Governing AI-Drafted Referral Letters Sent to Outside Providers
Govern AI-drafted referral letters in four separate layers: written policy on what data may be used and disclosed, runtime enforcement that allows, redacts, blocks or escalates during the drafting interaction, human approval before anything leaves the organization, and retained evidence of what was permitted, produced and approved. Technical controls cover the first part of that path. Training, legal basis and retention schedules remain organizational work.
Why a referral letter is a consequential external output
A referral letter is different from most AI-drafted text because it leaves the organization and reaches a party you do not control. Once an outside provider has it, an error or an over-disclosure of health information cannot be recalled. The drafting assistant may have been given more patient data than the letter needed, may have stated a detail the record does not support, or may have addressed the wrong recipient.
Most drafting assistants run without enforced controls on three points: what data enters the prompt, what content appears in the letter, and whether a person approved it before sending. Review after the fact does not close that gap, because the letter has already been transmitted. Afterward, the organization often cannot show a regulator or auditor what the AI was permitted to include, what it actually produced, and who released it.
The legal frame starts with the general rule in 45 CFR Part 164: a covered entity or business associate may not use or disclose PHI except as permitted or required by subpart E of that part or by subpart C of part 160. Each category of information in a referral letter therefore needs a disclosure basis that your legal team has confirmed. As of November 2026, the HIPAA Security Rule proposal published in January 2025 and the December 2025 proposal to revise the ONC transitions of care criterion are proposals. Treat them as pending, not binding.
Step 1: Define policy and data-minimization rules
Start by listing every data element a referral letter might contain: demographics, diagnosis, medications, history, prior results, free-text notes, and contact details. For each element, record the disclosure basis the organization relies on, with legal review. Then define what each recipient type may receive. A specialist receiving a treatment referral and an administrative recipient confirming an appointment should not receive the same fields.
The result is a policy that a system can evaluate, not a paragraph in a handbook. Each rule should state the condition and the action: this field is allowed for this recipient type, this identifier is redacted, this category of content is blocked, and this combination requires human approval. Rules written this way can be applied to the prompt before the model sees it and to the draft before anyone sees it.
Also design for the failure modes you expect from drafting tools: invented clinical details, disclosure to the wrong recipient, and omitted context that changes meaning. Each maps to a different control. Prompt restrictions limit what the model can draw on. Recipient checks address misdirected letters. Mandatory review addresses invented or missing content, because those errors usually need a clinician's judgment to catch.
Step 2: Separate what runtime can enforce from what stays organizational
Runtime enforcement acts in the execution path of the drafting interaction, before the letter exists in final form or leaves the organization. It can only enforce rules that are explicit enough to evaluate. Everything else depends on people and process. Recording which is which lets an auditor see what was enforced and what relied on human judgment.
- Data entering the prompt: Runtime. Allow, redact or block fields outside the permitted set for the recipient type.
- Content in the generated letter: Runtime. Redact or block identifiers and content categories the policy disallows.
- Approval before release: Runtime and organizational. The interaction can escalate and hold for approval. The quality of the review depends on the reviewer.
- Evidence generation: Runtime. Policy applied, decision taken and output are captured as enforcement happens.
- Disclosure basis and policy content: Organizational. Legal and compliance decide what the rules should be.
- Training, vendor review, retention schedule, halt decisions: Organizational. Assign owners and document the decisions.
Step 3: Require human approval before external release
No source requires a human approval step for every AI-drafted letter, so this is a recommended organizational control. For letters that leave the organization, it is a sensible default. Decide whether approval applies to every external letter or only to letters meeting defined conditions, such as a new recipient, a sensitive category of content, or a redaction having occurred during drafting.
The approval must sit between drafting and transmission, not after it. In practice the interaction is held, the approver sees the draft and what policy did to it, and release happens only on an explicit action. Escalation for human approval is the runtime mechanism. Competent review is the organizational part: approvers need to know what to check, including facts against the record, the recipient, and whether anything unnecessary remains.
Assign ownership for exceptions and for the decision to pause the drafting tool. NIST AI 600-1 includes an action to devise a plan to halt development or deployment of a generative AI system that poses unacceptable negative risk. A documented halt plan for the referral use case fits that action, with a named person who can invoke it.
Step 4: Retain evidence of what was allowed and what happened
An auditor asking about a specific letter will want four things: the policy in force at the time, what data was permitted into the prompt, the output the AI produced, and who approved the release and when. Capture these for each letter, and keep the policy version alongside the decision so the record shows what was permitted, not only what occurred.
Evidence that is assembled later from scattered systems is weaker than evidence produced as part of enforcement. If the decision to redact a field or hold a letter for approval is itself recorded at the moment it is made, the record shows what the controls did, not what the organization believes they did.
Set retention periods with your legal and compliance team and apply them as a configured setting, not an informal habit. Also decide how evidence is exported and mapped to HIPAA so that a request from an auditor can be answered from one place. Because letters contain health information, apply to the evidence store the same access and residency requirements you apply to the source records.
Where Trussed fits
Trussed enforces policy in the execution path of individual AI interactions and tool calls: allow, redact, block, rate limit, and escalate for human approval. For a referral drafting workflow, that means the rules from Step 1 are enforced at runtime, not reviewed after the fact, and an escalation can hold a letter until an approver acts.
Enforcement generates audit-ready evidence mapped to frameworks including HIPAA, with configurable retention and self-managed deployment in the customer cloud. Scope matters. Trussed enforces and records at the level of individual interactions and tool calls. It does not replace the organizational work described above: the disclosure basis, the policy content, reviewer training and the retention decision remain with your compliance function.
Whether or not you use Trussed, the evaluation question for any tool is the same: is the decision made before the letter leaves, in the path of the interaction, or only recorded afterward?
Four layers of referral letter governance
Define policy
Disclosure basis and data-minimization rules by recipient type.
Enforce at runtime
Allow, redact, block or escalate during the drafting interaction.
Approve before release
A named person authorizes each external letter.
Retain evidence
Policy applied, output produced, approver and release action.
Review your referral letter controls
Walk through how policy enforcement and audit-ready evidence apply to AI-drafted external correspondence in your regulated enterprise.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.