Epic UGM 2026: What Agent Factory and Curiosity Mean for Your Health System

August 28, 2026 Written By Brady Defibaugh, Vice President, EHR and RCM Services

Epic’s UGM in Verona put AI agents squarely on the health system roadmap. The headlines are about Agent Factory and Curiosity. The harder question is what happens after a health system turns them on.

Anyone who has lived through an EHR optimization project knows the pattern: the build can be technically sound and still struggle because ownership is fuzzy or the workflow on paper is not the workflow on the floor. Agents raise the stakes by surfacing information they can act on it.

Here is what to start doing now, well before the platform shows up.

What Epic Announced at UGM 2026

Agent Factory is a no-code platform for building AI agents inside Epic workflows. It starts with roughly 120 existing AI capabilities, and general availability is planned for 2027. Here is the part that changes how you staff and govern. Operational teams can configure agents themselves, without waiting behind a data science project.

Curiosity uses Cosmos data to predict outcomes such as readmission and stroke risk. It is being validated at about 20 organizations.

The 2027 timeline sounds like runway. It is shorter than it looks. Epic also introduced Ergo, its AI-native clinical interface, arriving in November. It puts AI-forward workflow in front of clinicians this year. Turning technology on is fast. Governance, workflow redesign and earned trust are not.

Why This Is More Than Another AI Feature Release

Every EHR release now seems to include an AI feature. Agents are different because they can move from recommending work to performing it.

A clinical decision support alert makes a suggestion and waits. An agent might reschedule an appointment, draft a denial appeal, or initiate patient outreach. If the output is wrong, the problem may no longer be a recommendation someone ignored. The action may already have happened.

The Real Readiness Question

The question is not simply, “Can we turn this on?” The point of a no-code platform is that many organizations will be able to.

The better question is whether the health system can govern it. Say an agent misses an escalation, or sends outreach to the wrong patient cohort. Somebody owns the response, somebody can stop the workflow and somebody explains it to the family. Name those people now.

If those answers are unclear, the organization is not ready yet. Accountability needs to be assigned before the first agent goes live, not after the first incident.

The Six Readiness Decisions

Six decisions separate an agent that earns trust from one that creates cleanup work: which workflow goes first, where a human stays in the loop, how the model gets validated locally, who is accountable, whether the data can support it, and how the workforce is prepared to supervise it.

Choosing the First Workflow

The best first agent is boring. That is the point.

Look for clear inputs, clear outputs, a named owner and a result you can measure inside a quarter. If the use case takes a page to explain, start somewhere else. Same if six departments must agree on what success means.

Good candidates:

• Scheduling, including no-show outreach or waitlist backfill. High volume, low risk, easy to measure.
• Denials, such as drafting appeals from existing documentation while a person reviews the result before submission. Epic reported its Denial Appeals Assistant live at more than 330 organizations. Appeals were created 23% faster for medical necessity denials.
• Documentation, including summarization and inbox drafting where clinicians retain final control.
• Readmission outreach, where a prediction can route a patient to a human care team for follow-up.

Diagnosis, medication and triage are poor places to learn how your organization manages agents. Those use cases may come later, after the operating model has earned trust.

The early results support the boring approach. ECU Health used an agent to summarize Transfer Center requests and estimated roughly 20 hours saved per week. Narrow scope, a named owner and a number leadership can verify.

Human-in-the-Loop Design

“Human in the loop” only becomes useful when it is translated into workflow decisions.

Which decisions require approval? A sensible starting point is to require review for anything that reaches a patient or payer. Adjust from there based on observed performance.

When should the agent escalate? Define the confidence thresholds and edge cases where it must stop and hand the work to a person.

Who can override it? The off-switch belongs to a named operational role at the workflow level. It should not be buried inside an IT support process. A charge nurse should not need to file a ticket to stop an outreach workflow at 2 a.m.

Clinical and Operational Validation

Epic’s validation work is important, but it does not replace local validation. Models still have to meet the reality of your patient population, documentation practices and payer environment.

Before launch, test with local data and populations. Look at accuracy, bias, safety and workflow impact. Does it perform well here? Does performance vary across populations? What is the worst plausible failure? Does it reduce work or merely move it to a different team?

Validation also continues after go-live. Payer rules change, documentation patterns shift and models get updated. Set a review schedule, watch for drift, and decide in advance what would trigger a return to the human-only workflow.

AI Governance and Accountability

Governance does not need to begin as another large standing committee. It does need clear ownership.

The CIO owns the platform, integration and vendor relationship. The CMIO defines clinical appropriateness. Compliance addresses regulatory exposure and documentation. Quality measures outcomes and reviews safety. Security owns identity, access and data boundaries. Each agent also needs one operational owner who is responsible for how it behaves in the actual workflow.

Start with two practical artifacts. The first is an audit trail. It records what the agent did, what information it used, and when a person approved or changed the action. The second is an incident procedure covering how to detect, stop, communicate and review a failure. Neither should be invented during an incident.

Data and Integration Readiness

Agents amplify the quality of the data and workflows they inherit. A stale problem list, duplicate patient record or poorly scoped service account can turn one error into a repeated one.

Do not attempt to clean the entire enterprise before beginning. Audit the data elements the first agent will actually use. Use a disciplined Epic workflow design and validation process. Map where its actions land in the work people really perform, including common workarounds. Give the agent the minimum access needed for that job and no more.

Workforce and Change Management

People need to understand what each agent does, what it will never do and how to override it. That requires thoughtful EHR training and user engagement, not an announcement email.

The new skill is supervision: knowing when to trust an agent, when to question it and when to stop it. Teams should also watch for automation that quietly creates more work. If every action generates another verification task or inbox message, the workflow has not really been automated. Measure the net workload, not the number of tasks the agent touched.

What Health Systems Should Do Before 2027

The first steps are straightforward:

• Select one valuable, limited-risk workflow with clear inputs and outputs.
• Assign one accountable operational owner.
• Define success numerically before launch: hours saved, denials overturned or no-shows reduced.
• Document approval requirements, escalation thresholds and an accessible override path.
• Run a controlled pilot and expand only when the evidence supports it.

Benchmarks already exist for the third step. Epic reported more than 40 groups cutting coding denials by 20% or more with its coding assistant. Pick a number in that range, write it down before launch, and hold the pilot to it.

By the time Agent Factory reaches general availability, the goal does not need to be a perfect enterprise AI strategy. A better goal is one useful agent that has earned the trust of the people responsible for its work.

Epic is building the factory. What it produces inside your health system will depend on the decisions you make before turning it on.