> ## Content Index
> Fetch the complete content index at: https://www.healthdatacon.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Agent-ready EHRs still need ordinary software to do much of the work
- URL: https://www.healthdatacon.io/articles/agent-ready-ehrs-deterministic-workflows-hlth-2026/
- Published: 2026-11-17T18:50:00.000Z
- Updated: 2026-11-19T04:29:26.000Z
- Description: Avon Health says custom workflows are driving interest in its APIs, while Epic, NVIDIA and HCA put permissions, evaluation and clinician control ahead of broader autonomy.
- Author: HDM Staff
- Tags: Health IT, Americas, HLTH

One of the strongest signals of demand for an AI-native electronic health record is coming from a less futuristic capability: application programming interfaces. Avon Health CEO Maitreyee Joshi said practices without engineering teams are using the company’s APIs and AI coding tools to build highly specific workflows, including complex scheduling rules. For repeatable tasks, she said, an AI agent can introduce unnecessary expense and delay.

“Agents are not the right solution to that either,” Joshi said during a November 17 discussion about agentic-native EHRs. Her example complicated the idea that a modern record should hand every task to an autonomous system. The opportunity she described was to let organizations adapt software more easily while keeping predictable work predictable.

CMS Innovation Center’s Amy Truong moderated [the Ferrari Stage panel](https://hlth.com/events/usa/agenda/2026/are-agentic-native-ehrs-the-future?ref=healthdatacon.io) with Joshi; Derek De Young, who works on AI and analytics platforms in Epic’s research and development organization; NVIDIA director of digital health Raghav Mani; and HCA physician and technology leader Dr. Mark Kaplan. They agreed that demand for AI was strong, but their discussion repeatedly returned to the machinery around an agent: permissions, tests, monitoring, clinical policies and a way to determine who is responsible when something goes wrong.

## Being native means having bounded access to actions

Joshi defined an agentic EHR as a platform where agents are treated as first-priority users, with permissions and capabilities analogous to those assigned to people. That can mean navigating a chart or initiating an order within the system’s controls. It does not mean that every component underneath the conversational interface is itself generative AI.

“All of those parts are actually very, very deterministic,” she said of the back-end functions that carry out tool calls. In her description, a model interprets what a user wants, while established software performs actions such as creating a laboratory order. Keeping those layers distinct reduces the number of places where unpredictable generation needs to govern an action.

Her API example pushed the distinction further. Practices may have a precise sequence for booking examinations or arranging follow-up that a general EHR does not support out of the box. Helping them build that sequence can improve fit without requiring an agent to reason through the same known rules on every occasion.

## Epic’s platform reaches beyond its own screens

De Young described Epic’s Agent Factory Connectors as a way for agents operating inside Epic to reach tools and information in other systems. A payer portal, registry or other application can otherwise require staff to leave one interface and repeat work elsewhere. A common connector offers developers a way to make a capability available to more than one AI application.

He said Epic was also working on incoming Model Context Protocol connections, with an initial release expected in the first quarter of 2027\. That was a forward-looking product timetable. Epic’s [October announcement](https://www.epic.com/epic/post/introducing-connectors-a-new-way-for-apps-to-communicate-with-epic/?ref=healthdatacon.io) separately says developer-created connectors will appear in Showroom starting in March 2027, with organizations able to review their reading and writing permissions before enabling them.

The infrastructure is as important as the features, De Young argued, particularly as agents gain the ability to work across applications. Feedback, evaluations and monitoring have to make problems observable after deployment. He described organizations running agents in a shadow mode, showing what they would do in production without allowing those actions to affect care.

“Deterministic logic is not going away,” De Young said. He also described a future in which agents help improve EHR configuration, with the resulting changes reviewed before entering production. “It's the cheapest way. It's the safest way to do something,” he said of deterministic logic. Those remarks pointed to agents helping build conventional automation as well as performing tasks directly.

## Trust can be lost faster than it is established

Kaplan, a nephrologist who previously worked on technology and a new EHR at DaVita, emphasized the clinician’s position in an automated workflow. He said clinical decision-making still requires a human in the loop. For an ambient documentation tool, that includes a physician’s ability to resolve nuance or contradiction and decide what the final record should say.

“The minute you have inaccuracy that a physician sees, it's really hard to regain their trust,” Kaplan said. He connected that concern to extensive testing before and during pilots. Adoption therefore depends on accuracy that clinicians can experience, rather than a demonstration that works in a favorable example.

“Accountability before autonomy,” he said, returning to the phrase as the panel discussed readiness. His proposed division of responsibility placed clinical decisions with providers, implementation with health systems and software development and integrity with vendors. He explicitly presented that as his perspective, not a legal determination settling liability.

De Young said actual arrangements can be less tidy. A vendor may supply the platform while a health system supplies clinical policies, protocols and the visit types it enables. That combination creates a middle ground where responsibilities need to be defined rather than assumed from the label on the technology.

## Controls belong outside the model, too

Mani described NVIDIA’s [OpenShell](https://www.nvidia.com/en-us/ai/openshell/?ref=healthdatacon.io) as an example of a secure runtime that can limit agents’ access and monitor their activity. His point was that a useful safety boundary can come from the environment in which an agent operates, alongside the model’s own behavior. It gives developers a place to scope tools and capabilities.

“You can very severely scope the capabilities of that agent,” Mani said. He also urged developers to share more evaluation work publicly so model builders can improve against problems that matter to deployed systems. Those proposals address infrastructure for safer use; they do not establish that a particular clinical application has been validated.

Simulation was another option he identified for testing interactions before broader deployment. A simulated patient can help an organization examine responses across different circumstances, though he cautioned that it is not a guarantee. “There's enough to do even with partial autonomy,” Mani said, pointing to applications where a clinician already reviews the result.

## Risk and value determine what moves into care

Truong connected the technical discussion to federal questions about safety and payment. The FDA’s [August discussion paper](https://www.fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices?ref=healthdatacon.io) sought input on generative-AI medical devices, including risk assessment, evaluation and monitoring. It was a discussion document, not final guidance or authorization for autonomous clinical practice.

Panelists argued that a revenue-cycle task and an autonomous clinical visit require different controls because their potential consequences differ. Kaplan also insisted on value alongside safety, including how time returned to staff is actually used. The discussion supplied no single threshold that would make an entire EHR ready for autonomy; it described a series of decisions about particular tasks, evidence and responsibility.

That leaves buyers with a more specific test than whether a product calls itself agentic-native. They need to know which actions an agent can take, which rules remain conventional software, how clinicians can correct it and what happens when it drifts. Joshi’s experience with APIs suggests that a valuable next-generation EHR may succeed partly by making familiar automation easier to build, while reserving agents for work that genuinely needs their flexibility.