On September 23, we hosted the second annual Verticalist Summit in San Francisco (a half-day, invite-only event for founders, operators, and investors building Vertical AI). Over the past two weeks, we’ve shared the first two sessions from the conference: Nic Poulos from Euclid on The State of Vertical AI, and our go-to-market panel, Scaling Vertical AI, on how to scale beyond founder-led sales and why you’re probably hiring too many reps.
Today we’re sharing our AI in Healthcare panel: where AI is already delivering value in healthcare, what the labs will and won’t take, and why adoption is so slow but so sticky. Charles Hudson (Precursor Ventures) moderated with panelists Christophe Rimann (Camber), Mike Kopko (Pearl Health), and Daniel Kador (OpenAI). Here are three moments you shouldn’t miss:
4:17 — Why nobody chases the last $6 on a $100 claim, and why an agent should
6:56 — OpenAI’s B2B healthcare lead on why “use chat, good luck” doesn’t cut it
11:55 — The RFPs that never get answered
See our write-up below for a breakdown of the five key learnings that every Vertical AI founder should take away from the panel:
AI makes the long tail worth collecting
Give back cash and time, and you get more care
Even the labs say workflow wins
Your buyer may be the job you’re replacing
Clinical AI is already here, with a human in the loop
And in the coming weeks, stay tuned for two more sessions from the Summit:
Nov 24: AI-Native Marketplaces.
Moderated by Mike Duboe (Greylock) with Matt Greenberg (Handshake), Ludo Antonov (SuperMe) and Andrew Silard (OpenRouter).Dec 1: The Investor View on Vertical AI.
How investors are underwriting Vertical AI, moderated by Omar Elayat (Euclid) with Dave Yuan (Tidemark), Brian Feinstein (Bessemer), Hunter Walk (Homebrew) and Nikhil Basu Trivedi (Footwork).
Thanks to our sponsors for making the day happen: Stifel, Latham & Watkins, DialStack, Founder Shield, and Tight.com.
Why Healthcare, Why Now?
The US spends around $5T a year on healthcare, and that number keeps climbing. Employer healthcare costs are up around 10% a year. As Pearl Health CEO Mike Kopko put it, almost nothing else in your life grows that fast — not even your investment accounts.
You wouldn’t pick healthcare as one of the first domains for AI to get into. It’s a regulated industry, it’s slow, and the stakes are life and death. Yet Dan Kador argues the crisis is precisely what’s created the opportunity. “The magnitude of the spend means there’s enough margin to go around for people to make money.”
Charles Hudson of Precursor Ventures moderated. Christophe Rimann is CEO of Camber, which uses AI to help providers get paid by insurance. Mike Kopko is CEO of Pearl Health, which manages about $6B of Medicare spend on roughly half a million seniors. Dan Kador is head of healthcare engineering at OpenAI and previously led engineering at Abridge. He introduced himself as the panel’s boogeyman.
Top Five Takeaways from the Panel
Five takeaways that stood out to us from the panel:
1. AI makes the long tail worth collecting
Revenue cycle management is a $250B to $500B-a-year market, and it’s predominantly labor spend (much of it outsourced to BPOs). For a hospital chasing a $100K claim, 95% collection still leaves you with real money, so it’s worth having someone on the final 5%.
The long tail is quite different. You’ve got thousands of smaller providers who are billing the $100 claims and have become very accustomed to collecting 94% of each claim. In most cases, it’s not economical to have a human go after the last 6%. Christophe noted that in many instances it’s actually a negative ROI.
An agent will change that math. It can afford to go and call a payer for $5. “Every last dollar actually makes sense to go catch,” as it all hits margin. For a small practice, that gap between 90% and 95% collection is the difference between staying open and going bankrupt.
That’s the part we think most founders miss. Most healthcare AI pitches are about reducing the cost of admin. Camber’s is an uplift pitch: revenue that was never worth a human’s time to collect.
2. Give back cash and time, and you get more care
Charles asked where all the AI savings go. Christophe’s answer started with working capital. Payers take 60 days or more to pay most providers. Camber’s customers get that down to about 20 days.
That’s important because demand is far outstripping supply. Camber started in the autism space where Christophe tells us demand for autism care has gone up by approximately 4,000% over the last five years. Providers simply can’t scale fast enough when their cash is tied up with the payers.
Time is the other constraint. Providers spend about a quarter of their time on admin, and it takes five to ten years to train a new one. In his view, giving that 25% back to existing providers is by far the biggest lever for increasing care in the US.
Dan saw the exact same thing at Abridge. Clinicians told Abridge it had saved their marriages and let them see their children again. Some had been contemplating leaving and now felt they could actually do it for another five years.
