If a model can rebuild my product over a weekend, what am I worth? Well, hopefully the presumption here isn’t true in your case. But general sense of angst increasingly competent frontier models provokes in founders and investors alike is real. Historically common answers to defensibility — switching costs, painful integrations, platform effects — are in question. But in this era, you need to look beyond protecting the revenue you already have, to moats that win you the next customer. In a market where velocity is paramount and code can be cloned, the industry network effects that drive referrals, bookings, expansion have never been more important. Our guest today argues it might be all that matters in the coming era of AI.
The Takeaway
Switching costs, hardware and integrations protect the revenue you have. They don’t win the next customer, and AI is making migrations cheaper.
Run the codebase test every quarter: if a competitor had your entire codebase tomorrow, what would they be worth versus you?
Lean into what makes you more valuable to customer N+1 vs. the clone: network density, data on rare events, purchasing scale.
Pick verticals where 20 years of experience beats 2, where someone has to sign off on the work, or where scale wins the next customer. An aging incumbent alone now draws a wave of accelerator-funded startups.
Early-stage companies can still steer toward a network or data advantage. Later-stage ones may need to acquire whoever owns it in their vertical.
The Episode
Coyne Lloyd was on the founding team at Snapdocs, the mortgage closing platform that now supports roughly 1 in 4 U.S. mortgage transactions and raised a $150M Series D at a valuation over $1.5B in 2021. He went on to become a partner at Fractal Software and has since helped incubate Layer (embedded accounting and bookkeeping for SMB platforms) and Reach (embedded marketing tools), both sold through Vertical SaaS platforms. He’s also an angel in companies like Moxie and Vesta. He’ll share his view on why the AI copy threat is real, and why the defensibility that’s left looks a lot like what made Snapdocs work long before LLMs.
The Codebase Test
On February 3, Anthropic’s Claude Cowork legal plugins helped knock $285B off software stocks in a day — Thomson Reuters specifically fell 18%. Since then, it’s become clear the clear and present danger of immediate SaaS replacement was a bit hyperbolized. In fact, since late March, software stocks have largely recovered (as we covered in SaaSpocalypse Now What?) and median net revenue retention has held. Still, replacement fears loom. Bain now claims to vibe-code replicas of software acquisition targets during diligence to test how easily the core product can be copied.
In our episode today, Coyne shared an example. The founder of a multi-billion dollar Vertical SaaS company was negotiating to buy a smaller player in his space, and the seller kept dragging out the deal for a better price. Then a big model release landed, and the buyer “just sat down over the weekend and coded up a pretty passable version of the acquired app.” He went back to the table with a lower number. The customer base was still worth something to him. The software was worth a lot less.
So what’s worth more? Some, convinced software itself is the problem, point to hardware. Hardware, as Coyne put it, “is a moat in the sense that you will retain your cash flows. It is not going to be easy for someone to rip you out. But it is not a moat in the sense that it is an accumulating advantage.” This isn’t 100% true, if you include a handful of historic B2C hardware stories, like the iPhone — Apple truly did turn design and manufacturing into a snowballing edge, even if the software was important too and its more important eventual edge was a network effect. B2B hardware, however, rarely has the potential for such user-driven breakouts. (We covered the bull case for B2B hardware in Hardware as an AI Moat.)
Coyne doesn’t think thorny integrations are a moat for long, since computer-use agents can work around a closed API. Similarly, customer data migrations — long one of the top three headaches for any vertical CEO — are getting easier (a point we made in Reducing Switching Costs with AI). If migration gets cheap, switching costs shrink with it.
In AI, the promise of data moats is clear. Getting there can be trickier. Even when you do have sufficient data, you have to decide what to do with it. Luke mentioned he’s had several founders call him about seven-figure offers for their datasets. Google paid $10M for Spirit Airlines’ data out of bankruptcy, and SimpleClosure has run more than 150 data-sale deals for shut-down startups. This is an opportunity for Vertical AI startups that do have a product that can clearly benefit from a reinforcement learning loop, or similar. We believe the best platforms, however, are creating data assets that don’t exist, because if you can buy it, so can someone else. Moreover, currently purchasable data may be a deflationary asset, as sample efficiency1 improves.
In his 7 Powers, Hamilton Helmer holds that every durable advantage needs both a benefit and a barrier. Switching costs, for example, “hold no Benefit with potential customers.” Coyne posed a simple thought exercise for founders to gauge their benefits and barriers in this AI era, with software easier than ever to replicate. Here’s the Codebase Test:"
Think about if your competitors had a complete copy of your codebase, but they didn’t have your data or the network or scale of people that are on your thing, what would your competitors be worth versus you?
Case Study: Attack of the Clones
Snapdocs — where Coyne was an early employee — offers a good case study. The original product coordinated remote mortgage closings: the workflow of getting a notary or attorney to a borrower, with the right packet, at the right time. Built pre-ChatGPT, it was initially only “marginally better” than the software that pre-existing title and escrow offices already used.
