Healthcare AI runtime control

    Blocking AI-drafted patient messages that change medication doses

    Treat dose-change prevention as a pre-send enforcement decision, not a prompt instruction or a post-send review. Define a dose change in written policy, evaluate each draft in the path between generation and delivery against separate drug, dose, unit, frequency and change-verb signals, and return allow, redact, block or escalate for human approval. Record every decision with the policy version and any approver, so you can show the control operated.

    Place the enforcement point in the execution path

    1. 1

      Allow

      The draft carries no dose-change content, or only restates an existing instruction, and proceeds to send.

    2. 2

      Redact

      Remove the offending span and let the rest proceed. Use sparingly, because a message with a removed dose can still mislead.

    3. 3

      Block

      Stop delivery outright where the draft proposes a change no clinician has approved.

    4. 4

      Escalate

      Hold the draft and route it to a clinician for approval, edit or rejection before anything is sent.

    Evidence to record for every decision

    • Draft identifier and timestamp
    • Policy and rule version that applied
    • Outcome: allow, redact, block or escalate
    • Signals that triggered the rule
    • Approver identity and decision for any exception or escalation
    • Retention period and export path for auditors

    Why review after the fact does not work

    A draft that adds, removes or alters a medication dose, frequency or instruction can reach a patient as clinical direction nobody authorized. If the control is a periodic review of sent messages, it can only report on the message after the patient has read it. Prompt instructions that tell the model not to discuss doses reduce the likelihood of the behavior but do not give you a decision you can enforce or prove.

    The control therefore has to sit where the message can still be stopped. That is a runtime enforcement problem: an explicit decision, made on each draft, before delivery, with a record of what was decided and under which policy.

    Define what counts as a dose change

    Start with a written definition that clinical leadership signs off on. A workable definition covers any draft that adds, removes or alters a medication dose, a frequency, or an instruction about how or when to take a drug. Include instructions such as stopping, skipping, doubling or splitting, since these change therapy without naming a number.

    The harder boundary is restatement. A draft that repeats an instruction already in the patient's record is different from one that proposes something new. Decide whether your gate can tell the two apart. If it cannot, treat any dose-bearing content as in scope and accept more escalations. Test the definition against sample messages from your own portal traffic before you write any rules, and keep the examples with the policy so reviewers apply it consistently.

    Design rules around separate signals

    Evaluate each element as its own signal rather than one broad pattern. Federal data standards already treat dose and dose unit of measure as distinct elements of the medications data class, and a rule schema can follow the same separation. This is a design choice rather than a mandated format, but it makes rules easier to test and explain.

    Combine signals in a rule. A drug name alone is not a violation, and neither is a number. A drug, a dose value, a unit and a change verb appearing together is a strong candidate.

    • Drug names: Normalize brand, generic and ingredient names to a standard vocabulary such as RxNorm so one rule covers every way a drug can be written.
    • Dose values and units: Match numbers with units such as mg, mcg, mL and tablets, including spelled-out numbers and ranges.
    • Frequency: Match patterns such as twice daily, every eight hours, at bedtime and as needed.
    • Change verbs: Match increase, decrease, double, stop, skip, switch, start and resume, along with phrasings that imply them.
    • Context where available: Patient-specific data such as age or lab values can sharpen decisions. Whether the gate can reach it is an architecture decision to settle early.

    Escalation, false positives and false negatives

    Escalation needs an owner. Define which clinician role receives held drafts, how long a draft may wait, and what the patient sees while it waits. Because blocking or delaying a message can raise information blocking questions, have counsel assess whether the preventing harm exception conditions can be met. The exceptions are voluntary, and practices that do not qualify are judged case by case.

    Expect both error types. Pattern matching is predictable but misses paraphrase. Terminology lookups improve drug matching but not intent. Classifier-based checks can catch looser wording but are harder to explain and need their own validation. Measure false positives and false negatives on your own message corpus, since rates elsewhere will not transfer. Tune toward escalation for high-risk rules, and treat a missed change as the more serious error.

    Re-test when models, prompts, templates or message workflows change, and on a fixed schedule. A January 2025 HHS proposal to amend the Security Rule would require testing of technical controls at least every 12 months. As of November 2026 it is a proposal, not current law, but it is a reasonable benchmark.

    Where Trussed fits

    Trussed enforces policy in the execution path, so a draft is evaluated at runtime, not reviewed after the fact. It applies allow, redact, block and escalate for human approval to individual interactions and tool calls, and generates audit-ready evidence from those enforcement decisions, mapped to HIPAA. It can be self-managed in your cloud, with configurable retention and data residency. Trussed runs in production with healthcare and pharma organizations.

    The rule content, the definition of a dose change and the clinical sign-off remain yours. The design above holds whichever tool sits at the gate.

    Pre-send gate for AI-drafted patient messages

    Draft generated

    The model drafts a portal reply or in-basket message.

    Policy evaluated

    Drug, dose, unit, frequency and change-verb signals are checked before send.

    Outcome applied

    Allow, redact, block or escalate for clinician approval.

    Evidence recorded

    Policy version, outcome, timestamp and approver are retained.

    Review your pre-send enforcement design

    Talk through where a dose-change gate fits in your message path and what evidence it should produce.

    Talk to an Expert