Subscribe | Advertise | PRO Member Hub

Before asking what an agentic EHR might become, it is worth stepping back and asking: how did the health record become the place where so much of healthcare happens?

For our readers enmeshed in the healthcare space, ideas like “ambient scribing,” “health record interoperability,” and “agentic health records” may already feel like common vernacular. But much of what we discuss as an inevitability in the modern healthcare workflow has not necessarily been that way for very long.

Since the times of ancient Egypt (at least), physicians have recorded their work. In those early days, documentation was personal and intended to facilitate pattern recognition. It was written by physicians, for physicians. As our ability to standardize and disseminate knowledge progressed, physicians moved from writing their own individual catalogs of cases to developing more formal casebooks.

The emergence of the modern hospital transformed what was individual into an institutional requirement, facilitating care across increasingly complex systems. These systems also began tracking and organizing other parts of a patient's care: orders, lab and imaging results, medication administration, and much more. With the introduction of insurance, the record became a legal and financial document too, serving as the “ground truth” for what had happened to the patient.

While this was initially done on paper, the advent of widespread computerization heralded a transition. We were promised that electronic health records would speed up and simplify healthcare.

Today, the health record has become entwined not just with clinical documentation, but also with order entry (labs, imaging, referrals), billing, and communication between all levels of medical staff. Moreover, with the rollout of modern patient data-sharing rules, the record is no longer just for medical providers. It is for patients too, as they increasingly read and digest what happens within the EHR.

What began as a place to write down important lessons for self-learning has become a database of patient information and a fundamental component of how we practice medicine and receive healthcare. The EHR is no longer simply a digital version of the paper chart. It has become the infrastructure through which much of healthcare operates.

But the EHR is also widely considered a pain point and a contributor to physician burnout. A 2024 AMA report highlights that physicians spend around 13 hours on indirect patient care and 7.3 hours on administrative tasks out of a 57.8-hour average workweek. EHRs are the home and route for much of this non-patient-facing workload. A 2017 study of primary care physicians found they spent an average of 5.9 hours interacting with the EHR during an 11.4-hour workday, including documentation, orders, inbox management, and other administrative tasks. In other words, more than half their working day involved the electronic record.

Despite how central they've become, EHRs still require clinicians and staff to initiate, navigate, and complete much of the work themselves. They may be digital, but they are not necessarily digital-first tools designed around improving patient care.

Which brings us to our question today: could the EHR, as a system of record, also become an agentic system that actively helps deliver care?

Throughout this piece, we'll use “agentic” to describe AI tools capable of taking steps within a workflow rather than simply generating information for someone else to act on. This can range from preparing orders and coordinating administrative tasks to executing more complex clinical workflows. The degree of independence, and where human review is required, varies considerably. While the technology underpinning what is “agentic” can vary based on the use case, most often this process involves language models and the interplay between structured and unstructured data. 

We'll look at the tools already entering the chart, what research suggests may come next, and how startups and incumbents are approaching the transition.

Ambient documentation tools like Ambience, Abridge, Dragon Copilot, and Nabla are already helping clinicians reduce documentation time, with some reporting substantial time savings. Dragon Copilot alone has been deployed across more than 150 health systems.

Ambience, which supports more than 200 specialties, reports that institutions using its technology see over 3x return on investment, with an 80% adoption rate during scheduled visits. At HIMSS 2026, Microsoft expanded Dragon Copilot's scope beyond note-taking into order capture and nursing workflows, signaling that these tools are moving past documentation and into action.

The early evidence on clinician experience is promising, too. One multisite study across six health systems found that self-reported clinician burnout declined from 51.9% to 38.8% after 30 days of ambient AI scribe use. While the study was observational, the findings suggest that these tools may meaningfully reduce administrative burden.

Taken together, these developments suggest that clinicians are beginning to trust AI to sit inside the record, describe it accurately, and participate in their workflows. The logical next step is allowing AI to do more with that information. But trusting a system to draft a note is quite different from trusting it to act on what that note says.

A recently published paper in Nature introduced MIRA, an AI agent capable of operating inside a sandboxed EHR environment. MIRA can pull a patient's history, order and interpret labs and imaging, generate a differential diagnosis, and formulate a treatment plan, including prescriptions, procedures, and admissions. In the research evaluation, its performance was compared directly with physicians, with results suggesting that AI agents can already perform some complex clinical workflows at a competitive level in a controlled environment. At least in a research setting, it feels like the technical capability for agentic action inside the EHR is beginning to exist.

But technical capability is only part of the story. These controlled evaluations are not the same as allowing an AI system to autonomously make decisions in live patient care.

For example, Dr. Michael Han, Chief Medical Officer at Ambience Healthcare, pointed out:

A piece in npj Digital Medicine, "Tracing the Pen," makes the case that as AI-generated and human-generated content start blending together in the same chart, data provenance becomes a real problem. Basically, if you can't tell who, or what, wrote a note or placed an order, you lose the ability to understand how a decision was made and who is accountable for it.

