AB 3030 vs SB 1120: How Patient Disclosure and Utilization Review Rules Map to Different Runtime Controls
AB 3030 and SB 1120 are separate control problems. AB 3030 applies to health care providers that use generative AI to produce patient clinical communications, so the enforcement point is the outbound message, where disclosure can be checked. SB 1120 applies to health plans and insurers that use AI in utilization review, so the enforcement point is the decision or action, and who makes it. Each needs its own policy, its own runtime control and its own evidence. This page is not legal advice.
Evaluating runtime enforcement approaches
- Can it verify that a required disclosure is present in each AI-generated patient clinical communication before it is sent?
- Can it prevent an AI agent from executing a denial, delay or modification of care without licensed clinician approval?
- Does it record the escalation, the reviewer and the reviewer's decision for each action?
- What evidence does each control produce for an individual interaction or tool call, and can it be exported for an auditor?
- Are policies versioned and tied to the statutory requirement they address, with a way to update them when guidance or amendments change?
- Which obligations does the vendor treat as out of scope for runtime enforcement, and what must your organization supply?
Why one compliance item is the wrong unit
Teams often file both laws under a single heading, "healthcare AI compliance." That framing hides the fact that the laws regulate different entities, different workflows and different kinds of AI output.
AB 3030 added Health and Safety Code section 1339.75, effective January 1, 2025. It addresses disclosure when specified health care providers use generative AI to generate communications to a patient pertaining to patient clinical information. Enforcement runs through licensing provisions by facility type: a violation by a licensed health facility is subject to the enforcement mechanisms in Article 3 of Chapter 2, and a violation by a licensed clinic is subject to Article 3 of Chapter 1.
SB 1120 sets requirements for health plans and insurers that use AI for utilization review and utilization management. Under the Governor's office description, use of AI, an algorithm or other software must be based on a patient's medical or other clinical history and individual clinical circumstances as presented by the requesting provider, and must not supplant health care provider decision making. The California Attorney General's advisories describe physician supervision of AI tools and limits on using AI or automated decision systems to deny coverage.
As of November 2026, no amendment, delay or repeal of either law has been identified. A single organization may fall under both, which is why each system should be scoped to one statute first: provider-side patient messaging or plan-side utilization review.
AB 3030 and SB 1120 side by side
The table separates the two laws by who is regulated, what triggers the obligation and where a technical control can act. The control column is analytical design reasoning, not a statutory requirement for specific technical measures.
- Regulated entities: AB 3030: specified health care providers, enforced through health facility and clinic licensing. SB 1120: health plans and insurers.
- Workflow: AB 3030: generative AI patient communications about clinical information. SB 1120: utilization review and utilization management.
- Trigger: AB 3030: generative AI produces a patient clinical communication. SB 1120: AI, an algorithm or other software informs a utilization review decision.
- Core obligation: AB 3030: disclose the use of generative AI. SB 1120: base decisions on the patient's clinical history and circumstances, do not supplant provider decision making, and keep licensed physicians supervising AI tools.
- Runtime enforcement point: AB 3030: the outbound message. SB 1120: the approve, modify or deny action, often exposed to an agent as a tool call.
- Policy action: AB 3030: allow, redact or block. SB 1120: block automated adverse actions or escalate them to a licensed clinician.
AB 3030: controls on the outbound communication
AB 3030 is about the content of what reaches the patient. The relevant control sits between the model and the delivery channel, such as a patient portal message, email or text. At that point a policy can check whether an AI-generated clinical communication carries the required disclosure.
The policy outcomes are simple. If the disclosure is present and the message is otherwise permitted, allow it. If the message contains information that should not leave the boundary, redact it. If the disclosure is missing, block the message or hold it until the disclosure is added. Rules should be built from the statute text in Health and Safety Code section 1339.75 and counsel's reading of it, including how provider review of a communication is treated.
The evidence an auditor would reasonably expect is per message: that the communication was AI-generated, which disclosure check ran, the result, the action taken and the policy version in force. A log of sent messages alone does not show what the system was prevented from sending. Records of blocked and held messages matter as much as records of delivered ones.
SB 1120: controls on decision authority
SB 1120 is not mainly about wording. It concerns whether AI, rather than a licensed clinician, determines the outcome of a medical necessity review. The control therefore sits at the action. An AI agent may summarize a case or recommend an outcome, but an action that denies, delays or modifies care should not execute automatically.
When the adverse-determination function is exposed to an agent as a tool call, that call is a concrete place for a pre-execution policy decision. The policy can block the call outright or escalate it for human approval, routing it to a licensed clinician. The clinician's decision then determines whether the action proceeds. Approvals and recommendations that do not adversely affect the patient can follow a different path under the same policy.
Escalation has an operational cost. Human-approval queues must fit within applicable decision timeframes. For example, Insurance Code section 10123.135 sets timing rules for approve, modify or deny decisions based on medical necessity, including an outer limit of five business days from receipt of the necessary information where a shorter timeframe does not apply. Queue design, staffing and routing need to be tested against those timeframes before enforcement is turned on.
Expected evidence per action: what the AI attempted, the policy decision applied (block or escalate), the identity and role of the human reviewer, the reviewer's decision and the time it was made. This shows that the clinician decided and that the AI did not.
What runtime controls can and cannot prove
Runtime enforcement can show what happened at the level of individual interactions and tool calls. It cannot prove obligations that depend on organizational process outside the AI execution path. Who qualifies as a licensed reviewer, how a plan's review criteria are set and how policies are approved internally are examples. Those require credentialing, governance and documentation from the organization. Runtime evidence complements them and does not replace them.
The evidence also sits in the context of HIPAA 164.312(b), which addresses audit controls: mechanisms that record and examine activity in systems that handle protected health information. Per-interaction records of AI activity give that expectation a concrete form for AI systems. HIPAA does not prescribe the AI-specific mapping, and no tool by itself delivers compliance with either California statute.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with actions that include allow, redact, block and escalate for human approval. Enforcement is at runtime, not reviewed after the fact, and it generates audit-ready evidence for individual interactions and tool calls, mapped to HIPAA among other frameworks. Trussed runs in production with healthcare and pharma organizations and supports self-managed deployment in the customer's cloud, with configurable retention and data residency. It does not provide end-to-end tracing of agentic workflows or transaction-level governance.
Two statutes, two enforcement points
AB 3030: patient communications
Check the outbound AI-generated message. Allow, redact or block. Evidence shows the disclosure check result.
SB 1120: utilization review
Check the adverse-determination action. Block or escalate to a licensed clinician. Evidence shows the policy decision and the reviewer's action.
Shared foundation
Per-interaction and per-tool-call records of what the AI attempted, what policy applied and what happened.
Map each statute to its own enforcement point
Talk through how output checks for patient communications and human-approval escalation for utilization review would work in your execution path, and what evidence each would produce.
See How Evidence Is Produced