OpenAI and Epic Just Killed a Thousand Healthtech Startups

On September 1, OpenAI announced that hospitals can now plug their Epic systems into ChatGPT. A doctor can ask what’s changed since the patient’s last visit, which labs to look at before the appointment, and whether anyone added a new medication or sent the patient to a specialist. The same release shipped a Healthcare Public Data plugin that connects to nine official sources, including PubMed, DailyMed, CMS Coverage, and ClinicalTrials.gov. The launch partners are UCSF, Cedars-Sinai, HCA, Memorial Sloan Kettering, Boston Children’s, Baylor Scott and White, and AdventHealth.

For the last two years, this exact product has been the most common healthcare AI pitch in my inbox. Doctors spend too much time digging through the chart, so a startup connects to Epic, reads the record with a language model, and hands the doctor a summary before the visit. Dozens of funded companies were built on that one idea, and hundreds more were trying to get funded on it. Every one of them had to do the same slow work of getting Epic access and getting a hospital to say yes. OpenAI skipped all of that. It went straight to Epic and to seven of the biggest health systems in the country, and it shipped the summary as a feature of a product those hospitals already pay for.

Reading the patient’s chart and telling the doctor what’s in it is now a feature. Pre-visit summaries, medication checks, handoff notes, the three things you need to know before you walk into the room. Two days ago that was a company. It had conferences, a Slack group, and a set of valuation multiples that people quoted at each other. Today it’s a checkbox that a workspace admin turns on, and the integration work that used to be the hard part is a support article. The cost of turning it on just went to roughly zero, and anything whose whole reason for existing sat above that cost has a problem.

Nobody Built Anything

OpenAI didn’t invent anything here, and I think that’s the whole story. There’s no new model in the announcement. There’s no breakthrough in clinical reasoning and no clever new way to search records. It’s an integration plus a set of connectors to public datasets that have had open APIs for years. Any competent five-person team could build those connectors, and several hundred five-person teams did. What none of them could build was the sentence “healthcare organizations can now connect Epic environments to ChatGPT.”

That sentence is a business development call that got returned.

I’ve been using reconstruction time as my test for proprietary data claims. If a competitor wanted the same data you have, how long would it take them to build it? I still like the question, but it has a blind spot. It only measures technical distance. It says nothing about who can get a meeting. Whatever the reconstruction time was on these Epic integrations, it never got tested, because the gap closed from a direction the test doesn’t look at.

What these companies were really selling, mostly without saying it out loud, was distribution. For about two years the model vendors and the systems of record had no reason to talk to each other, so somebody had to do the work of connecting them. That meant writing the FHIR plumbing, which is the code that pulls patient records out of Epic in the standard format hospitals use to exchange data (FHIR stands for Fast Healthcare Interoperability Resources) and pushes results back in. It meant sitting through the hospital’s security review and surviving procurement. And it meant signing a BAA, a Business Associate Agreement, which is the contract HIPAA requires before any vendor is allowed to touch patient data and which makes that vendor legally responsible for protecting it. All of that was real work and it deserved to get paid. It was also a bet that two large companies would keep not cooperating, and that kind of bet only pays until they do. Once OpenAI and Epic decided it was worth picking up the phone, the two years of plumbing stopped being a moat and became a support article.

There’s no Epic quote in the announcement. The testimonials are from UCSF Health’s CEO and AdventHealth’s Chief AI Officer, both of them customers. The company that owns the actual distribution said nothing. That reads to me like a deal where Epic is happy to rent the same lane to anybody who asks, which is worse for the startups than an exclusive deal would have been. An exclusive at least tells you the door is shut. Open access to the door means the door was never the asset.

Healthcare is just the vivid version, since the regulatory moat was supposed to be the deep one. A week earlier, Salesforce and Anthropic announced Claudeforce, launching with 37 prebuilt sales skills covering meeting prep, deal health, and pipeline review. Same shape. The model vendor and the system of record shake hands, and the middleware sitting between them stops having a reason to exist. That’s the SaaS-pocalypse thesis, delivered by the incumbent who captured it instead of the startup that expected to.

