Blocking AI from Denying, Delaying, or Modifying Care at Runtime Under SB 1120
To enforce SB 1120 at runtime, classify every model output, agent action, and system-of-record write that could deny, delay, or modify care on medical necessity grounds. Block or hold those outcomes in the execution path until a licensed clinician approves them. Record the policy version, the enforcement decision, and the approval so you can show what the AI was allowed to do and what it did. Counsel should confirm legal sufficiency.
- {"title":"Model responses","description":"Classify output for deny, delay, or modify content before it reaches a user or downstream system."}
- {"title":"Agent actions","description":"Evaluate each planned tool call and state change against the adverse outcome rules."}
- {"title":"System-of-record writes","description":"Refuse AI-only writes that finalize an adverse determination."}
- 1. The model or agent produces a candidate outcome, and the policy classifies it before it takes effect.
- 2. An adverse classification holds the action and blocks any write to the system of record.
- 3. The case enters the clinician queue with its urgency, deadline, and the AI-assisted material attached.
- 4. A licensed clinician approves, modifies, or rejects, and the approval is bound to that specific determination.
- 5. Only the approved determination is released, and the release call is itself checked against policy.
- 6. Approaching deadlines trigger escalation to another reviewer, not a silent timeout.
Evidence to retain for an audit
- Policy version and rule identifier in force at the time of each decision
- Input classification: what the output or tool call was judged to be
- Enforcement decision: allow, redact, block, or escalate, with timestamp
- Approver identity, credential reference, decision, and timestamp
- A link from each approval record to the exact determination it released
- Retention period and export format that match your regulatory and contractual obligations
What SB 1120 restricts, and what can be enforced at runtime
SB 1120 (2023-2024), titled Health care coverage: utilization review and known as the Physicians Make Decisions Act, was passed by the California Legislature in 2024. It restricts health insurers and disability insurers from using AI, algorithms, and similar tools as the sole means to deny, delay, or modify care based on medical necessity. The California Attorney General describes the requirement as physician supervision of AI tools, and describes the law as amending the Knox-Keene Act and the Insurance Code. As of November 2026, no later amendment, delay, or repeal is reflected in the material reviewed for this guide.
The restriction is narrow in a useful way. It targets three adverse outcomes arising from medical necessity decisions, not AI use in general. That makes it a candidate for runtime enforcement, because the outcomes can be described as conditions on what a model, agent, or tool call is permitted to produce or do. Drafting, summarizing, retrieving records, and flagging cases for review are different activities from issuing a final adverse determination, and a policy can treat them differently.
This guide covers individual interactions and tool calls. Runtime enforcement evidence shows what a system was permitted to do at each step. It does not establish that an organization is covered by the statute or that it complies with it. Those are questions for counsel.
Translate the restriction into enforceable policy rules
Start with an inventory of utilization review workflows that use AI, and mark each step that produces or influences an approve, modify, deny, or delay outcome. The statutory language uses approve, modify, and deny as its decision verbs, and the summary of the Act adds delay. That gives you a working vocabulary for classifying outputs. Mapping the vocabulary to specific fields in your systems is an implementation choice, and you should have legal review the resulting definitions of "sole means" and clinician supervision before they become policy.
Write each rule as a testable condition with a defined action. A rule that cannot be evaluated against a single interaction or tool call cannot be enforced in the execution path.
- Adverse outcome classification: Any output or payload that denies, delays, or modifies a request on medical necessity grounds is classed as adverse. Classification is evaluated before the outcome takes effect.
- No AI-only final determination: A write that finalizes an adverse determination is blocked unless it carries a clinician approval tied to that specific case.
- Escalate, do not guess: Outputs that cannot be confidently classified as adverse or non-adverse are escalated to a human reviewer rather than allowed.
- Allow non-adverse work: Summaries, record retrieval, and drafts that are not issued as determinations are allowed, and each is labeled as AI-assisted.
- Redact at the boundary: Draft determination fields or rationale text are redacted from payloads headed to systems or recipients that could treat them as final.
- Policy versioning: Every rule carries a version, an owner, and an effective timestamp, so each enforcement decision can be traced to the rule in force.
Evaluating in-path enforcement, including Trussed
Whatever tooling you choose, ask four questions. Can it classify an output or tool call as an adverse determination before the outcome takes effect? Can it block or hold that outcome and route it to a clinician? Can you show the enforcement at the tool-call level? Are the policy versions tied to each decision and each approval record? Also check how the approval workflow handles urgent timelines and reviewer availability, and what the retention and export options are.
Trussed is one example of this approach. It enforces policy in the execution path across models, agents, tools, and MCP servers, with actions to allow, redact, block, rate limit, or escalate for human approval, scoped to individual interactions and tool calls. Policy is enforced at runtime, not reviewed after the fact. Enforcement generates continuous audit-ready evidence, mapped to frameworks including HIPAA, NAIC, NYDFS, the EU AI Act, and NIST. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. Trussed serves regulated enterprises, with customers across insurance, pharma, healthcare, and education, and has run in continuous production in a regulated enterprise for more than 18 months.
The design above does not depend on any one product. The test is whether an AI-originated adverse determination can take effect without a clinician, and whether you can prove the answer.
Runtime path for an AI-originated adverse determination
Classify
Identify outputs and tool calls that read as deny, delay, or modify.
Enforce
Allow, redact, block, or escalate before the outcome takes effect.
Approve
A licensed clinician releases or rejects the held determination.
Evidence
Policy version, decision, and approval are recorded and linked.
Map your utilization review decision points to runtime policy
Review where AI-originated outcomes could become adverse determinations in your workflows, and see how enforcement and evidence work at the tool-call level.
Explore Agent and Tool-Call EnforcementHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.