You're Not Ready for Epic’s Agent Factory

John Lee is an emergency physician and Epic consultant at HIT Peak

Advisors. Taryn Shipley is an Epic architect at Sutter Health and principal at Lean Health Tech.

Somewhere in America right now, a small army of healthcare executives are in a frenzy. They heard a pitch about Epic's Agent Factory. Maybe it was at HIMSS26 in Las Vegas. Maybe it came from a vendor demo, a LinkedIn post, or an Epic advisory council they sat on last quarter. Whatever the source, they or the people around them are now convinced that their health system will soon have AI agents autonomously managing medication reconciliation, processing prior authorization at light speed, and rebuilding Epic workflows faster than their entire IT department.

Almost none of that is going to happen anytime soon. There is quite a distance between that vision and what the product can actually do now or in the near future.

THE HYPE MACHINE IS RUNNING. THE PRODUCT IS NOT.

Depending on who you ask, Agent Factory is still a pre-alpha product. The current interface is a drag-and-drop visual builder that is familiar to anyone who has used other ubiquitous automation software like n8n or Zapier. There are flowchart-style boxes connected by arrows with mechanisms to attach context and reasoning. The "no coding required" pitch is especially appealing because Agent Factory is embedded in Epic's Chronicles. Therefore, you do not have to manually wire APIs the way you would if you tried to use a third-party automation tool. This is the real architectural moat. Because you are doing this within Epic, Epic-specific guardrails are baked in.

As slick as this sounds, our estimate is that a realistic timeline for Agent Factory to reach meaningful operational momentum at leading health systems is two to three years. Widespread adoption will take considerably longer. Anyone familiar with Epic's adoption history will recognize that many applications took much longer than that to reach critical mass. Agent Factory is fundamentally more complex. That may lead one to believe that it will be adopted more slowly, but the hype cycle is going to push at least some organizations into decisions they are not ready to make.

First of all, they may not have the underlying infrastructure, governance, and skillset to fully leverage the tool. More than 85% of Epic's customers now use AI in some capacity, but using ambient documentation, voice recognition, or predictive models is not the same thing as deploying agentic workflows. The activation costs go well beyond licensing and token spend. Organizations need a team that is fluent in both Epic and large language models. They need data and AI governance structures that are simultaneously meticulous, semantically rich, and agile. The health systems that navigate these requirements first will be positioned to win when the tool is ready. The ones chasing the FOMO will arrive at a platform they are not prepared to use.

Probably the most fundamental skill an organization will need is deep familiarity with Epic's underlying record and data structures. You will need to know the Clarity and Caboodle tables to create insights from your data, and the Chronicles master files to connect those insights to action. The visual interface will lower the execution barrier, but it does not lower the knowledge requirement. If anything, Agent Factory will magnify the knowledge requirement. The flip side is that it will magnify the skills of those who already know Epic well. In other words, if you are already skilled at configuring Epic, Agent Factory will be a force multiplier. Using agentic tools to compensate for skimping on your analyst and informatics workforce is like putting a turbocharger on an engine that has a cracked engine block.

THE WORK DOESN'T GET EASIER. IT JUST CHANGES.

Epic introduced three persona-based agents at HIMSS 2026: Art for clinical workflows, Penny for revenue cycle, and Emmie for patient engagement. The ROI figures are real and worth taking seriously. Clinicians are completing discharge summaries 20-30% faster with Art. At The Christ Hospital, Art's extraction of incidental findings from radiology reports drove a 69% early lung cancer detection rate against a national average of 46%. At Summit Health, Penny reduced prior authorization submission time by 42%, with 92% of AI-generated responses accepted without human edits. At Rush University Medical Center, Emmie reduced billing-related customer service messages by 58%.

These numbers are a testament to the power of AI, but do not provide the full picture.

In every one of those use cases, the licensed provider or staff member is still the accountable intermediary. Art drafts the note. The physician signs it. Penny generates the appeal. The coder reviews it. Human-in-the-loop is a structural element of these workflows.

