Review Tiers for AI-Drafted Patient Messages by Message Type
Define message-type categories, score each against five risk factors (clinical content, PHI presence, required patient action, error reversibility, regulatory exposure), and assign each category a review tier: automated release with policy checks, sampled review, or mandatory clinician or staff approval. Add escalation triggers that move a message up a tier, enforce the tier in the send path as allow, redact, block or escalate, and record the decision per message so an auditor can see what applied and what happened.
- Clinical content appears in a message type that is not classified as clinical.
- PHI appears that falls outside the fields permitted for that message type.
- The message requires a patient-specific action, such as a dose change, a deadline or a follow-up visit.
- The message is marked urgent or time-sensitive in existing message metadata.
- The classifier finds no matching type, or the type and content disagree.
- A policy check fails or cannot complete.
- Message identifier, message type, and the rule or context that produced the classification.
- Tier applied, policy version, and which escalation triggers fired.
- PHI detected in the draft, and whether it was redacted, retained or caused a block.
- Policy decision: allow, redact, block or escalate.
- Approver identity, timestamp, and any edits made before release.
- Release time and channel, plus the outcome of any sampled review.
Why one review standard fails
Healthcare organizations use AI to draft appointment reminders, billing notices, medication instructions, test result communications and portal replies. The clinical, privacy and regulatory risk of these messages differs widely. A single standard fails in one of two directions. If every draft needs human approval, the review queue becomes the bottleneck and the value of drafting erodes. If every draft is released automatically, high-risk messages reach patients with too little oversight.
Tiering is therefore a policy design problem before it is a staffing problem. The questions are which message types exist, which risk each carries, what review that risk justifies, and how the decision is enforced and proven for each individual message.
Two reference points frame the design. ONC's SAFER guidance recommends that AI-generated recommendations be reviewed and approved by humans who take responsibility for them before they are sent to patients. It also recommends separating critical and time-sensitive messages from routine or information-only ones. SAFER is guidance, not regulation, and the HIPAA Security Rule at 45 CFR 164.312 is regulation that applies to covered entities and business associates. Keep the two distinct when you document why a tier exists.
Message types mapped to review tiers
The table is a starting model, not a standard. SAFER's approval recommendation is framed around AI-generated recommendations, so whether a given administrative message falls under it is an organizational determination. Document that reasoning for each category rather than assuming it.
Assign the tier by the highest-risk factor present, not by an average. A message with low clinical content but a patient-specific deadline still carries the risk of the deadline.
- Appointment reminders and logistics: Low clinical content, limited PHI, easily corrected. Tier 1: automated release with policy checks that confirm only permitted fields appear.
- Billing notices and statements: Financial and PHI exposure, moderate regulatory exposure. Tier 2: sampled review, with higher sampling after template or model changes.
- Routine administrative portal replies: Variable content that is usually non-clinical. Tier 2, with escalation if clinical content is detected.
- Medication instructions: Clinical content, patient action required, errors hard to reverse. Tier 3: mandatory clinician approval before send.
- Test result communications: Clinical content, sensitive PHI, possible follow-up action. Tier 3: mandatory clinician or designated staff approval.
- Clinical portal replies: Free-form clinical advice or triage. Tier 3: mandatory clinician approval, with urgent items flagged as such.
Classify by rule, then check the content
Message-type classification should be deterministic where possible. Derive the type from the originating workflow, template or portal context rather than inferring it from the draft text with a model. A rules-based classifier is easier to explain to an auditor and easier to test.
ONC's HTI-1 preamble commentary points the same way for a related question. It says rules-based algorithms that route secure messages by message type would likely not be Predictive DSIs, while NLP models trained on message content and past responder behavior would likely be. That commentary concerns certification criteria and says "would likely". It is not a determination about HIPAA obligations, but it is a useful reason to document how your classifier works.
The type sets the baseline tier. A content check on the draft then tests whether the draft still matches the type. A reminder that contains a medication change, or a billing notice that includes diagnosis detail, should not keep its Tier 1 or Tier 2 status.
Enforce the tier in the execution path
A tier that exists only in a policy document is a statement of intent. To be defensible, the tier decision has to be made for each message before it is sent, in the path the message takes. Four outcomes cover the cases. Allow releases the draft. Redact removes PHI that the message type does not need and releases the remainder. Block stops the draft. Escalate holds it for human approval. Where review is required before sending, the approval step must sit in the send path, so that a draft cannot reach the patient without the approval being recorded.
This is the point where review after the fact stops being enough. Reviewing sent messages can find a problem, but it cannot prevent a PHI exposure or an unreviewed medication instruction from reaching a patient. Trussed enforces policy in the execution path at the level of individual interactions, with allow, redact, block, rate limit and escalate for human approval. Enforcement is at runtime, not reviewed after the fact. Trussed runs in production with healthcare organizations and supports self-managed deployment in the customer cloud, with configurable retention and data residency, which matters when drafts contain PHI.
Tier assignment can also be carried in urgency or priority fields that EHR messaging already uses, so reviewers see the tier in the queue they already work in.
Where per-message tiering stops
Tiering governs individual messages. It does not replace the organizational processes around them. SAFER frames patient and clinician awareness that AI is used, EHR documentation of patient-specific communication, and named accountability for approvals as policies and procedures. Reviewers who approve AI-drafted content must take responsibility for it, and that needs an owner, training and a defined escalation path.
Sampled review has its own limits. The sampling rate is a judgment, and any approval step can turn into a formality if volume is high or reviewers see little they disagree with. Set sampling rates deliberately, raise them after template, model or policy changes, and review the sampled outcomes as data about whether a category belongs in its tier.
Monitoring also bears on breach discovery. Part 164 treats a breach as discovered when it is known, or would have been known with reasonable diligence. Whether monitoring of drafts and sent messages affects when an organization would have known is an inference to confirm with counsel.
Finally, as of November 2026, the proposed Security Rule amendments published on 6 January 2025 remain proposals. Treat the current rule text as governing, and revisit the tier model if the rule is finalized.
Tier model at a glance
Tier 1
Automated release with policy checks
Tier 2
Sampled human review after policy checks
Tier 3
Mandatory clinician or staff approval before send
Escalation
Triggers that move a message to a higher tier
Evidence
Per-message record of tier, decision and approver
Turn tier policy into runtime enforcement
See how tier decisions are enforced per message in the execution path and captured as audit-ready evidence for HIPAA.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.