How to Block Sensitive Diagnoses in AI-Drafted Messages Sent to Proxy and Caregiver Accounts
Enforce policy at runtime, between the model's drafted output and message delivery. Scan the drafted text for sensitive diagnosis categories, evaluate the recipient's role and authorized scope, then allow, redact, block or escalate for human approval. Record the policy version, recipient role, detection result and action for every message. Prompt instructions and after-the-fact review cannot do this, because the draft is generated and sent in real time.
Criteria for validating the implementation
- Every drafted message passes through enforcement before delivery, with no bypass path for any channel.
- Test cases cover each category using direct terms, paraphrases, abbreviations and implied references.
- Test cases cover each role, including unresolved or stale role data, which must take the restrictive default.
- False negative and false positive rates are measured on your own corpus and reviewed against an accepted threshold.
- Escalated messages follow the defined hold, approve and reject paths, and the approver is recorded.
- Each enforcement record shows policy version, role, detection result and action, and can be exported for audit.
Where disclosure happens in the drafting flow
A typical drafting flow retrieves context from the chart, builds a prompt, calls a model and hands the output to a delivery service. Disclosure of a sensitive diagnosis occurs at the last step, when text containing it is delivered to a recipient who may not be authorized to see it. The cause can sit earlier. Retrieved context may include a mental health, HIV, substance use or reproductive health diagnosis, and the model may repeat it or imply it in a summary, a visit recap or a reminder.
Two common controls fail at this point. Prompt instructions are probabilistic: they ask the model not to mention certain topics, but nothing verifies the output. Post-send review finds the problem after the message has been read. The Privacy Rule's general limit, as of November 2026, is that a covered entity or business associate may not use or disclose PHI except as permitted or required by Subpart E of 45 CFR Part 164 or by subpart C of part 160. HHS also describes the minimum necessary standard as requiring covered entities to limit unnecessary or inappropriate disclosure. A control that acts only after delivery does not limit the disclosure it is meant to prevent.
The reliable control point is the boundary between the drafted output and delivery. The system knows the draft text and the recipient at that moment, and it can still stop the message.
Define a recipient-aware policy
A recipient-aware policy has three inputs: the sensitive categories, the recipient role and scope, and the action for each combination.
Start with explicit category definitions. The brief's examples are mental health, HIV status, substance use and reproductive health. Write each category as a definition your detection layer can be tested against, including the diagnoses, medications, procedures and abbreviations that signal it. Confirm with counsel which categories carry legal protection in your jurisdictions and on what terms.
Next, model the recipient. Patient, personal representative, proxy and caregiver should be distinct roles, and each should carry an authorized scope. The authority behind each role comes from legal review and from the account data your identity systems hold, so decide who owns that data and how fresh it must be at evaluation time. A stale or missing role should resolve to the most restrictive treatment, not the most permissive.
Finally, map every category and role pair to an action and record the legal basis for the mapping. Set the threshold at which a case goes to a human reviewer, and name the authority who approves escalated messages. This keeps the policy reviewable as a document, separate from the code that enforces it.
- Category definitions with example terms, abbreviations and implied references
- Role model: patient, personal representative, proxy, caregiver, with scope per role
- Action per category and role: allow, redact, block or escalate
- Escalation threshold and named approving authority
- Default for unknown or unresolved recipient role
Choosing allow, redact, block or escalate
Each action has a cost, and the right choice depends on the category, the role and how much the message depends on the sensitive content. Treat the choice as a risk response applied per case rather than one global setting.
- Allow: Use when the recipient is authorized for the category. Lowest friction, but only as reliable as the role data behind it. Record the decision as evidence, not only the denials.
- Redact: Use when the message has legitimate value without the diagnosis. Risks: surrounding text may still imply the diagnosis, and removal can change clinical meaning.
- Block: Use when the recipient is clearly not authorized and the sensitive content cannot be separated. Risks: a legitimate message may be delayed, so define what the sender sees and what happens next.
- Escalate for human approval: Use for ambiguous detection or unclear authorization. Adds latency and reviewer workload, so define whether the message is held, and for how long, while it awaits approval.
Detection limits and failure modes
NIST AI 600-1 includes an action to detect PII or sensitive data in generated output, and that is the right place to apply detection: on the drafted text itself. Generative output can be inaccurate, and a draft may state or imply a diagnosis that is not in the record. Scanning only the retrieved source data would miss that case.
Detection methods have known limits. Government guidance cautions that keyword blocking is limited for terms with both harmful and benign uses and is easy to evade. NIST makes that point in the context of synthetic content, not PHI, but it carries over by analogy: a term list can miss paraphrases, abbreviations and implied disclosures, and it can over-block benign mentions. An FTC report to Congress says AI usually does not detect particular content well, and that blocking more content can suppress legitimate communication.
Plan for both error types. A false negative is a disclosure. A false positive is a blocked or redacted message that a caregiver needed. Rules and classifiers can be combined, with rules covering the unambiguous terms and a classifier covering paraphrase, but no published accuracy figures for diagnosis categories in patient messages should be assumed. Measure detection performance on your own message corpus and treat the results as validation evidence.
Evidence an auditor will look for
Enforcement in the execution path produces evidence as a by-product, because the system that makes the decision also writes the record. That is different from reconstructing events from application logs after the fact.
A useful enforcement record answers four questions: which policy applied, which recipient role was evaluated, what was detected, and what action was taken. As a design recommendation, each record should carry the policy version, the recipient role and scope as resolved at that moment, the detected categories, the action, the time, and for escalations the approver and outcome. Confirm required fields and retention periods with your compliance team. Keep records of allowed messages as well as blocked ones, since an auditor needs to see what was permitted and why.
Trussed is the runtime enforcement layer for this pattern. It enforces policy in the execution path and can allow, redact, block, rate limit or escalate for human approval, at the level of individual interactions and tool calls. It generates continuous audit-ready evidence from that enforcement, mapped to HIPAA among other frameworks, and it can be self-managed in the customer cloud with configurable retention and data residency. Its scope is the individual interaction, not transaction-level or workflow-level governance.
Enforcement point in the message-drafting flow
Draft
The model generates message text for a portal or similar channel.
Evaluate
Detect sensitive diagnosis content and resolve the recipient's role and scope.
Enforce
Allow, redact, block or escalate for human approval before delivery.
Record
Write the policy, role, detection result and action as evidence.
See runtime enforcement for AI-drafted messages
Talk to an expert about enforcing recipient-aware policy in the execution path and the audit-ready evidence it produces.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.