Mike was more dubious about where the gains actually land. “Everybody wins except those who don’t adopt new technologies,” he said. Affordability isn’t really improving, and he sees it as much a policy issue as a technology issue.
3. Even the labs say workflow wins
This was the question from Who Wins the AI App Layer? (asked of the lab itself): which healthcare problems will model providers take on, and which will remain with applications?
Dan’s answer was straightforward. The labs are going down the vertical path (that’s literally what his job at OpenAI is) because a purely horizontal approach won’t work. Some of what the health systems and payers do will be handed over to the models, where you just need smarts. Healthcare workers, however, live in workflow. “You can’t just be like, ‘Just use chat, good luck.’”
An audience member asked what happens when the token subsidies end. Dan thinks the subsidies are overstated in the media and that labs have every incentive to keep models as cheap as possible. Open weights will keep putting pressure on prices. His advice to founders was to pick the model that sits on the cost/performance curve for that particular use case.
Christophe added a healthcare-specific case for open-weight models. Foundation models have seen far less healthcare data, so hallucination rates are high. Camber spends a massive amount of time on validation.
4. Your buyer may be the job you’re replacing
Mike’s view, from his days building networks at Oscar before Pearl, is that the interoperability layer isn’t ready for what AI can already do. The health systems simply aren’t ready to put their brand on the line. Integrations take months and years, not days. “Anyone who tells me it’s easy, I don’t believe them.”
Then there’s the org chart. Healthcare is one of the biggest employers in the economy, and some of those people work in the departments that actually buy the software. Mike said he can’t possibly count how many competitive offers never get surfaced because they go through a group that knows the product could end its division. Some RFPs just never get answered. Plan your sales timeline around that.
Trust takes a long time too. Mike sees three layers. The certifications (SOC 2, HIPAA, HITRUST, etc.) are the baseline; they’re a real cost of doing business. A breach is game over, so security has to be taken seriously from day one. The human layer is much easier with people who’ve been in healthcare before, as data errors here aren’t just ‘forgotten’ like they are in a social app.
He also had a very strong view: HIPAA is probably due for a rethink. If you loosened it up and, in the process, got more data exchange and saved 100,000 lives, would people accept that trade-off? He’d like to have that conversation.
The upside of all that friction is durability. Innovation takes a long time to get absorbed into healthcare, according to Mike, but once it is, the scale, permanence, and returns are huge.
5. Clinical AI is already here, with a human in the loop
The final audience question was about when AI moves from being used on the admin side to actually doing the doctoring. Mike drew the self-driving car analogy: one fatality in a self-driving car stops everything, whereas human drivers kill people by the minute. He believes the technology is pretty much there in terms of clinical AI, and that most clinical AI applications already beat doctors on standard questions and standard diagnoses. Absorption is the bottleneck, of course; in his insurance business, state-by-state rules hold up every new application. His advice was to go for an outcomes mindset, because guaranteeing good outcomes at scale is the hardest thing in healthcare.
Dan added that clinical AI is already widely used, almost invariably with a clinician in the loop. His example is from a major health system in Israel. Patients applying for disability benefits following a period of mental health hardship are typically required to have a psychiatrist or psychologist make a determination. They can now opt to speak with an AI agent instead, which will review their experience and make a recommendation. If they’re not happy with the outcome, they can always go on the waitlist for a human, which is a month or two out.
The patient chooses. That’s a design pattern well worth stealing.
Charles closed by asking what’s preventing the adoption of AI in healthcare. The answers were IT governance, the EHRs, and the sheer complexity of the environment.
What to Do With This
This, of course, doesn’t mean cost savings aren’t important. Pearl Health’s entire model is about reducing the cost of care, and Mike’s point is absolutely correct: affordability is the bigger issue, and it’s as much a policy issue as it is a technology issue. For most Vertical AI founders selling into healthcare, though, the panel is very clear. The easiest sale is simply revenue you’ve never been collecting. The hardest is a product that replaces the buyer’s own team.
Here’s a quick heuristic for Vertical AI founders. List the work in your vertical that’s never worth a human’s time (the $6 balance, the follow-up no one does, the claim that’s too small to even warrant a response). Price each of those at agent cost, not labor cost. Wherever the ROI turns positive, you’re selling new revenue rather than cost savings, and that’s an easier conversation to have with a buyer concerned about their own department.
Finally, if you’re building clinical or patient-facing AI, then make the agent optional and keep the human path open. Let the patient make the choice and let adoption speak for itself before you ask a health system to put its brand on the line.