Who Owns the 0.9 Percent

Sitting in the middle of the OpenAI release is the number that tells you where the money actually ends up. Physicians reviewed ChatGPT’s answers across 27 clinical use cases. Out of 4,363 ratings, 99.1% came back rated safe.

So about 39 didn’t.

As an eval result, that’s strong, and I don’t want to be cheap about it. As a business fact, it raises a question, which is who owns those 39. The release answers that question carefully and at length. Governed workspace. Signed BAA. Role-based access, single sign-on, audit logs. ChatGPT summarizes the chart and points back to the source. The doctor reviews the output. The doctor signs the note.

So the liability sits exactly where it already sat, with the doctor. That’s the only version of this product that ships, and I’d have built it the same way. The tool makes it faster and easier for a physician to get to a judgment, and the physician still has to make the judgment. It also tells you who survives, and it gives you a better test than the one most people are using. Forget proprietary data for a second and ask whether you’re standing somewhere the platform has decided not to stand.

The platform declines to stand in two places. The first is the regulated liability itself. Cleared decision support. Ambient documentation where the generated note goes into the record as the record. Coding and billing work where being wrong means a clawback or an audit, and you’re the one who eats it. If your product can be wrong and somebody other than you pays for it, you were a feature.

The second is a workflow the platform can’t reach from the outside. “We integrate with Epic” is now a plugin, so that one’s gone. What’s left is the last mile into the payer, the lab, the device, the state registry, the county system that still runs on a fax machine. Those integrations are ugly and specific, and nobody at a frontier lab wants to own them, which is the whole point. Ugly and unwanted is a business.

What doesn’t survive is a folder of markdown files and a custom RAG pipeline behind a chat window. That was a viable company for about eighteen months… and to be fair, it was a good eighteen months.

Why I’m Still Not Writing AI Tooling Checks

People ask why I’m not writing more AI checks, so here’s the actual reason.

Tooling doesn’t settle into a stable shape until the thing underneath it stops moving. Right now the models get better on something like a quarterly cadence, and every meaningful improvement changes what the product wrapped around the model should look like. Context windows grew, and a whole generation of chunking strategies stopped mattering. Tool use arrived and did the same thing to orchestration frameworks. Now native connectors have arrived, which is the memo you’re reading. Every layer of tooling built during the climb is a bet that the climb stops right where the builder happens to be standing.

If you’re building AI tooling today, my expectation is that you rebuild your workflow roughly every six months, and that most of what you ship is temporary structure holding a spot until the model absorbs it. That’s how the railroad equipment business worked before the track gauge got standardized, and plenty of people made money in it. It’s just a bad thing to underwrite at a price that assumes it lasts.

The Phone Call Was the Product

I’ll grant one thing to the companies I just declared dead. Enterprise healthcare deployment is where fast things go to die, and eighteen months from now these Epic rollouts could easily be stuck in security review, change management, and the part where you convince a fifty-eight-year-old cardiologist to change how she preps for clinic. An announcement about distribution is a roadmap, and roadmaps slip. If this one slips badly enough, the startups get handed free top-of-funnel by the biggest AI company in the world, and the value moves to whoever does the last-mile implementation work. I’m watching for that. I’m not betting on it.

Because even in that world, the thing that got announced on September 1 is still true. Nobody built anything. The connectors were always easy, the public data was always public, and the summary was always going to become a feature the moment the model vendor and the system of record found it convenient to shake hands. The startups in the middle were selling a phone call they couldn’t make, and now the phone call has been made. Whatever the rollout timeline turns out to be, the price of “we read the chart and tell the doctor what’s in it” has been set, and it’s zero.

What’s left to pay for is the part the platform won’t touch. Somebody still has to sign the note and take the blame when it’s wrong, and somebody still has to build the ugly connections into the payers, labs, and county systems that Epic doesn’t reach. Those are the only two places in this market where being wrong costs you something, and that’s exactly why they’re still worth money. Everything else in the stack, the chart summary and the connectors and the chat window on top, is now a feature bundled into two subscriptions the hospital already pays for, Epic and ChatGPT. There’s no third invoice left for a startup to send.