California SB 1120 | Implementation Guide

    How to prove a licensed physician made the final medical necessity decision under California SB 1120

    Proof requires more than a reviewer sign-off in a case management system. For each denial, delay, or modification, you need a record of who the licensed reviewer was, what clinical information and AI output they saw, when they approved, and evidence that the AI step could not finalize the outcome without that approval. The most direct way to produce this is to enforce the approval requirement in the execution path and capture evidence from each enforcement and approval event.

    Enforcing physician approval in the execution path

    1. 1

      1. Classify the output

      Identify AI outputs that recommend denying, delaying, or modifying care for medical necessity reasons.

    2. 2

      2. Block finalization

      Prevent any AI-originated finalizing action, such as a status write or a determination letter, from executing.

    3. 3

      3. Escalate for human approval

      Route the case to a reviewer who holds the required licensure or competence for the clinical issue.

    4. 4

      4. Present the record

      Show the reviewer the clinical information and, separately, the AI output, and keep a copy of what was shown.

    5. 5

      5. Record the approval

      Capture reviewer identity, licensure check, decision, and timestamp as a single event tied to the case.

    6. 6

      6. Release and retain

      Permit the finalizing action only after approval, and store the enforcement and approval events with the case.

    Evidence elements to capture for each determination

      What SB 1120 requires

      SB 1120 (2023-2024), the Physicians Make Decisions Act, is codified in Health and Safety Code 1367.01 for health care service plans and Insurance Code 10123.135 for insurers. As of November 2026, it is enacted and in effect from January 1, 2025, and no later amendment has been reported.

      The core restriction is about who decides. No individual other than a licensed physician or a licensed health care professional competent to evaluate the specific clinical issues involved may deny or modify a request for authorization for reasons of medical necessity. The category is broader than "licensed physician" alone, so evidence design should capture licensure or competence for the specific clinical issue rather than assume a physician in every case.

      When a plan or insurer uses an AI, algorithm, or other software tool for utilization review or management based in whole or in part on medical necessity, or contracts or works through an entity that does, it must comply with the section and ensure listed conditions. Two matter most here. The tool must not supplant health care provider decisionmaking, and it must base medical necessity decisions on the enrollee's medical or other clinical history, other relevant clinical information in the record, and the individual's clinical circumstances. Patient data also may not be used beyond its intended and stated purpose, consistent with the Confidentiality of Medical Information Act and HIPAA, as applicable.

      The California Department of Justice legal advisory frames the same point: these tools cannot be used to deny, delay, or modify care based, in whole or in part, on medical necessity unless the conditions are met. This is not legal advice, and counsel should confirm how these provisions apply to your operations.

      Statutory requirement versus audit evidence

      The statute names the decision-maker and the non-supplanting condition. It does not prescribe an evidence format. What follows is a reasonable expectation of what an auditor would look for, not a statutory list.

      An auditor would likely want to link each denial, delay, or modification to an identified licensed reviewer, see what that reviewer was shown, including any AI output, and confirm that the tool did not issue the determination itself. A policy statement that physicians make final decisions addresses none of these. A case management sign-off shows that someone clicked approve. It does not show that the AI step was unable to finalize the outcome, or that the reviewer had the clinical record and room to disagree.

      Because the statute's scope is "in whole or in part," evidence should cover AI-assisted steps that influence a determination, not only steps where the AI makes the final call. A recommendation that a reviewer routinely accepts without material review is the case your evidence needs to be able to distinguish from independent judgment.

      Delegates, communications, and the evidence store

      The statute reaches plans that contract or work through entities using these tools. If a delegate or contracted AI vendor performs utilization review steps, the same evidence needs to exist for their determinations and be retrievable by you. Assign one owner for evidence completeness across the plan and its delegates, and test retrieval before an audit asks for it.

      Check evidence against the communication requirements as well. Decisions on prior or concurrent authorization requests must be communicated to the requesting provider within 24 hours of the decision. Decisions that deny, delay, or modify a service must be communicated to the enrollee in writing within two business days. Written communications to providers must name the responsible health care professional and give a telephone number. The Insurance Code adds a clear and concise explanation of reasons, a description of the criteria or guidelines used, and the clinical reasons for the decision. The Insurance Code and Health and Safety Code language differ, so verify against the section that applies to each line of business.

      Retention periods, data residency, and access controls for the evidence store should be set with counsel and applied consistently, with access limited given the patient data involved. Trussed enforces policy in the execution path at the level of individual interactions and tool calls, with actions that include allow, redact, block, and escalate for human approval. Each enforcement event produces audit-ready evidence. It can be deployed self-managed in the customer cloud with configurable retention and data residency. It does not provide workflow-level governance or end-to-end tracing of an entire utilization review process, so the case-level record and delegate coverage remain your design work.

      Assessing your current workflow

      Walk through one recent denial, delay, and modification and try to answer five questions from records alone. Who was the licensed reviewer, and how do you know their licensure or competence applied to this clinical issue? What did they see, including any AI output? Could the AI step have finalized or communicated the determination without a recorded approval? Was the approval timestamp consistent with the 24-hour and two-business-day communication windows? Can the same evidence be retrieved for decisions made by a delegate?

      Each question you cannot answer from records is a gap. The most common gap in this design is the third: a sign-off exists, but nothing demonstrates the AI step was prevented from acting alone. Closing it requires an enforcement point before finalization, not a better policy document. Prioritize that control first, then complete the remaining evidence fields.

      From AI recommendation to documented physician decision

      AI recommendation

      A recommendation to deny, delay, or modify care is generated.

      Block and escalate

      Finalization is held until a licensed reviewer is asked to approve.

      Reviewer decision

      Identity, licensure, timestamp, and information presented are recorded.

      Release and retain

      The determination proceeds and the evidence is stored for audit.

      Check your workflow against the evidence checklist

      Talk through where finalizing actions occur in your utilization review workflow and how enforcement and approval evidence would be captured.

      Explore Agent and Tool-Call Enforcement
      Next step

      How does your AI governance program compare?

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