfolio.

Velocity can solve a lot, but it doesn’t solve everything. First-generation Vertical AI wedges (e.g. voice intake, scheduling, document summarization, etc.) work. Customers want them. They generate real results, real dollars. As we argued in Dude, Where’s My Moat?, however, the real question is whether you’re converting that speed into a structural advantage — data gravity, workflow loops, platform lock-in — before those wedges get commoditized and competition nukes your pricing power.

In some categories — especially in popular areas of horizontal AI — the window is closing faster than many think. Not just because LLMs are improving quickly, but also because VC consensus is accelerating. Brendon pointed out that last year, ~41% of all venture dollars concentrated into roughly ten companies. That capital consolidation has a second-order effect: it manufactures consensus. When a handful of funds control a disproportionate share of capital, the entire ecosystem — founders, downstream investors, even LPs — begins optimizing for what those funds want to see. As we explored in The Fundability Trap, the growing conflation of fundability and investability is a structural byproduct of this concentration.

The perennial problem with VC consensus is that it rewrites history. The narrative always gets manufactured after the winner emerges. Which means the broader ecosystem, paying attention to the here and now, will never learn from the dangers of bandwagoning onto a hype cycle. From 2006 to 2011, VCs poured $25B into cleantech — over half of that capital was lost. Around 2015, VCs trying to ride the wave of two truly contrarian on-demand commerce titans (Uber and AirBnB), burned billions more (Homejoy, Washio, Shyp, Luxe, SpoonRocket). Amidst the ~2020 D2C boom, breakouts Casper, Allbirds, Peloton, Blue Apron all grew on copious VC funding, went public, and then collapsed. In 2022, crypto funds raised over $18B, led by a16z ($4.5B) and FTX ($2B) — meanwhile, the vast majority of 2025 tokens are underwater. The story should be familiar: chasing the success of a model that worked in a big way leads to a consensus frenzy, which ultimately destroys massive amounts of capital. Yet because each hype cycle has its own new justification — and no one can ever definitively say where we are in any cycle — there is always an appeal to consensus.

So what does this mean for ambitious founders trying to answer the Claude 14 question? It means the answer probably isn’t do whatever is popular and working right now. And it certainly isn’t try to provide the same value that OpenAI or Anthropic are. The answer isn’t obvious. It probably requires the kind of earned domain insight that a three-month market scan can’t replicate. And it demands a path to defensibility that goes beyond “we’ll move fast” — that you need to do regardless.

Investors in AI face a different version of the same question. Do you back what everyone else is and jump on the consensus bandwagon? Or do you ignore the skepticism and FOMO that accompanies contrarianism? History would suggest being ahead of consensus is the only true path to VC alpha — while it pays to be a passenger on a beta train, hopping on when it’s already reached full speed is a dangerous game.

This week, we switched up the format. Vignesh & Brendon interview Nic on the above, and much more. Check out the full episode on Youtube, or wherever else you watch and listen. See you next week!

This episode was originally filmed as the debut of the DPI Podcast (Deconstructing Pre-seed Investing) — a new series from Sierra Ventures co-hosted by Vignesh Ravikumar and Brendon Schmidt that brings on top pre-seed investors to break down how they identify breakout companies, what themes they’re excited about, and how founders can find the right partners to build generational businesses. Check out the DPI Pod on Spotify, Apple, or YouTube. Huge thanks to Vignesh, Brendon & the Sierra team for having us on!

Subscribe to Verticals to get new episodes every week, available wherever you watch or listen.

Youtube • Spotify • Apple • Amazon

This week on Verticals:

• Why 99% accuracy is the same as wrong in high-stakes industries

• The “exoskeleton” model — AI suggests, humans stamp, clients trust

• Selling into CapEx budgets and why it changes everything

• Is regulation actually a moat?

• Monetizing via share of upside vs. work-to-be-done

Check out the full episode on YouTube.

There’s a popular assumption circulating in AI right now: that the remaining accuracy gap between foundation models and human experts is closing so fast that it barely matters. In consumer or SMB applications, that may be true. But in verticals where a single error can kill a $100M deal (or worse) — energy, infrastructure, defense structured finance — 99% accuracy is functionally identical to wrong. The last 1% is the entire business.

Jacob Sandry knows the cost of that last mile personally. Before co-founding Euclid Power, he helped build Goldman Sachs’ renewables investment fund from zero to $3B deployed. At 26, he was walking into Goldman’s investment committee recommending $500M commitments on solar portfolios. “If I had a mistake in my model or my presentation and said, ‘sorry, Claude did that’ — immediately fired,” he told us. There is no margin for error when you are underwriting 6% unlevered returns with no recourse — meaning the trust threshold for AI is extremely high.

Today on Verticals, Jacob shares how he’s solving that problem for his customers as CEO of Euclid Power through a mix of AI, services, expert approvals, and outcome guarantees — and how why earns them significantly more pricing power.

Watch the full episode here

This episode is brought to you by Parafin: the embedded capital infrastructure behind Amazon, DoorDash, Gusto and many more. If you have SMB customers, learn more about how Parafin can grow your retention, revenue, and TAM today — click here or below.

The 1% problem is an operating reality for a growing type of Vertical AI startup building in what we’d call high-consequence workflows: energy project finance, clinical decision support, construction compliance, legal diligence. In these markets, the buyer doesn’t just want speed — they want someone accountable for the answer. And 100% accountability is something an LLM structurally cannot provide today. Training data lags, hallucinations, inconsistencies can all result in an imperfect outcome that is effectively worthless — if a customer has to review results for even a small mistake, they might as well handle it in-house.

Euclid Power’s response is what Sandry calls an exoskeleton model. AI handles the extraction, pattern-matching, and first-pass analysis across thousands of documents in a renewable energy data room — interconnection studies, land agreements, environmental reports, engineering specs. But a domain expert reviews every output, validates the conclusions, and stamps the deliverable. The client gets a diligence report they can take to their investment committee, not a chat interface they have to babysit.

Euclid also just acquired Thresh, an AI-native competitor whose pitch was compelling: go from data room to IC memo automatically. The technology was strong. Customers liked it. But the market was limited because when the AI still made mistakes, users had to redo the work anyway. “There’s some sort of cap on the market for it,” Sandry said. “You’d be like, well, this still gets things wrong. So then I still have to go back and do the work.” The AI-only product hit a ceiling not of capability, but of trust.

This has implications well beyond energy. As we explored in Dude, Where’s My Moat?, brand and trust are among the most durable moats at scale for vertical AI companies. In high-stakes verticals, trust isn’t a nice-to-have bolted on after product-market fit — it’s the product itself. The companies that figure out how to deliver AI-powered speed with human-grade accountability have an opportunity to play a different role with customers. Instead of being an OpEx line item, you can earn the economics of advisors and consultants who already charge on percentages of deals or longer-term profits. In industries like renewables, half of project spend goes to admin, legal, and compliance — but the big prize still sits in the long-term appreciation of a project. The potential is an order of magnitude in relative ACV.

The AI services debate often gets framed as a transitional phase — a necessary evil until models get good enough. Jacob’s experience suggests it may be more of a long-term feature than a bug. In verticals where the last 1% is everything, the human in the loop might turn out to be a moat. Watch the full episode to hear why.

Subscribe to Verticals to get new episodes every week, available wherever you watch or listen.

Youtube • Spotify • Apple • Amazon