Implementation Guide

    Governing AI-Drafted Letters of Medical Necessity

    Govern AI-drafted letters of medical necessity by mapping each stage of the drafting workflow (input assembly, model call, record retrieval, draft output, clinician approval, submission) to one control and one evidence artifact. Enforce the data and release controls where the model and tools are called, using allow, redact, block and escalate for human approval. Keep clinical accuracy, policy authorship and payer criteria selection as procedural controls owned by people. Retain per-interaction evidence of the policy applied, data removed, approver decision and output released.

    Map each drafting stage to a control and its evidence

    1. 1

      1. Input assembly

      Control: an allowlist of permitted data fields per letter type, with a decision per field to allow, redact or block. Evidence: the fields submitted, and the fields redacted or blocked.

    2. 2

      2. Model call

      Control: only approved models receive the request, and PHI filtering is applied before the call. Evidence: the policy version applied, the model called and the outcome.

    3. 3

      3. Tool or record retrieval

      Control: restrict which tools and records a call may reach. Treat this as separate from the model call, since PHI can enter through retrieved records. Evidence: each tool call, the records requested and the decision.

    4. 4

      4. Draft output

      Control: output constraints, such as requiring claims to trace to the patient record and flagging citations for verification. Evidence: the output produced and any constraint result.

    5. 5

      5. Clinician approval

      Control: the draft is held until a named clinician records an approve, edit or reject decision. Evidence: approver identity, decision and time.

    6. 6

      6. Submission

      Control: release is permitted only when an approval record exists for that letter. Evidence: the final text released and the approval it references.

    Evidence to retain for each AI interaction

    • Policy name and version applied to the interaction
    • Data fields allowed, redacted or blocked, and the reason
    • Models and tools called, and records retrieved
    • Escalation event, if one was triggered
    • Approver identity, decision and timestamp
    • Output released, and the submission it was attached to

    Separate policy definition from policy enforcement

    A letter of medical necessity supports a prior authorization request or an appeal of an adverse determination. When generative AI drafts it, the organization takes on three exposures: PHI reaching a model or tool without authorization, unsupported clinical claims or fabricated citations appearing in the text, and a letter leaving the organization without clinician review. These are risks to design against. They are not incidents this guide asserts have occurred at any given organization.

    Most programs start with a written policy. The weakness is that a policy document does not act on any individual model call. NIST AI RMF Govern 1 expects policies, processes, procedures and practices to be in place, transparent and implemented effectively. That last phrase is the practical test: for each rule, can you say where it is applied when a draft is generated?

    Three roles should stay distinct. Compliance defines the policy. A runtime layer enforces the parts that can be expressed as rules on an interaction. A clinician reviews and approves the content. Each should leave its own record for each letter.

    Regulatory and framework anchors

    Govern 1.1 asks that legal and regulatory requirements involving AI are understood, managed and documented. Assign an owner and keep the mapping current. Compliance and legal should map each workflow stage to the organization's HIPAA obligations, including whether a given model or tool is permitted to receive PHI at all. Govern 1.2 integrates trustworthy AI characteristics into policies and procedures, and Govern 1.3 sets the level of risk management activity by organizational risk tolerance. In practice, classify letter types by risk. An appeal of an adverse determination warrants tighter approval requirements than a routine request. The NIST Generative AI Profile (action GV-1.1-001) adds that generative AI use should align with applicable laws, including those on data privacy.

    Appeals matter here for a specific reason. Under 29 CFR 2590.715-2719, adverse benefit determinations include those involving medical judgment, such as medical necessity, appropriateness, health care setting, level of care or effectiveness. Because the decision under appeal turns on medical judgment, clinician accountability for the letter's content should be explicit in your policy.

    Payer-side context is also useful. A CRS report describes CMS requirements that Medicare Advantage medical necessity determinations consider the enrollee's medical history, physician recommendations and medical notes, and a CRS product states that an expected adverse medical necessity determination must also be reviewed by a physician or other qualified professional. Those provisions govern payers, not provider-drafted letters. They are still a reasonable reason to ground every claim in your letters in the patient record.

    What can be enforced at runtime and what stays procedural

    Controls that can be written as rules on an interaction belong in the execution path. These include filtering PHI before a model or tool call, restricting which models, tools and records a call may reach, rate limiting, and holding a draft until an approval is recorded. Each resolves to a concrete action: allow, redact, block, or escalate for human approval. Because the rule runs at the point of the call, it applies to every interaction, not only to those someone remembers to check.

    Other controls cannot be reduced to a rule and should remain procedural. The clinician's judgment on whether the letter is clinically accurate and persuasive is the clearest example. Policy authorship, workforce training, and the decision about which payer criteria apply also stay with people. Runtime enforcement can require that a clinician decides. It cannot make the decision.

    The tradeoff is friction. Strict blocking and mandatory escalation slow drafting, and over-redaction can remove clinical detail the letter needs. Set the allow, redact or block decision per field and per letter type, and revisit it using reviewer feedback rather than loosening it ad hoc.

    Where runtime enforcement fits

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. For this workflow, that means PHI can be allowed, redacted or blocked before a model or tool call, tool and record access can be restricted, and a draft can be escalated for human approval before release. Enforcement generates audit-ready evidence, mapped to HIPAA and NIST, so the policy applied and the action taken are enforced at runtime, not reviewed after the fact. Deployment is self-managed in the customer cloud, with configurable retention and data residency.

    These capabilities address stages one through three, and the release gate at stages five and six. They do not replace clinician review or the compliance team's ownership of policy. Correlating interactions to a specific letter remains a workflow design decision on the organization's side.

    When evaluating any approach, ask where each policy is enforced and what is merely documented, whether removed PHI is recorded, whether a draft can be held for approval with the approver and decision captured, and how long evidence is retained and retrievable.

    Drafting workflow: control and evidence by stage

    Input assembly

    Permitted data fields defined; PHI allowed, redacted or blocked.

    Model call and retrieval

    Models and records restricted per policy; each call recorded.

    Draft output

    Output constraints applied before the draft reaches a reviewer.

    Approval and submission

    Draft held until a clinician decision is recorded.

    Review your drafting workflow against these controls

    Talk through where policy is enforced today, where it is only documented, and what evidence you can produce for a given letter.

    Talk to an Expert