On the other hand, implementing more complex, multistep AI agents in healthcare should not be a blind "set it and forget it" process. These are complex tools and will require an equivalent amount of governance. Some of the ROI from the agent must be redirected toward maintaining compliance and governance for that agent. Organizations need to be mindful of this consideration in their resource planning. This is called out in regulations as well. The February 2026 HIPAA Security Rule updates explicitly classified agentic AI systems as a distinct regulatory category, requiring AI-specific risk analyses, enhanced access controls, and documented oversight procedures for any agent interacting with protected health information.

YOUR WORKFLOWS MAY BE DYSFUNCTIONAL

The most predictable mistake organizations will make with Agent Factory is automating their current processes as-is.

Most health systems run on workflows that are functional but imperfect. Clinicians and front-line operations have adapted around the gaps. Analysts have built workarounds. The friction is tolerable enough that redesigning from scratch has never quite been worth it.

Now imagine you automate the dysfunction.

The agent runs. The workflow executes at scale, reliably, without anyone doing the manual steps that used to generate complaints. Nobody gets annoyed, because nobody is doing the work anymore. The workflow gets swept under the rug. The dysfunction is hidden but also systematically incorporated. What you have done is embedded and amplified dysfunction. And you're also churning tokens unnecessarily.

Dysfunctional processes are only part of the problem. Research documents that 20-40% of critical EHR data points are missing in typical records. Anyone who has spent time inside a live EMR implementation would tell you that number is conservative. Missing data, inconsistent coding practices, and clinically relevant information buried in unstructured free text is epidemic. Even more concerning is the opposite: inaccurate information created as the output of workflows designed to encourage practices like copy-paste. Agent Factory does not solve any of that. An agent is only as reliable as the data it is fed. Automation does not fix the data. It just moves more dysfunction faster.

We are both deeply invested in Epic and in using data to improve the healthcare system, so naturally, we were excited when we first heard about Agent Factory. Taryn is now working with the tool, and through ongoing conversations about her experience, one thing has become clear: the tool is not a magic wand. We are both close enough to the work to understand the current state of most organizations' data foundations. Fragmented, inconsistent, and incomplete data foundations will be a major constraint to the effective use of Agent Factory. Organizations that have invested the time and effort in building solid data foundations will have an even greater strategic advantage.

HERE’S WHAT MOST EXECUTIVES ARE IGNORING

Healthcare has a long and humbling history of failed standardization efforts: alerts clicked through without being read, order sets based on dubious evidence, documentation requirements that people click randomly just to get on with their day. What AI offers, if used correctly, is a different mechanism entirely.

Take a workflow that is genuinely well-designed with true first principles in mind. Fully optimize it and only then use automation to make it significantly faster and easier than the alternative. Automating content adulterated by current workflow compromises will embed the dysfunction. The key is to recognize that we created many of these data collection compromises because we had poor alternatives. Think of the required questions on our orders, odd flowsheet documentation shoehorned in because a quality director wanted someone to document it.

The opportunity is to use AI to address the first-principles reason a workflow exists, not to automate the compromise that replaced it. Instead of adding a required field in an order requiring a reason for an antibiotic, use LLM-powered semantic reasoning to gather that information and put it into a place in the record where it is actually useful. Instead of creating a library of HCC codes that get attached to patient encounters like merit badges, a high-fidelity picture of a complex comorbid versus simple patient can be assembled every year with RAG-powered attribution. You are no longer forcing people toward the right workflow. You are removing the friction that was keeping them on the wrong one.

There is a related point worth naming: some interventions that could meaningfully improve care quality are not happening because the manual effort required is genuinely prohibitive. Clinicians and IT teams have limited capacity. If automation can lower the demand on the humans creating and maintaining the system, it will unlock improvements that no amount of mandate-and-measure would have produced.

