Insights

AI Governance

Agents reading the chart need more than FHIR, and more than a login

Josh Mandel argues AI agents can now review a patient chart, but the data they need sits partly outside the FHIR APIs EHRs must offer, and getting vendor approval can take months. Whatever access route wins, the practice still has to decide what the agent may read, what it logs, and who signs.

November 18, 2026 · 6 min read

Agents reading the chart need more than FHIR, and more than a login

Two things an AI agent needs to prepare a wellness visit, the patient's message history and the practice's health maintenance list, are outside USCDI, the data set certified FHIR APIs have to serve. That is one of the points Josh Mandel makes in a post on X and the longer piece on his blog. He writes that no version of USCDI through v7 includes either.

Mandel's headline claim is that AI agents can now reliably review a patient's chart, but agentic access needs data and workflows beyond USCDI FHIR APIs, and under current rules a practice can spend months working through EHR partner fees, IP terms, and app registration. His piece is a regulatory proposal. Mine is narrower. Whatever happens with the rules, a practice that lets an agent into the chart has to decide, in writing, what that agent is allowed to do. That is a governance document, and most practices do not have one yet.

What Mandel argues

I want to represent his argument accurately and not stretch it.

In his piece, agents that operate software the way a person does, reading the screen, clicking, and typing, now work well enough to read through a chart before a visit, draft the note, and queue orders for a clinician to sign. He illustrates the access problem with a story he labels as fiction: a small clinic, a four-person startup, and an EHR vendor, where every step the vendor takes is one the current rules allow. In his stylized calendar, the process runs 96 business days and ends with an integration that still cannot see the message history or the screening schedule.

His proposal is that certified EHRs should let a customer have software of its choosing operate the EHR's ordinary screens, under a clinician's login or an account the customer sets up for it, without a separate license, fee, or vendor approval. He describes two setups. In one, the agent works inside a clinician's own session with that clinician's access. In the other, the customer provisions the agent like a scribe or medical assistant: a named account, a role limited to the job, and a sponsoring clinician who reviews and signs, with the agent's actions logged separately and revocable on their own. He says the second will often be the better model.

He also answers the "hallucinating agent" objection by saying the agent can only do what its account's role permits, and that whether drafts need review before signing is a clinical governance question for the organization, the same one it answers for scribes and students.

That last point is where this post picks up.

The governance gap after access

The regulatory fight is about whether the practice can let the agent in. Even if Mandel's proposal were adopted tomorrow, the practice would still own the next set of decisions: what this agent may read, what it may change, what is recorded, and who answers for it.

Health systems already make these decisions for people. A scribe gets a role, a supervisor, and a scope. A medical student's notes require cosignature. Those rules live in credentialing policies and EHR security templates. An agent needs the same treatment, with a few additions because software does not get tired, works at night, and can be copied.

For HealthTech founders building these agents, this is also the document your customers will ask for. Having a draft ready shortens the conversation.

A permission template for agent chart access

One page, one agent, one use case. Fill it in before go-live and review it on a schedule.

  1. Purpose. One sentence describing the job, such as preparing a pre-visit summary for annual wellness visits. If the purpose needs two sentences, it is two agents.
  2. Identity. Its own named account, or a clinician's session? If its own account, who created it and who can disable it?
  3. Sponsoring clinician. The person who reviews the agent's output and answers for what is signed, by name or by role.
  4. Read scope. Which parts of the chart, and which patients. For a visit-prep agent, perhaps only patients on tomorrow's schedule, and only notes, labs, outside records, messages, and the health maintenance list.
  5. Write scope. Which actions it may take, stated as allowed actions rather than forbidden ones. Draft but not sign. Pend orders but not release them. Never send messages to patients.
  6. Hours and volume. When the agent runs and roughly how many charts per run. Anything far outside that is an alert.
  7. Audit trail. What is logged for every action, where the log lives, who reviews it, and how often.
  8. Data handling. Whether chart content leaves the EHR, where it goes, which subprocessors see it, and whether anything is retained after the run. Tie this to the business associate agreement.
  9. Human sign-off. Which outputs need review before they reach the chart or a patient, and how a reviewer flags an output as wrong.
  10. Stop and review. Who can turn the agent off, how fast, and the date of the next review of this page.

What to watch for in practice

The template fails in two predictable ways.

The first is scope creep. The agent works, someone asks it to do a second job, and the role gets widened instead of a second agent being created. Six months later nobody can say what the account is allowed to do or why.

The second is a sign-off that stops being real. If a clinician signs forty pre-visit summaries in ten minutes, the review step exists on paper only. Watch for it in the audit log and talk about it in the review meeting.

Neither problem depends on how the agent got in. Both are on the practice.

Where to start

If you are a practice or health system piloting any chart-reading agent, fill in the ten lines above for it this week. If you cannot answer read scope, write scope, and who signs, pause the pilot until you can. If you are a founder, send the filled template to your next prospect before they ask. Mandel's piece is worth reading in full for the access argument. The permission page is the part you control today.