OpenAI Just Put ChatGPT Inside Epic. Hospitals Are Already Paying for Two Other AI Tools That Do the Same Job.

On September 1, OpenAI announced that healthcare organizations can connect their Epic systems to ChatGPT for Healthcare. A clinician with the right permissions can pull appointment notes, lab results, medications and specialist documentation into a chat window and ask what changed since the last visit, which labs to check before today's appointment, or which follow-ups never got closed. In some setups ChatGPT sits inside the Epic interface itself, so the doctor never leaves the chart. OpenAI shipped a plugin the same day that reaches nine official data sources including PubMed, ClinicalTrials.gov, DailyMed, RxNorm and CMS Coverage. TechCrunch, Fierce Healthcare and Becker's all had it up within hours.

So did the take. A thousand health tech startups just died. Every company pitching "AI that plugs into your EHR and saves clinicians time" has been steamrolled by the biggest name in AI working with the biggest name in medical records.

Except that this already happened, in February, and OpenAI wasn't involved.

Epic already did this in February

Epic released AI Charting in February, a native feature that listens during the visit and drafts the note and suggested orders. HIT Consultant and HLTH both read it the same way at the time. Health systems could now get a built-in option that cost less, was easier to govern, and kept patient audio out of an outside cloud, which meant standalone vendors would have to be dramatically better than Epic and not just a little better. STAT and MedCity News landed in the same place.

The usage figures are what make it stick. By February, Healthcare Dive reported, the chart summary feature inside Art, Epic's clinician-facing AI, was running more than 16 million times a month, roughly triple its November volume. Penny, the revenue cycle tool, was live at more than 200 organizations. Emmie was answering patient messages in MyChart. Epic told Fierce Healthcare that 85 percent of its customers were already using generative AI across the three products.

Sixteen million chart summaries a month is the thing OpenAI supposedly took over this week. Epic had it in the winter, running across a customer base that covered more than 42 percent of the acute care hospital market and close to 55 percent of acute care beds as of KLAS's 2024 count, reported by Healthcare Dive.

The integration is read-only, which limits what it threatens

One detail got very little pickup. ChatGPT can read the record but cannot write to it, so nothing it produces stays behind once the clinician closes the window. OpenAI says so directly, and TechCrunch confirmed it.

That matters for working out who was actually threatened. Summarizing a chart is now available from at least three vendors, one of which the hospital already owns. Acting on a chart is a different business. Ambient documentation that drafts a note a physician signs, coding software that produces a claim, prior authorization tools that generate a submission: all of that requires writing to the record, which is where the liability sits, and none of it changed on September 1. OpenAI's own post describes the new features as designed to complement existing EHR workflows.

If you are building or funding in this category, the useful question is whether your product reads or acts. Read-only products are now up against a feature the customer already pays for. Products that act still have a business, though a harder one than they had in 2024.

What this looks like to the person writing the checks

A hospital CIO woke up this morning already paying Epic for AI that summarizes charts, with doctors using it. Many are also partway through a contract with an ambient documentation vendor like Abridge, Ambience, Nabla or Suki, all of which spent two years building pre-visit prep and chart synthesis into their products to justify prices that can reach seven figures a year. Investors put close to $1.6 billion into ambient AI companies in 2025 through the end of September, according to PitchBook data cited by Fierce Healthcare, and a lot of that went into feature expansion. Now a third vendor offers to read the same chart, from inside the same screen, as part of an enterprise agreement the organization may already hold for other reasons.

No hospital is going to run three chart readers. RamaOnHealthcare flagged the overlap on day one, calling it a third AI layer stacked on the same chart with CIOs left to sort out which one keeps the job.

The criteria are already published. When Epic shipped AI Charting, HIT Consultant and HLTH laid out why health systems would revisit outside contracts: ambient deals can run into millions a year, a native tool means patient audio never travels to a third-party cloud, and every additional vendor is another contract and another integration to maintain. Each of those arguments gets stronger when the vendor under review is the third one rather than the second.

The startups saw this coming in February. Heidi CEO Tom Kelly told MedCity News that his company was built as an AI partner for clinicians rather than a feature living inside an EHR, and that visit notes are one piece of a broader set of workflows it supports. Abridge did not respond to the same request for comment. Kelly is making the only argument available to him. It concedes the summarization fight and moves the fight somewhere else, which is a smaller claim than $1.6 billion in funding assumed.

Epic said nothing

Before anyone writes that OpenAI has partnered with Epic, look at who is quoted.

Epic is not. The testimonials in OpenAI's announcement come from Suresh Gunasekaran, CEO of UCSF Health, who described his organization as a pilot partner still exploring what the integration can do, and from AdventHealth's chief AI officer, whose comments are about ChatGPT Work and Codex rather than the medical record. No statement from Epic appeared anywhere in day-one coverage.

That is not evidence of a rift, but it does point to a distinction. Epic runs an open integration program, and by its own account in April, more than 1,000 apps are listed in its Showroom marketplace, up by over 200 in a year, with patient-directed data exchanges through Epic's APIs running around 3.7 billion a month. Building on those rails is a well-worn path and does not require a deal. "OpenAI struck a partnership with Epic" and "OpenAI became app number 1,001" are very different stories, and only the second is consistent with what has been published. Until Epic describes the arrangement itself, the first should not be repeated as fact.

Either way, Epic owns the data, the interface, the marketplace and the customer relationship. Every foundation model company needs Epic. Epic does not need any particular one of them. Anthropic has been working a different angle, embedding Claude into Elation Health's EHR for primary care chart summaries and reaching patient-held records through HealthEx under TEFCA rather than going through the dominant enterprise vendor at all.

About those safety numbers

OpenAI says physicians rated 99.1 percent of responses safe across 4,363 ratings covering 27 clinical use cases. On the new public data connectors, Becker's reports accuracy ratings running from 93.2 percent for CMS Coverage to 98.6 percent for DailyMed. These are internal evaluations, not peer-reviewed studies and not clinical outcomes.

They should still be reported, and pushed on. What counts as "safe" in this framework, and who wrote the definition? Four thousand ratings is a small sample against the deployment being described. And the 6.8 percent of coverage answers that physicians did not rate good or better reads one way in a benchmark table and another way when someone acts on a coverage determination.

What would change this

The biggest variable is write access. If ChatGPT can put a draft note or an order into the record, this becomes the story people are writing prematurely today, and everything above needs revisiting.

Scale is the second. UCSF is one pilot site, and if HCA's footprint goes live the pilot framing stops holding.

Then there is Epic. If the company tightens its integration terms, prices Art more aggressively, or announces something exclusive with another model provider, that will tell you how it read September 1. STAT reported from Epic's August annual meeting that customers were already asking what the company's AI strategy actually is.

None of that changes where things stand today. Epic took this market in February. What happened this week is that a very large company got in line for a job the customer already has covered.