THIS IS NOT AN ANALYST PROBLEM

A part of the Agent Factory readiness conversation concerns analyst staffing and knowledge. There is an assumption that if you have deep enough technical expertise, the rest will follow.

That is not necessarily the case.

The most difficult part of deploying agentic workflows is not the technology rollout. The real challenge is successfully coordinating organizational change. And it fails in predictable ways depending on which leg of the stool is missing.

A health system with strong analysts and weak operational buy-in produces technically correct agents that nobody uses. A health system with executive vision and weak technical depth produces announcements that do not survive implementation. A health system with operational alignment and technical capability but no leadership cover produces pilots that never scale, because every cross-departmental dependency becomes a negotiation that nobody has authority to adjudicate.

Progress only happens when technical, clinical, and operational gaps are addressed. The profile that matters most in this environment is the person who can hold clinical, technical, and operational judgment simultaneously. Someone who understands what a clinician or other end user is trying to accomplish, can translate that into what is actually buildable in Epic, and can anticipate the operational consequences of how it gets built. Those people are rare. Health systems that cultivate those skills will move faster than those that do not.

Consider something as simple as the reason for exam field on a radiology order. A clinician who knows that a nurse's ED triage note captures that information over 90% of the time sees a prepopulation opportunity. An analyst who knows Chronicles can build it. Without both perspectives present, that insight either never surfaces or takes weeks of friction to execute. The closer those two perspectives get to living in one person, the faster things move. Agent Factory compresses that friction to minutes, but only if someone already understood the problem worth solving.

WHAT EXECUTIVES SHOULD ACTUALLY BE ASKING

The question dominating leadership conversations — from the CIO's office to the CMO's — is when we can go live with Agent Factory. It is the wrong question.

Adam Farren, CEO of Canvas Medical, said plainly after HIMSS 2026 that he does not think hospitals are ready to take advantage of the platform yet. The gap between what Agent Factory requires and what most health systems currently have is substantial.

Instead of fixating on a technology timeline that Epic largely controls, leaders can direct attention toward organizational readiness. If you are an organization itching to start using Agent Factory, ask yourself which workflows have significant variation and a high number of manual steps. Would those workflows be worth automating in their current state, or do they need to be redesigned first? Does your team understand your Epic build deeply enough to know what they are connecting and what happens when an edge case falls outside the agent's parameters? Have you thought about your process for identifying when an automated workflow fails, especially when the friction that once exposed issues no longer exists? Have you completed an AI-specific risk analysis — not as a compliance checkbox, but as a genuine audit of where your data has gaps that will constrain what agents can reliably do?

We both firmly believe Epic is building something with genuine long-term value. In fact, we believe that developing the skills within the next two to three years to use Agent Factory will be a key critical competency for any Epic organization. Having this competency will allow Epic organizations to orchestrate everything connected within a single ecosystem without APIs. Think of all the Epic applications and master files that can be stitched together with plain English. Then include the insights from Clarity and Caboodle. Finally, think about insights derived from data from over 300 million patient records in the Cosmos dataset. You can build powerful AI on top of Epic. You cannot build what Epic is building from inside it.

But these advantages will only accrue to organizations that have done the data governance work, built the team depth, and designed their workflows before trying to automate them. The competitive advantage of the Epic ecosystem will not be available equally to all Epic customers. It is available to those who have built the foundation to use it.

That work starts now. The agents come later.

THE PROMISE IS REAL. THE PRECONDITIONS ARE NOT OPTIONAL.

John Lee is an emergency physician and Epic consultant who helps health systems bridge the gap between Epic’s capabilities and operational reality. He specializes in data architecture, registry optimization, and making Epic’s tools actually deliver results.

If you need help configuring your Epic environment to support these capabilities, connect with him on LinkedIn or via his website.

Taryn Shipley, MBA, CSSBB, is the founder and principal Epic consultant at Lean HealthTech. She specializes in AI, analytics, and population health—and in closing the gap between insight and execution. Connect: LinkedIn | Website | Email