In a world where notes often get carried forward, an AI-generated summary that contains errors, obscures uncertainty, or mischaracterizes the original record could have consequences well beyond the initial interaction.

And once AI begins taking actions, rather than simply generating content, new questions emerge. Who approves those actions? Who is responsible when something goes wrong? What happens when an agent encounters conflicting information, or when the task requires a judgment it cannot confidently make?

Perhaps AI technology is arriving faster to EHRs than the governance, oversight, and liability infrastructure needed to trust it.

On the commercial side, a wave of companies got their start on the documentation side of things and are now pushing further into action. Abridge, Nabla, and Ambience all began as ambient documentation tools and have since expanded into areas such as coding assistance, clinical documentation integrity, and revenue cycle workflows. At this point, we think it is fair to say that the line between "documentation AI" and "revenue cycle AI" is starting to blur.

Cohere Health and Hippocratic AI are approaching prior authorization from different angles, using AI to help gather clinical evidence, prepare documentation, support submissions, and coordinate follow-up. The precise scope of automation, and how much human oversight is needed, varies by platform and workflow. 

Companies like Abridge and Tennr are also increasingly positioning their products around completing work that historically required substantial administrative teams, rather than simply helping those teams work faster. Consider a referral that has been placed but never scheduled. Today, someone may need to check its status, identify missing information, contact the receiving provider, coordinate with the patient, and follow up until the appointment is confirmed. An agentic workflow could potentially handle many of those steps, escalating exceptions to a human when needed.

That is a meaningful distinction. The goal is no longer just generating a better summary of an unfinished task. It is helping ensure the task actually gets completed.

Startups in this space are benefiting from regulatory tailwinds, too. CMS's electronic prior authorization rules, ONC's information blocking enforcement, and TEFCA adoption are all pushing the industry toward greater data exchange and interoperability, conditions that can make AI-driven workflows easier to implement.

Effectively, many of these players are seeking to expand an existing trusted relationship- AI working within or alongside the chart- into new kinds of actions. Instead of trying to build that trust from zero, they are extending it into adjacent workflows. VCs would describe this as a wedge.

Another crop of startups is approaching the problem from the opposite direction: instead of adding AI to existing EHRs, they're trying to build systems designed around AI from the beginning.

Avon Health is one example of what this looks like in practice. The company has built an AI-first EHR with ONC certification and agentic capabilities across intake, eligibility, scheduling, medical coding, result triage, referrals, medication reconciliation, and more.

The idea is to rethink the EHR as more than a place to store patient information. Instead, the record becomes the foundation for a set of connected workflows that AI can help execute.

Other companies are taking a more incremental approach. Avo is attempting to embed AI directly into established EHRs like Epic, MEDITECH, and athenahealth. Avo recently raised a $10M Series A, citing a 35% reduction in documentation time and positive clinician feedback. These are fundamentally different strategies. One seeks to rebuild the EHR itself, while the other aims to make existing systems more intelligent and actionable.

There is also a third group, composed of companies building specialized AI agents that work across the fragmented systems healthcare organizations already use. This is to say there's officially a YC pipeline. Novoflow is building "AI employees" that use a mouse and keyboard to navigate legacy hospital software, sidestepping the closed API problem instead of waiting for EHR vendors to open up. Pelica is building an AI-native operating layer for value-based care organizations, unifying claims, EHR, pharmacy, lab, and ADT data behind a set of copilots. Lastly, Arintra converts patient charts into insurance claims, reporting 96% coding accuracy, an 11% reduction in undercoding, and a 43% reduction in coding-related denials. 

These companies aren't necessarily trying to become the next Epic. Some are rebuilding the EHR, some are building on top of it, and others are finding ways to automate work across systems that were never designed to communicate seamlessly. The EHR automation turf war is likely entering a new wave, but the opportunity isn't simply to make administrative workflows cheaper.

As Avon Health's Maitreyee Joshi explained:

This is an important point. Healthcare doesn't just suffer from too much administrative work. It also suffers from an inability to efficiently complete all the work we already know should happen, for every patient. What this means practically is that work that could, or ought, to be done by a professional is fielded to patients and their families, all exacerbating long-standing inequities.

An agentic EHR could begin to change that equation.

Not wanting to be caught flat-footed, the legacy players have embarked on the same mission. Epic, Oracle Health, and Microsoft are all embedding more AI capabilities directly into their existing platforms. Epic, for example, has continued investing in AI capabilities and integrations, including its recently announced connection with ChatGPT for Healthcare.

These companies have advantages that are difficult to replicate. They already sit at the center of clinical workflows, hold critical patient data, maintain relationships with health systems, and have the infrastructure to distribute new capabilities at scale. This creates an interesting competitive dynamic.

