Keeping Unsigned AI Draft Notes Out of the Legal Medical Record
Treat the legal record boundary as an enforced architectural control, not a policy statement. Define draft and authenticated states with your HIM or records custodian, inventory every path an AI tool can use to write to the chart, restrict those paths to a draft or working state, require clinician sign-off outside the AI tool, and keep evidence of what the AI was permitted to write and what it actually wrote. Runtime policy enforcement covers individual interactions and tool calls within that design. EHR configuration, records policy and clinical workflow ownership sit outside it.
Where drafts can cross into the record
- 1
Ambient scribe write-back
A scribe integration posts a generated note into the EHR when the encounter ends, sometimes under a service account rather than a clinician identity.
- 2
Direct EHR write APIs
A note-drafting assistant calls a create or update endpoint. The draft state depends on what that endpoint and its scopes permit.
- 3
Agent tool calls
An agent with a documentation tool can write to the chart or a document store as a side effect of a task.
- 4
Document repository uploads
Generated summaries land in a shared repository that other systems index or release.
- 5
Clinician copy-forward
A clinician pastes draft text into a signed note. This is a legitimate path, but it moves the review obligation to the human.
- 6
Amendment of existing entries
An AI tool modifies a note that was already authenticated, which raises authorship and amendment history questions.
Define the boundary before you enforce it
Enforcement needs a precise definition to enforce. Records definitions, including what counts as the legal health record, vary by organization and jurisdiction. The boundary therefore has to be set by your HIM function or records custodian with legal counsel, and this guide does not give legal advice. What architects can contribute is an explicit state model that engineering can implement.
A workable model has at least three states. A draft is content generated by an AI tool that no clinician has adopted. A reviewed draft has been edited or accepted by a clinician but not yet authenticated. A legal entry has been authenticated by an identified clinician in the system of record. Only the last state is a candidate for the legal record, and the transition into it should be an act by a person, not a side effect of an API call.
The idea that a signature expresses an individual's intent to authenticate is familiar from the electronic-signature definitions in 21 CFR Part 11, which describe a signature as executed or adopted with the present intention to authenticate a writing. Part 11 is scoped to FDA record requirements, so it is not the standard for provider notes. It is still a useful design reference for why sign-off must be tied to an identified individual.
The boundary also has disclosure consequences. As of November 2026, 45 CFR 164.524 gives individuals a right to inspect and obtain a copy of PHI in a designated record set, and 164.526 gives them a right to request amendment of PHI in that set, with action required within 60 days of a request. Whether a given AI draft, transcript or prompt falls inside a designated record set is a question for counsel and HIM, which is a further reason to keep drafts in a separate, deliberately governed location.
A layered control approach
No single control holds the boundary. Vendor defaults can change, EHR configuration can drift and reviewers miss things. Layer controls so that a failure in one is caught by another.
The first layer is write-path restriction. Give AI tools the narrowest credentials and scopes available, and use a draft or working-state target where the receiving system supports one. Where it does not, route output to storage outside the record repository. The second layer is draft-state handling: drafts carry an explicit marker, live in separate storage, and are never indexed for release as record content.
The third layer is the sign-off requirement. Enforce it outside the AI tool, with the signing identity captured by the EHR's own authentication mechanism, so the AI cannot attest on a clinician's behalf. The fourth layer is retention and disposition. Agree draft retention periods with the records custodian, and treat transcripts, prompts and intermediate outputs as a separate disposition decision rather than assuming they follow the note.
The fifth layer is evidence, covered below. Test the blocked path as deliberately as the approved one: attempt a draft write without sign-off and confirm it is stopped and recorded.
- 1. Write-path restriction: Least-privilege credentials, draft-only targets, no write access where it is not needed.
- 2. Draft-state handling: Explicit draft marker, separate storage, excluded from record indexing and release.
- 3. Human sign-off: Authentication by an identified clinician in the EHR, outside the AI tool.
- 4. Retention and disposition: Rules set with the records custodian, including transcripts, prompts and intermediate outputs.
- 5. Audit evidence: Retained policy and a record of each attempted and completed write.
Where runtime policy enforcement fits
Runtime enforcement applies to individual AI interactions and tool calls as they happen. Trussed enforces policy in the execution path: when an AI tool attempts a write to the EHR or a document system, the call can be allowed, redacted, blocked or escalated for human approval, and this is enforced at runtime, not reviewed after the fact. In this design that means a write attempt to the record target that lacks the required approval can be stopped before it reaches the chart, and a write to a draft location can be permitted.
This is the layer that closes the gap between what a vendor integration is configured to do and what it is actually allowed to do on a given call. It does not replace the other layers. It does not configure draft states inside the EHR, it does not set records management policy, and it does not own clinical workflow. Those need named owners: the EHR team for configuration, the records custodian for definitions and disposition, and clinical leadership for the review and sign-off process.
There is a tradeoff to plan for. Escalating for human approval adds a step, and if it is applied to too many calls it will push clinicians toward workarounds. Scope escalation to writes that target the record, and let draft-location writes proceed under allow rules.
Evidence an auditor or records custodian will ask for
The evidence goal is to show what the AI was permitted to write and what it actually wrote. Treat the elements below as design recommendations. Which elements your organization requires is a matter for your records custodian, compliance team and counsel.
The permitted side is policy: which tools may write to which targets, in what state, and under what approval condition, with a version history so you can show the rule in force on a given date. The actual side is a record of each attempted and completed write: the target system, a reference to the content, the policy decision (allowed, redacted, blocked or escalated) and the approver where one applied. EHR audit trails then show the authenticating clinician and the signature event, and the two sources together tell the full story.
Runtime enforcement generates the decision records for individual interactions and tool calls as audit-ready evidence, which Trussed maps to frameworks including HIPAA. It can be deployed self-managed in the customer cloud with configurable retention and data residency, which matters when the evidence itself contains sensitive content. It does not produce the EHR's signature records or your retention schedule for draft artifacts.
- Policy in force: Write permissions per tool and target, with version history.
- Attempted writes: Target system, content reference and timestamp for every attempt.
- Decisions: Allow, redact, block or escalate outcome for each attempt.
- Approvals: Approver identity and time for escalated writes.
- Signature events: Held in the EHR, tied to the authenticating clinician.
Who owns which control
Assigning ownership early prevents the common failure in which each team assumes another one holds the boundary.
- Legal record definition and draft states: Owner: HIM or records custodian with counsel. Runtime enforcement does not set this.
- EHR draft-status and role configuration: Owner: EHR platform team. Runtime enforcement does not configure it.
- Allow, redact, block or escalate on a write attempt: Owner: security and architecture, enforced in the execution path of the AI call.
- Clinician review and sign-off workflow: Owner: clinical leadership. Authentication is captured by the EHR.
- Draft, transcript and prompt retention: Owner: records custodian. Retention of enforcement evidence is configurable by the deploying organization.
Draft-to-record lifecycle and its control points
Generate
The AI tool produces draft content in a working state.
Write attempt
Each write to the EHR or document system is checked at runtime.
Hold
Drafts stay in separate storage with a defined retention and disposition rule.
Sign
A clinician reviews and authenticates the entry in the EHR.
Evidence
Policy, decisions and actual writes are retained for audit.
Map your AI write paths to enforceable controls
Walk through your documentation tools, their write scopes and the evidence your records custodian will need, and see where runtime enforcement applies.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.