Then, Snapdocs built a network around their key user, the notary. They enabled notaries to manage work from every title company and lender in one place, and gave them professional profiles. Soon, notaries spent more time on Snapdocs than any other platform. It naturally evolved into a marketplace. Today, Snapdocs counts 140K+ notaries, 130+ lenders and 37K+ title branches, and more than 8M cumulative signings. If we apply the codebase test: a clone starts with zero notaries, and hence zero appeal to the lenders who only care about that next close in rural Montana tomorrow.
Next, Snapdocs leaned into its data asset. As Coyne described it, closings are often “held hostage.” A notary agrees to $200, gets the documents signed, then claims the price was $1,000, hours before a deadline that cancels the whole transaction if missed. It’s relatively rare — but that’s the point. “If you don’t have visibility into just about every notary and attorney in the country, you don’t know who the bad actors are.” Each closing made the risk model better for the next customer. Coyne says Snapdocs is able to charge 5-10x competitors per closing, purely for the certainty that it wouldn’t get held up. “You could have copied everything about Snapdocs’s codebase and we still would have been a many multiples premium.”
One of the greatest assets of Vertical AI businesses is their ability to double down on industry networks to create network and data assets beyond the code. Procore, for example, priced on unlimited users so general contractors would invite subs and owners — a tightly interwoven multi-party stakeholder ecosystem. By 2019, the average customer brought 170+ outside participants onto the platform, and more than 60% of its 1.3M users were collaborators. Moxie, a med spa platform Coyne invested in, is the largest buyer of injectables in the country thanks to its demand base. That economy of scale allows them to offer procurement pricing to its customers that offsets a healthy portion of what it charges practices. Ideal assets are both barriers and benefits. Founders can ask the simple question: does it make us more valuable to customer N+1 than a code clone would be?
Picking a Defensible Vertical
Coyne shared a view of a fellow incubator that picking verticals has become harder:
7 years ago, you could basically just go to any vertical, see some really legacy incumbent that Constellation… had purchased, and go build something newer, cleaner, with maybe a mobile interface also. And you would be able to take the market… Are there going to be 3 companies in every one of the next 4 Y Combinator batches that are hitting just as hard and have all raised $10M and are bidding you down?
We see the crowding in our own pipeline: in nearly 20% of startups we’ve met at Euclid this year, we’d already met 3+ with near-identical product visions.
So what makes a vertical worth picking today? Coyne’s filter is to ask “what are the fields where someone with 20 years of experience is better than someone with 2 years of experience?” Those are the fields where edge cases pile up, and where years of workflow data teach a model things a clone can’t learn from a manual, close to what Helmer calls process power.
Verticals that require a human to sign off on the work have an inherent advantage over merely complex ones. This concept of liability can be formal (you need a licensed architect to validate a plan by law) or informal (a vertical startup would rather not own customers’ books because getting it wrong has legal consequences). Coyne credited the latter as a big reason vertical embedded startups have been working, despite the “you don’t own the customer” critique. We covered it extensively in Why Your Voice AI Agent Isn’t Enough. An agent can grind through complexity — it can’t (yet) take on liability in the way a human can, and often must.
The Takeaway for Vertical Founders
Whether you’ll keep winning is about the next customer. And compounding moats, increasingly, are found outside the code itself. The Codebase test is designed to probe what assets you’re actually building: assume a competitor has your entire codebase tomorrow, and list what still makes you the better choice for the next customer.
For Snapdocs it was a network of notaries and the data on each that won them trust. For Moxie it’s buying power. You have time to steer an early company toward something that compounds: a marketplace inside the workflow, data on events your competitors can’t see or purchase, etc. But you need a plan for both barrier and benefits, and investors want to understand it earlier. So ask yourself: what will we have beyond the code that wins customer N+1.
See you next week.
Key Moments from this Episode
00:00 — Intro
03:07 — Why Vertical SaaS founders are panicking about AI
05:12 — How Snapdocs charged 5-10x more than competitors
08:43 — The thought experiment every SaaS founder needs
09:29 — Why hardware lock-in is an overrated moat
13:14 — How Snapdocs built its marketplace and data moat
19:09 — Can proprietary AI workflows create process power?
21:56 — Why your biggest AI threat might not be OpenAI
29:32 — Should SaaS founders sell their proprietary data?
32:33 — Will AI destroy embedded software businesses?
38:58 — Can startups beat incumbent Vertical SaaS companies?
45:49 — Why software graveyards are ideal rollup markets
49:24 — What Coyne got wrong about AI rollups
52:14 — Should Vertical SaaS price on outcomes?
54:48 — The private vertical software company Coyne is bullish on
Sample efficiency is how much data a model needs to learn a task. It’s improving over time. Meta’s LIMA (2023) was rated as good as or better than GPT-4 in 43% of human comparisons after fine-tuning on just 1,000 curated examples. In 2025, s1 and LIMO reached competition-level math reasoning with 1,000 and 817 training examples, respectively (LIMO outperformed models trained on 100x more data).