Epic is building AI into its EHR across the board: a clinician assistant called Art that drafts notes and surfaces patient information, a patient assistant called Emmie inside MyChart, a revenue-cycle assistant called Penny, and predictive models trained on its Cosmos dataset. For hospitals, “good enough” AI bundled into the EHR will often beat better standalone tools.
When Epic took the stage at its user group meeting in August 2025, the most important announcement was not any single feature. It was the shape of the whole: an AI assistant for clinicians, one for patients, one for the billing office, a native ambient documentation tool built with Microsoft, and foundation models trained on Cosmos, Epic’s pooled dataset drawn from hundreds of millions of patient records across its customers. Taken together, it was a message to every AI startup selling into hospitals: much of what you sell, we intend to include.
This piece argues that the message is credible, and that the consequences for the health AI market are underrated. Platform vendors do not need to build the best AI. They need to build AI that is good enough, already integrated and already paid for. For a large share of use cases, that combination beats a better point solution. The startups that survive will be the ones that sell something the platform structurally cannot.
What AI features does Epic offer?
Epic’s AI portfolio spans three audiences, and the naming makes the strategy explicit.
For clinicians, Art is positioned as an assistant that can summarise a chart, pull up trends such as blood pressure over time, draft notes, help with orders and, through the ambient tool, document a visit from the conversation. Earlier generative features, including drafted replies to patient portal messages and chart summarisation, were already in use at many customers before 2025.
For patients, Emmie sits inside MyChart, helping people understand results, schedule and navigate their care.
For the back office, Penny targets revenue cycle work, such as drafting appeal letters for denied claims and supporting coding.
Underneath sit predictive models. Epic has long shipped models for deterioration, sepsis, readmission and no-shows. The newer ones draw on Cosmos, which gives Epic something no startup can easily replicate: longitudinal data across a large share of American patients, already de-identified and aggregated inside its own ecosystem.
Why does Epic’s market position matter so much for AI?
Because the EHR is where the clinician’s attention is, where the data is and where the purchasing decision already sits.
KLAS Research’s 2026 market share report found that Epic held 43.7% of US acute care hospitals and 56.9% of beds in 2025, having gained 77 hospitals that year while Oracle Health lost 56. Epic has added a net 568 hospitals since 2021. Its footprint in large academic and integrated systems, which are the biggest buyers of AI, is larger still.
Federal data shows what that means for AI adoption. The US health IT office (ASTP/ONC) reported that in 2024, 90% of hospitals using the market-leading EHR vendor had adopted predictive AI, compared with 50% of hospitals on other vendors’ systems. Among all hospitals using predictive AI, 80% used models from their EHR developer. The EHR is not just one channel for hospital AI. It is the main one.
Big technology companies sit one layer below. Microsoft supplies cloud infrastructure and ambient technology, and partners with Epic rather than competing against it. Google, Amazon and the frontier model labs sell models and cloud services to EHR vendors, health systems and startups alike. Oracle, which bought Cerner in 2022, is trying to rebuild its EHR around AI to stop losing hospitals. The pattern is that big tech increasingly supplies the raw capability while the EHR vendor controls the last mile into the clinician’s screen. Whoever owns that last mile captures most of the value, which is why the platform question is really an EHR question.
Why would “good enough” AI beat better AI?
Because the buyer is not comparing models. The buyer is comparing total cost and total hassle.
Consider what a hospital must do to adopt a standalone AI tool: run a vendor security review, sign a business associate agreement, negotiate a contract, build an integration, train staff to use a second interface or a new sidebar, and set up monitoring. Each step takes months and involves people who are already overloaded. An EHR-native feature skips most of this. It uses existing data flows, appears where clinicians already work, and often arrives as part of an existing licence or a modest add-on.
A point solution therefore has to be not just better, but better by enough to justify all that friction. For some use cases it will be. For many, a 5% accuracy advantage will not overcome a six-month integration project.
Ambient scribes are the test case. Startups built the market. Abridge, one of the leaders, raised $300 million in June 2025 at a valuation of about $5.3 billion, roughly doubling in four months, helped by being among the first ambient tools integrated into Epic’s workflow. Microsoft’s Nuance, now Dragon Copilot, competes directly. Then Epic announced its own native ambient capability built with Microsoft. The standalone scribes did not disappear. Several have large enterprise contracts and strong clinician loyalty. But the ceiling on their pricing, and the default choice for the next hospital that has not yet bought, shifted overnight.
A worked example: the hospital’s choice
Take a hypothetical 12-hospital system on Epic that has been piloting a third-party AI tool for drafting responses to patient messages. The tool works well. Clinicians like it, and the vendor’s drafts score slightly better than Epic’s native feature in the system’s own blind review.
Now the renewal arrives. The third-party tool costs a significant annual fee, requires a separate security review every year, and needs the integration team to test it with each Epic upgrade. The native feature is already live, is included in the system’s existing agreement, and is updated automatically.
