Testing utilization review algorithms for fair and equitable application under SB 1120
Start by inventorying every AI, algorithm or software tool used in utilization review, including tools run by contracted entities. Then compare approval, denial, modification and delay outcomes across documented patient and request segments, using stated statistical methods. Next, test that every adverse outcome reaches a licensed clinician before issuance, and retain plans, results, revisions and decision-level records. Testing shows how the tool behaves. Runtime enforcement and audit-ready evidence show what it was allowed to do and what it did on live requests.
Step 3: Run segmented comparisons of outcomes
- 1
Define the population
Fix the date range, request types and tools covered, and document any exclusions.
- 2
Define segments
List demographic and clinical segments with the rationale for each and the minimum volume for a valid comparison.
- 3
Compute outcome rates
Calculate approval, denial, modification and delay rates per segment and request type.
- 4
Apply the stated tests
Run the pre-selected significance tests and threshold sensitivity checks, and compute override rates.
- 5
Investigate disparities
Document explanations, such as case mix, and decide whether the tool, its configuration or its use needs revision.
- 6
Record revisions
Keep the results, decisions and any changes made, tied to the review that produced them.
Step 5: Decide what evidence to retain
- Tool inventory, including vendor-operated tools and their outputs
- Test plans with segment definitions, measures and statistical approach
- Results of each periodic review and the revisions made afterward
- Escalation records linking each adverse outcome to its clinician reviewer
- Decision-level records of the policy applied and the action taken on live requests
- Vendor contract terms covering access to testing evidence and inspection
What SB 1120 asks you to be able to show
As of November 2026, SB 1120 (2023-2024) amended the Knox-Keene Act and the Insurance Code. A health care service plan or disability insurer that uses an AI, algorithm or other software tool for utilization review or management based in whole or in part on medical necessity must meet a set of conditions. The same applies when it contracts with or works through an entity that uses such a tool. The California Attorney General's healthcare advisory summarizes the limit this way: these tools cannot be used to deny, delay or modify health care services based, in whole or in part, on medical necessity. A determination of medical necessity must be made by a licensed physician or a licensed health care professional competent to evaluate the clinical issues involved.
Two further conditions shape the testing program. The tool's performance, use and outcomes must be periodically reviewed and revised to maximize accuracy and reliability. The tool must also be open to inspection for audit or compliance reviews. The Attorney General has also said it is investigating potential discrimination by AI algorithms used by California healthcare entities, so disparity testing records may matter beyond SB 1120.
This points to two separate questions. First, is the algorithm applied fairly and equitably across populations and request types when tested? Second, can you show that the tested policy held in production? Pre-deployment fairness metrics answer only the first. Keep the two bodies of evidence separate in your documentation.
Step 1: Scope what is tested
Build the test inventory from the statute's reach, not from your own architecture diagram. Every AI, algorithm or software tool used for utilization review or management based on medical necessity is in scope. That includes tools operated by contracted entities, because the conditions apply to a plan that works through another entity.
For each tool, record what it outputs and where those outputs go. Distinguish approvals from denial, modification and delay outputs, because the statutory wording treats those adverse actions separately. Also record the request types it handles, such as prior, retrospective and concurrent review, and which human roles receive its outputs. A tool that only prioritizes a work queue and a tool that recommends a denial carry different risk and need different tests.
Assign an owner for the quality assurance process that assesses and evaluates compliance with the utilization review section. Vendor contracts should secure access to testing evidence and inspectability, since the plan remains responsible for the conditions.
Step 2: Choose fairness and disparity measures, and know their limits
The most direct measure is outcome-rate comparison. NIST SP 1270 describes traditional statistical bias testing as looking at differences in predictions across protected classes such as race or sex. For utilization review, compare approval, denial, modification and delay rates across demographic segments and clinical segments, and by request type. Clinical segmentation matters because a disparity that appears across a demographic group may be explained by case mix, or may not be. Document why each segment was chosen.
NIST SP 1270 also notes that courts and regulators have used or considered significance testing to assess whether a disparity was likely caused by the challenged practice rather than by chance or a nondiscriminatory factor. Specify your significance approach before running the comparison, and note how small segments are handled, since low volumes produce unstable rates.
Two further methods are commonly used and are worth considering as candidates. Threshold sensitivity analysis varies the score or rule cut-off at which a case is flagged for adverse action and checks whether outcome gaps widen or narrow. Override-rate analysis compares how often clinicians reverse or modify the tool's recommendation, by segment. A higher override rate in one segment can indicate the tool performs differently there. Both need validating against authoritative guidance before you rely on them.
The limits are real. SP 1270 is general AI bias guidance, not SB 1120-specific, and it does not prescribe metrics, segments or thresholds for utilization review. Outcome-rate parity does not prove outcomes are correct, and observed outcomes can reflect upstream clinician behavior. Record your chosen measures and the reasons for them as design decisions.
Step 4: Test that adverse outcomes reach a licensed clinician
Fair rates do not satisfy the requirement that medical necessity determinations be made by a licensed physician or competent licensed health care professional. Test this separately. For every denial, delay or modification output, trace the path from tool output to issued decision and confirm a qualified clinician reviewed it before issuance. Record any path where it does not, including batch processes, system integrations, fallback modes and vendor-run workflows.
This is also where testing reaches its limit. A test run before or during a review period samples behavior. It does not show that the routing policy held on every live request afterward. If routing depends on staff following a procedure, evidence of compliance depends on their records.
Runtime enforcement addresses that gap. Policy enforced in the execution path can allow, redact, block or escalate for human approval at the level of individual interactions and tool calls. In a utilization review workflow, that means an adverse outcome can be held for clinician approval rather than released, with a record of the policy applied and the action taken. Trussed provides this kind of enforcement in the execution path. It does not replace fairness testing, and it does not test an algorithm before deployment.
Two questions, two bodies of evidence
Scope
Every utilization review tool in use, including vendor-operated tools.
Segmented outcomes
Approval, denial, modification and delay rates compared across defined segments.
Clinician escalation
Adverse outcomes verified as reaching a licensed clinician.
Production evidence
Decision-level records showing what the tool was permitted to do and did.
Connect tested policy to production evidence
Talk to our team about enforcing escalation policy in the execution path and producing audit-ready evidence for each interaction and tool call.
Talk to an Expert