Implementation Guide | Healthcare

    How to Review AI-Drafted Discharge Instructions Before Patients Receive Them

    Treat review as a control that is enforced before release, not a courtesy check. Define which discharge content is AI-drafted, apply fixed content and PHI criteria, route every draft to a named human reviewer for approval, release only on a recorded approval, and retain evidence of each decision. Enforcing policy at runtime, with allow, redact, block or escalate for human approval, makes the review step mandatory instead of optional.

    Route outputs through the workflow

    1. 1

      1. Classify the draft

      Identify the content types in the output and apply the matching review tier.

    2. 2

      2. Run PHI and content checks

      Detect PHI beyond what the communication requires and apply automated content rules before a human sees the draft.

    3. 3

      3. Act on findings

      Allow clean output to proceed to review, redact excess PHI, or block output that fails a rule.

    4. 4

      4. Escalate for human approval

      Route the draft to the named reviewer with the source record context and the checklist.

    5. 5

      5. Record the decision

      Capture approval, rejection or edits, with reviewer identity and timestamp.

    6. 6

      6. Release or return

      Release only the approved version. Return rejected or edited drafts to the start of review.

    • Accuracy: diagnoses, therapies, dosages, lab findings, allergy and contraindication information and follow-up plans match the patient record.
    • Omissions: no missing therapy or post-discharge instruction, side-effect or risk warning, or follow-up.
    • Safety: warning signs and instructions on when to seek care are present and clear.
    • Readability: legible, non-technical language, which aligns with what CMS discharge planning survey guidance looks for.
    • PHI: the draft contains only the PHI needed for this patient and this communication, and no other patient's data.
    • Personalization: instructions reflect this patient's condition and plan, not generic text.
    • AI version and date and time of use
    • Patient ID and the responsible clinical user
    • Reviewer identity and role at the time of approval
    • Edits made to the AI draft, kept alongside the final approved version
    • Approval, rejection or escalation decision with timestamp
    • Policy applied, including any redaction or block action taken

    Why informal clinician checks are not enough

    Many organizations already expect a clinician to glance at an AI draft before it goes to a patient. The problem is that this check usually happens outside the AI execution path. Nothing prevents a draft from being released without it, and nothing produces a consistent record that it occurred. For a compliance leader, that means the control cannot be enforced and cannot be evidenced.

    The risks are concrete. Medication, follow-up and warning-sign content can be wrong or missing. Drafts also carry PHI, so exposure can occur before anyone reviews the text. And when an auditor asks who approved a given instruction, an informal check offers no answer. HealthIT.gov SAFER guidance is direct on the first point: AI-generated recommendations should be reviewed and approved by humans who take responsibility for them before they are sent to patients. The same guidance expects a transaction log of AI use, including AI version, date and time, patient ID and the responsible clinical user.

    Start by scoping. List the discharge content types that AI drafts in your environment, such as medication changes, follow-up appointments, warning symptoms, diet and activity. Our recommendation is that all AI-drafted, patient-facing discharge content requires human approval. Risk tiers can set how deep the review goes, for example a fuller check for medication changes than for general activity advice, but they should not decide whether review happens at all.

    Assign reviewer roles and set release conditions

    The sources do not set a universal rule on who may approve, so this is a local decision that you should document. In one published protocol the treating physician performs the clinical validation. In another study, residents revised drafts and experts scored them. Whatever you choose, the approver must be a named person who accepts responsibility for the content, because SAFER guidance ties approval to humans who take responsibility.

    Separate two roles. The clinical reviewer checks content against the record. A privacy or compliance function owns the PHI criteria and the policy that governs release, and periodically reviews the records the process produces. Define what happens when the designated reviewer is unavailable, so that delay leads to an escalation path and not to an unreviewed release.

    Release conditions should be explicit and few. A draft is released only when the checklist is complete, a named reviewer has recorded an approval decision, and no open PHI finding remains. Any edit after approval returns the draft to review. Policies should also cover telling patients and clinicians, where possible, that AI-enabled systems are used, and delivery should respect confidential communication requirements under 45 CFR 164.522(b).

    Make review mandatory with runtime enforcement

    The difference between a courtesy check and a control is where it sits. If approval is a step in a separate process, a draft can reach a patient without it. If approval is a policy enforced in the execution path, the output cannot be released until the policy conditions are met.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with the actions allow, redact, block, rate limit and escalate for human approval. For discharge instructions, that maps onto the workflow above: output that passes policy proceeds, PHI that should not be present is redacted, output that violates a rule is blocked, and drafts that require a clinician's decision are escalated for human approval. Review is enforced at runtime, not reviewed after the fact. The scope is individual interactions and tool calls, so the policy applies to each draft or call as it happens.

    Enforcement also generates the evidence. Trussed produces continuous audit-ready evidence from enforcement, mapped to HIPAA, including the audit-control expectations at 164.312(b). It runs in production with healthcare organizations and can be deployed self-managed in the customer cloud, with configurable retention and data residency. That matters when drafts contain PHI that should stay within your environment. The clinical criteria and the choice of approvers remain your decisions; the runtime layer makes them binding.

    Review workflow for AI-drafted discharge instructions

    Scope

    Identify which discharge content is AI-drafted and in review.

    Check

    Apply content criteria and PHI checks to each draft.

    Approve

    Route to a named reviewer; release only on recorded approval.

    Evidence

    Retain reviewer, edits, decision, timestamp and policy applied.

    Put the review step in the execution path

    Talk to us about enforcing human approval for AI-drafted patient communications and producing audit-ready evidence of each decision.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.