The CMIO argues for the startup on quality. The CIO argues for Epic on cost and risk. The CFO sides with the CIO. The startup loses a customer who liked its product.
Nothing about this is irrational, and it will repeat across hundreds of systems. It is the platform dynamic in miniature: the incumbent does not need to win on quality, only to come close enough that switching costs decide the rest.
What can health AI startups still win?
Plenty, but not by competing head-on with features the platform will bundle.
The first opening is depth. An EHR vendor builds for its median customer across dozens of specialties. A startup focused on, say, oncology trial matching, cardiac imaging or surgical video can go far deeper than a platform ever will, particularly where specialist data and regulatory clearance are needed. FDA-cleared diagnostic AI is not something an EHR vendor casually ships.
The second is being multi-platform. Many health systems run more than one EHR across acquired hospitals, and payers, pharma and life sciences companies do not run Epic at all. A tool that works across systems serves a market no single EHR can.
The third is taking on risk. Startups that sell outcomes, not software, such as taking a share of recovered revenue or guaranteeing a reduction in denials, offer something a platform licence does not.
The fourth is speed. Platforms move on annual release cycles and must satisfy thousands of customers. Startups can go from idea to deployed feature in weeks, which matters in a field where the underlying models improve every few months.
The strongest counterargument to this piece is that platforms have tried to own everything before and failed. EHR vendors bundled analytics, population health and patient engagement tools a decade ago, and best-of-breed vendors still won many of those markets because the bundled versions were mediocre. Epic’s app marketplace and its partnerships with ambient vendors also suggest it wants an ecosystem, not a monopoly. That history is real. But generative AI differs in one respect: the underlying model capability is increasingly a commodity bought from the same few providers by everyone. When the startup and the platform are both calling similar foundation models, the startup’s advantage narrows to data, workflow and focus.
What this means for hospitals, startups and investors
For hospitals, the platform is a gift and a risk. Native AI cuts integration cost and speeds adoption. It also deepens dependence on one vendor, whose models you cannot inspect as easily and whose pricing you cannot easily escape. The safeguard is governance that treats native AI exactly like third-party AI: local validation, monitoring, and a willingness to switch a feature off. Being bundled does not make a model validated for your patients. Our argument that interoperability is a political problem more than a technical one applies here too: control over data flows is control over the market.
For startups, the test is simple. Write down what you sell, then ask whether Epic could ship 80% of it in 18 months. If the answer is yes, you need a different wedge, a different customer or a plan to be acquired. As our review of health VC in 2025 found, capital is concentrating in fewer companies, and platform risk is one reason.
For investors, platform risk belongs in diligence alongside clinical and regulatory risk. The question is not whether a startup’s model is better than the EHR’s. It is whether it is better by enough to survive procurement.
For policymakers, a concentrated EHR market that also controls AI distribution raises competition questions. Open interfaces, fair access to the app marketplace and data portability will matter more as AI becomes the main thing the EHR sells.
Frequently asked questions
What are Epic’s AI assistants called?
At its August 2025 user group meeting, Epic introduced three branded assistants: Art for clinicians, which summarises charts, drafts notes and supports ambient documentation; Emmie for patients within MyChart; and Penny for revenue cycle work such as appeal letters and coding support. They sit alongside Epic’s long-standing predictive models.
What is Epic Cosmos AI?
Cosmos is Epic’s aggregated, de-identified dataset drawn from patient records across participating Epic customers. Epic uses it to train predictive and foundation models for tasks such as forecasting readmissions and health events. Its scale gives Epic a data advantage that independent AI vendors find difficult to match.
Will Epic’s ambient scribe replace Abridge and other vendors?
Not immediately. Established scribe vendors have large enterprise contracts, strong clinician satisfaction and features beyond basic documentation. But a native option lowers the price ceiling and becomes the default for hospitals that have not yet chosen a vendor, which makes growth harder for standalone scribes.
Is EHR-bundled AI safe to use without local validation?
No. Bundled AI still needs local checks because performance depends on each hospital’s patients, documentation and workflows. A 2021 JAMA Internal Medicine study found that Epic’s earlier sepsis model performed much worse at one academic centre than reported. Hospitals should validate native tools as rigorously as third-party ones.
How can health AI startups compete with EHR vendors?
By offering specialty depth the platform will not build, working across multiple EHRs and non-provider customers, selling guaranteed outcomes rather than software, and moving faster than annual platform release cycles. Startups whose product is a generic feature of the EHR workflow face the greatest risk.
The EHR vendor does not need the best AI in healthcare. It needs AI that is close enough, already installed and already paid for. Startups that build their strategy around that fact will outlast the ones still arguing about benchmarks.
If your health system is on Epic, which AI tools have you replaced with native features, and which would you never give up? Let us know in the comments.