AI-native entrants can build around more modern architectures without the technical baggage of legacy systems. Specialized agents can move quickly on specific problems and sometimes work across multiple EHRs. But incumbents have the distribution, integration depth, and existing institutional relationships that startups often spend years trying to establish. It also raises the question of how much of this market will ultimately be controlled by the EHR itself, versus the companies that build around it.

Will health systems eventually purchase agentic capabilities directly from their EHR vendors? Or will independent companies continue to own specialized workflows where they can offer meaningfully better outcomes?

It's difficult to know how this plays out, but many folks have learned the hard way that you can't really ever count Epic out of the game.

While it is easy to create a landscape of companies working in this space, it is harder to resolve what these developing technologies actually mean for healthcare.

To make this practical, let's talk about an example of common friction in the hospital pipeline, and where value goes underdelivered. Take the example of a patient being discharged to a skilled nursing facility from the hospital. For reference, traditional Medicare alone covered roughly 1.5 million skilled nursing facility (SNF) stays in 2024, and SNFs are one of the most common post-acute destinations, receiving nearly one in five hospital discharges. The transition represents an important step in a patient's journey from hospital back to home, and errors in the process can often result in financial consequences for hospitals, patients, and the facilities. Currently, much of this work is supported by case managers who usually have a clinical background in nursing and then spend most of their time doing the administrative tasks of coordinating with facilities, patients, their families, and hospitals to ensure placement. It's a complicated dance, and anyone who knows a case manager or who has worked with them knows how much effort goes into managing these transitions of care. 

Now imagine a world where an agentic first approach to hospital discharge could revolutionize the way we place patients in skilled nursing facilities. As a patient prepares for discharge, an agent reads the clinical notes, summarizes the trajectory so far, and identifies what the next level of care will likely be. This will include physical therapy, occupational therapy, and any other clinical needs the patient has, like dialysis, and transportation to appointments. Then it interfaces with systems at skilled nursing facilities to identify suitable placements, taking into account the patient's clinical needs, insurance coverage, family location, and personal preferences, while coordinating with the care team and facilities to finalize placement. Finally, as a patient is discharged, the agent summarizes the inpatient medications and ensures that the discharge summary accurately reflects what the patient will need at the skilled nursing facility on the day they leave. If the patient ever has to come back to the hospital or when they go home from the skilled nursing facility, it makes sure to summarize the process again so that outpatient providers or the hospital staff can be clear on where the patient is in their trajectory. Earlier, when we discussed the work that “ought” to be done, and that which is fielded to next of kin instead, this example immediately highlights the obvious discrepancies in outcomes we all notice between patients who have someone to help them, and those who don’t. 

Much of the broader AI debate has centered on whether AI will eventually replace physicians or simply become another tool they use. But for agentic EHRs, the more immediate question may be simpler: what work should the health record be able to do on our behalf? We believe the example above represents the sort of work we should be executing with increasing autonomy.

However, a major stumbling block is that the actual infrastructure upon which these systems act is still not fully modernized. We live in a world where prescriptions can still be phoned into pharmacies, records are faxed (really?), and patients struggle to collect and interpret their complex health data. Putting a sophisticated AI agent on top of fragmented workflows doesn't automatically fix the fragmentation underneath.

That said, the opportunities are substantial. The next 3–5 years represent a chance to automate significant portions of billing, prior authorizations, documentation, referral management, and care coordination. These administrative and operational tasks may represent more achievable near-term wins, with measurable financial and clinical benefits. But there's another question worth asking: should all of these tasks exist in their current form at all?

As Dr. Han added:

That distinction feels important. If we simply use AI to accelerate the same inefficient workflows that made modern healthcare so difficult to navigate in the first place, we may miss the larger opportunity.

At the start of this piece, we wrote about the health record as a sort of “place of truth.” A singular compendium of information that ought to be where we turn to clarify what has happened, what is happening, and what should be done next. While the format and style of records have changed over time, they have fundamentally always been places for human data entry and retrieval. Records have been made by us, for us. Now, that seems likely to change.

We stand at a point where non-human actors will, for the first time, become deeply integrated into these same tasks. Rather than simply recording what happened, the EHR may increasingly participate in determining what happens next. As with the rise of the EHR itself, it seems inevitable that this will change how we practice medicine.

But the risks deserve just as much attention as the opportunities. Ambient scribing may mishear details that are carried forward unnoticed. Summarization may mischaracterize elements of a patient's history that are then propagated through future decisions. Automated billing will meet automated claim denial. And systemic biases in the way we build these tools may have unforeseen consequences for patients.

As much as we hope that the agentic EHR can revolutionize the mundane tasks created by computerization, there is every chance that poorly designed systems simply accelerate existing inefficiencies or push us further from the original purpose of clinical documentation.

Fundamentally then, the question is not just whether an AI agent can complete a workflow, but whether that workflow should exist, whether the resulting action improves care, and whether the system remains accountable to the people it serves.

Ultimately, the promise of the agentic EHR should be to bring us closer to what the health record was meant to support in the first place: understanding what the patient has been experiencing, and helping them.