Governing AI-drafted operative and procedure notes
Govern AI-drafted operative and procedure notes in two layers. Before deployment, decide the permitted use cases, approved models and tools, and the review policy. During each AI interaction, enforce policy at runtime: allow, redact, block or escalate for human approval. The surgeon's review and signature remains the accountability anchor. Runtime enforcement supports it by constraining what the AI can reach and do, and by producing evidence tied to the signed note.
Map control points across the drafting workflow
- 1
1. Input capture
Source material enters the workflow. Control: redact PHI that the drafting step does not need. Evidence: record of what categories were redacted.
- 2
2. AI drafting
A model produces the draft. Control: restrict to approved models and tools, and block out-of-policy actions. Evidence: policy applied, model and tool used, decision taken.
- 3
3. Clinician review
The surgeon checks the draft against what happened in the procedure. Control: escalation for human approval before the draft can proceed. Evidence: reviewer identity and time.
- 4
4. Signature
The surgeon attests and signs. Control: organizational attestation and sign-off. Evidence: signed note linked to the draft version and interaction records.
- 5
5. Amendment
Corrections after signature follow the organization's amendment procedure. Control: documented process owned by health information management and compliance. Evidence: amendment record tied to the original note.
- Document the permitted note types and the models and tools the workflow may reach.
- Map the five workflow stages and name a control owner and evidence artifact for each.
- Classify each control as runtime-enforced or organizational process, with a written rationale.
- Design the review step to surface laterality, implants, specimens, complications and counts.
- Specify evidence fields and agree retention with legal and compliance.
- Establish an incident process and a governed change procedure for the approved model and tool list.
Separate pre-deployment decisions from runtime enforcement
A policy document describes intent. It does not govern what happens inside a specific AI interaction. Governing AI-drafted operative notes requires two kinds of decisions, and conflating them is the most common design error.
The first kind is made once, before deployment, and revisited on a schedule. These decisions define which note types the AI may draft, which models and tools the workflow may reach, what the review policy requires, who owns each stage, and how changes to the approved list are approved. They are organizational decisions, and a governance committee owns them.
The second kind is applied during each interaction. These are inference-time actions: allow the request, redact PHI before it reaches a model or tool, block an out-of-policy action, or escalate for human approval. They turn the pre-deployment decisions into behavior. Without them, the approved model list is a document that nothing enforces.
The practical test is whether a rule can be evaluated against the data and actions in a single interaction. If it can, it belongs in runtime enforcement. If it depends on professional judgment, it belongs in process.
Keep the surgeon's signature as the accountability anchor
As of November 2026, 42 CFR 482.51 requires that an operative report describing techniques, findings, and tissues removed or altered be written or dictated immediately following surgery and signed by the surgeon. The rule names the surgeon as the accountable signer and does not say how an AI-drafted note should be handled. Counsel should confirm how it applies to AI drafts, but the design implication is straightforward: treat the AI output as an unsigned draft until the surgeon has reviewed and signed it.
This matters because drafts can omit or misstate procedural details. Laterality, implants, specimens, complications and counts deserve explicit attention in review, alongside the techniques, findings and tissues the rule names. A fluent draft can read as complete while being wrong in one of these fields, so the review step should be designed to surface them rather than rely on a general read-through.
Runtime enforcement does not validate clinical accuracy and should not be described as doing so. It can constrain the data and actions in each interaction. It cannot confirm that the right side was operated on or that the implant record is correct. That verification stays with the clinician, which is why attestation and sign-off are process controls and not features to be automated away.
Runtime enforcement versus organizational process
Use this split as a starting point when assigning each control. The boundary will vary by organization, but the reasoning should be documented.
- PHI exposure to models and tools: Runtime: redact before the data leaves the permitted boundary. Process: define which data categories each use case may send.
- Which models and tools the workflow can reach: Runtime: allow approved models and tools, block the rest. Process: approve and change the list through a governed review.
- Out-of-policy actions: Runtime: block the action or escalate it for human approval. Process: define the policy and the escalation triggers.
- Accuracy of the final note: Runtime: not validated. Process: surgeon review, attestation and signature.
- Training and amendments: Runtime: not applicable. Process: clinician training and the amendment procedure.
Define the evidence the program must produce
An auditor or regulator will ask what the AI was allowed to do, what it produced, and who approved it. Evidence should answer those three questions without reconstruction. For each interaction, retain the policy version applied, the model and tools used, the decision taken (allow, redact, block or escalate), the categories of content redacted or blocked, the approver and time of approval, and a durable link to the draft version and the final signed note. Record redaction by category, so the evidence record does not itself become a second store of PHI.
Retention needs a deliberate decision with legal and compliance. Where the EU AI Act applies, it requires high-risk AI systems to allow automatic recording of events over their lifetime, and logs must be kept for a period appropriate to the intended purpose, at least six months unless other law provides otherwise. That figure is a minimum. It is not a medical record retention period, and whether an operative-note drafting tool is high-risk needs a separate, documented determination. The Act also sets a serious-incident reporting outer limit of 15 days after the provider or deployer becomes aware of the incident, so the incident process should be able to meet that clock if the Act applies.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. That enforcement generates audit-ready evidence mapped to HIPAA, the EU AI Act and NIST, and it can be self-managed in the customer cloud with configurable retention and data residency. Trussed does not validate clinical accuracy, and it does not replace clinician attestation.
Governance model for AI-drafted notes
Decided before deployment
Permitted use cases, approved models and tools, review policy, ownership.
Enforced at runtime
PHI redaction, model and tool restrictions, blocked actions, escalation.
Owned by the clinician
Review, attestation and signature of the final note.
Retained as evidence
Policy decisions, redactions, blocks and approvals linked to the signed note.
Define control points you can evidence
See how policy enforcement in the execution path produces audit-ready evidence for AI-drafted clinical documentation.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.