How to prove the AB 3030 licensed provider review exemption applied to an AI-drafted message
To rely on the AB 3030 exemption, you need to show, for each specific message, that it was generated by generative AI and was read and reviewed by a human licensed or certified health care provider. A message-sent flag or a policy statement does not show that. A single per-message record should tie together the AI draft, the named reviewer and their licensure status, the review action and timestamp, the relationship between the sent text and the draft, and the routing outcome. Enforcing review in the execution path produces that record when the review happens, instead of reconstructing it later.
The five evidence elements each message needs
What the exemption asks you to demonstrate
California Health and Safety Code section 1339.75 (AB 3030) sets requirements in subdivision (a) for certain generative AI patient communications, including clear instructions describing how a patient may contact a human health care provider or other appropriate person. Subdivision (b) provides that if a communication is generated by generative AI and read and reviewed by a human licensed or certified health care provider, the subdivision (a) requirements do not apply.
An organization that relies on subdivision (b) is therefore relying on two facts about each message: it was AI-generated, and a qualifying human read and reviewed it. The statute defines health care provider by reference to a person who is licensed, so the reviewer's status is part of what you must be able to show. If you cannot prove the exemption for a given message, you are left with the subdivision (a) content as the only defensible position. As of November 2026, the published text of section 1339.75 shows no reported amendment, delay or repeal.
Enforcement exposure depends on facility type. Violations by licensed health facilities fall under the enforcement mechanisms in Article 3 (commencing with Section 1275) of Chapter 2, and violations by licensed clinics under Article 3 (commencing with Section 1225) of Chapter 1. This guide sets out an evidence standard and is not legal advice. Counsel should confirm how 'read and reviewed' applies to your workflows, including any expectation about timing.
Where common approaches fall short
Most deployments hold general workflow records. These are useful operationally, but each leaves a specific gap when the question is whether the exemption applied to one message. Showing that review preceded the send is a prudent evidentiary choice. The visible statutory text does not quote a timing rule, so confirm the position with counsel.
- Message-sent flag: Shows a message went out. It does not identify a reviewer, a review action or the final content relative to the draft.
- Policy statement that clinicians review drafts: Describes intent. It is not evidence that a particular message was reviewed by a particular person.
- General application logs reconstructed after the fact: May contain fragments, but linking draft, reviewer, action and final text across systems is manual, and gaps surface only during an audit.
- Approval recorded in a separate tool: Shows an approval exists, but it may not bind to the exact content that was sent, or to a credential check made at that moment.
- Per-message record created at enforcement: Captures the review event when it occurs, keyed to one message, so evidence does not depend on later reconstruction.
Producing the evidence by enforcing review in the execution path
The structural choice is between proving review after the fact and making release of the message depend on a review that generates the record itself. If the system holds an AI-drafted message until an eligible reviewer acts, the approval, edit or rejection becomes the evidence, created at the moment of enforcement.
Trussed enforces policy in the execution path and can escalate an individual interaction for human approval before it proceeds, producing audit-ready evidence as a by-product of that enforcement. Its scope here is the individual interaction or tool call that carries the message, not an end-to-end trace of a wider workflow. Reviewer credentials still need an authoritative source owned by your organization, and the policy should define which reviewers are eligible.
The routing design matters as much as the approval step. Decide in advance what happens when no eligible reviewer approves: the message is blocked, or it is sent with the subdivision (a) disclaimer and contact instructions. Recording that outcome closes the gap where a message leaves without either review or the required content.
Limits of an approval event, and what to set yourself
An approval event shows that a reviewer acted. It does not by itself prove a substantive read, and any control that records only a click has this limit. The words 'read and reviewed' are the statutory test, so define what counts as review in your environment, for example the reviewer opening the full draft and then approving, editing or rejecting it, and record that interaction. Where an edit occurs, retaining both draft and final text gives stronger evidence of engagement than an approval alone.
Tamper resistance and retention are also yours to set. Decide who can alter or delete evidence records, how alteration is detected, and how long records are kept. Retention periods are not prescribed in the statute's text, so assign an owner and document the rationale. Configurable retention and data residency, and self-managed deployment in your own cloud, help keep these records under your control.
The Attorney General's advisories say healthcare AI systems should be tested, validated and audited. Periodic audit of the review control itself, not only of the messages, is consistent with that general expectation.
A practical sequence for compliance teams
Work through these steps in order. The last one is the test that tells you whether the earlier ones produced usable evidence.
- Inventory communications: List which patient communications are generative-AI-generated, and which rely on the subdivision (b) exemption rather than subdivision (a) content.
- Define reviewer eligibility: Map eligibility to the statutory term for a licensed or certified health care provider, and name the credential source and its owner.
- Define what counts as review: Specify the reviewer actions that qualify and the timestamp that must be captured.
- Set routing rules: Decide whether unreviewed messages are blocked or sent with subdivision (a) content, and record the outcome for every message.
- Set integrity and retention: Assign ownership for alteration controls and retention of the per-message record.
- Sample and reproduce: Pick sent messages at random and try to reproduce draft, reviewer, action, timestamp and final content from records alone.
One message, one evidence record
AI draft origin
Identifies the message as generative-AI output and retains the draft.
Reviewer and licensure
A named person, checked against licensed or certified status.
Review action and time
Approve, edit or reject, with a timestamp.
Sent content link
Shows how the final text relates to the draft.
Routing outcome
Records whether the message was sent, held or sent with the required disclaimer.
Produce per-message evidence at the point of enforcement
See how policy enforced in the execution path, including escalation for human approval, generates audit-ready evidence for individual AI interactions in a regulated enterprise.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.