How to Prevent AI-Drafted Patient Messages from Disclosing Unreleased Results
Enforce result-release status in the execution path at two points: when context is retrieved or a tool is called, and when the draft is produced. Retrieval filtering keeps unreleased results out of model context. Output inspection catches mentions or implications that still appear. Each decision (allow, redact, block, escalate to a clinician) is recorded with its inputs, so you can show what was withheld and why. Prompt instructions alone do not provide this guarantee.
- {"title":"Retrieval context","description":"Results, notes or reports pulled into the prompt because they are relevant to the patient's message, regardless of release status."}
- {"title":"Tool and API calls","description":"FHIR or other API calls the assistant makes on its own, where the response may include results the patient cannot yet see."}
- {"title":"Clinician-visible chart data","description":"Fields the assistant can read because the drafting clinician can, such as result values, interpretations, pathology and imaging impressions, and free-text notes."}
- {"title":"Inference from partial context","description":"A draft that implies a result without naming it, for example reassurance such as \"your recent scan looks fine\", based on permitted context or on the shape of the conversation."}
- Layer 1: release-status check on retrieval and tool-call responses
- Layer 2: output inspection of the draft
- Layer 3: redact, block or escalate actions
- Layer 4: mandatory clinician approval before sending
Evidence to retain for audit
- AI version, timestamp, patient ID and responsible clinician
- Which items were retrieved or returned by tool calls, and the release status evaluated for each
- The policy and policy version applied, and the reason a hold was in effect
- The action taken at each stage: allowed, redacted, blocked or escalated
- The output inspection result and, for escalations, the reviewing clinician and their decision
- Retention settings that match organizational policy
Why this is a policy problem and not a prompt problem
A drafting assistant that reads from the EHR can see results that are final in the chart but still in a release hold, awaiting clinician review, or otherwise not released to the patient. If the draft mentions, summarizes or implies one of those results, the patient may learn it before the clinician has communicated it.
A prompt that says "do not discuss unreleased results" is an instruction to a probabilistic system. It does not remove the data from context, and it cannot be audited as a control. The reliable approach is to enforce policy where data is accessed and where output is produced, outside the model's own judgment.
The regulatory frame is narrow, and it cuts both ways. As of November 2026, the information blocking regulations, in effect since April 5, 2021, do not require actors to proactively release electronic health information. They do require a timely response once a patient requests it, and ONC notes that in practice a patient may see results in parallel with the ordering clinician. ONC also states that practices that interfere with, prevent or materially discourage access to EHI by automation technologies, including agentic AI, could implicate the rules. Following an individual patient's request to delay release would likely not be interference, but the delay should be no longer than necessary. Some states also require specific approaches or delays for life-altering diagnostic results, such as a new cancer diagnosis.
The design goal is therefore precise: withhold only what a defined hold covers, for only as long as it applies, and be able to show it.
Treating release status as a policy input
The controls above depend on a reliable release-status signal, and that signal may not exist. ONC reports that many EHRs lack built-in functionality to capture and honor individual patient preferences on result release, and that many organizations release all results automatically once they are available or a state-law delay expires. Verify the source field for each data feed rather than assuming a structured status is exposed to the assistant.
Start with an inventory of every source the assistant can reach: results, notes, imaging, pathology and tool endpoints. For each, record how release status is determined and whether the enforcement layer can read it. Then classify why a result is withheld, because the reasons carry different rules: a patient-requested delay, a state-law delay, or an organizational hold. Involve compliance counsel in defining hold rules, exceptions and state overlays.
Time-bound every hold. Track the start, the agreed conditions and the release trigger, since extended or unnecessary delays may be evidence of intent to interfere. Define behavior when status is missing, stale or conflicting across systems. Failing closed with escalation is a sound default, but weigh it against over-withholding, which carries its own information blocking exposure.
Some result categories deserve a stricter route. SAFER guidance notes that results such as HIV status or a cancer diagnosis are better discussed in person or by phone than by asynchronous electronic messages. Consider routing drafts that touch those categories to the clinician with no AI-generated result language at all.
Finally, test with scenarios where results are final but unreleased, including indirect references. Define pass criteria internally, and rerun the scenarios whenever the model, prompt, retrieval configuration or hold policy changes.
Where Trussed fits
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, on individual interactions and tool calls. For this use case, that means release-status policy can be applied when a tool call returns result data and when a draft is produced, with actions to allow, redact, block or escalate for human approval. Enforcement happens at runtime, not after the fact.
Enforcement generates audit-ready evidence mapped to HIPAA. Trussed supports self-managed deployment in the customer cloud, with configurable retention and data residency, which keeps PHI-adjacent evidence inside the organization's own environment. Trussed runs in production with healthcare and pharma organizations.
The pattern in this guide does not depend on any one product. The essential requirements are enforcement outside the model, at both retrieval and output, with a decision record for every interaction.
Two enforcement points, one evidence trail
Retrieval and tool calls
Check release status before result data enters model context.
Draft output
Inspect the draft for mentions, summaries or implications of withheld results.
Action
Allow, redact, block or escalate to clinician review.
Evidence
Record the decision, the policy applied and the responsible clinician.
Review your enforcement points before drafts reach patients
Walk through your retrieval paths, tool calls and evidence requirements with a Trussed expert.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.