Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice
California ADMT Risk Assessment: What Section 7150 Asks You to Cover Before First Use
The short answer: Before using automated decision-making technology for a significant decision, California requires a risk assessment covering the purpose, the data involved, the risks, and the safeguards (regulation section 7150). It is a before-use duty: the assessment belongs to the build, alongside the pre-use notice, the opt-out routes, and the record design — not to the aftermath of the first complaint. For businesses already using ADMT, the compliance phase begins January 1, 2027; a business starting after that date must comply, assessment included, from its first use.
Check your duties and dates — free Build your first notice — free
What is the ADMT risk assessment? It is the documented evaluation a business completes before ADMT use, structured by the four elements the regulation names: purpose, data, risks, safeguards. It is not the same as Connecticut's data-protection assessment or a vendor's model card — it is your assessment of your use, and it sits in your record.
Element one: the purpose, stated narrowly enough to test
A purpose like 'improve hiring' cannot be assessed; a purpose like 'rank front-desk applicants by predicted schedule reliability for interview selection' can. The purpose statement drives everything downstream: it decides which data is relevant, which risks are real, and which safeguards fit. Write the purpose as the decision, the population, and the tool's exact contribution — then check the tool actually does only that.
Purpose drift is the quiet failure mode. A screening tool bought for interview selection starts informing compensation bands; the assessment still describes interview selection. When the purpose changes, the assessment is stale, and the honest move is a fresh one. The kit's record keeps assessments with the decisions they cover, so drift shows up as a date gap anyone can spot.
Element two: the data involved, category by category
List the data the tool actually consumes: the categories, where each comes from, and which are the person's own provided data versus derived or purchased data. Connecticut's written-notice duty asks for the same inventory from the other direction (categories, how assessed, sources), which is a useful cross-check: if you could not complete Connecticut's notice fields for a tool, your California data element is not done either.
Derived data deserves its own line. A score, embedding, or inferred attribute is data the tool uses, even if no human typed it. The assessment should name the consequential derived fields — the ones that can change an outcome — because those are the ones an access request will later ask about under section 7222.
Worked example: Sequoia Grove Senior Living
Example — fictional business, real duty structure. Sequoia Grove is a fictional California senior-living operator adopting a scheduling-and-assignment tool that substantially replaces human decision-making for work assignment. Its first risk assessment contains:
- Purpose: one paragraph naming the decision (weekly care-staff assignment), the population (214 care staff), and the tool's contribution (draft assignments a supervisor approves or changes).
- Data involved: 9 input categories listed with sources — availability, certifications, and hours from its own systems; no purchased data; 2 derived fields named (fatigue-risk flag, continuity score).
- Risks: 5 named risks, including a proxy risk (availability patterns tracking caregiving status) and a stale-data risk (certification expiries syncing weekly, not daily).
- Safeguards: each risk mapped to a control with an owner and cadence — including a monthly sample of 30 assignments re-checked by a person and a same-day certification override.
- Filed: the assessment is version 1, dated before first use, stored with the decision record; a review is diarised for the tool's next model update.
Element three: the risks, named for your population
Generic risk lists ('bias, privacy, security') assess nothing. Name the risks for your decision and your population: the proxy variable that could stand in for a protected trait in your data; the stale field that most often makes your outcomes wrong; the population your training or vendor data represents worst. The checker data's California exception language points at the same two risk families the regulation cares most about — accuracy and discrimination — and your assessment should address both in your context, not in the abstract.
Also assess the process risks: what happens when the tool is down, when its output is missing fields, when a reviewer disagrees. A decision pipeline with no degraded mode will improvise one under pressure. Write the fallback into the assessment and the safeguards follow naturally.
Element four: safeguards that match the risks one-to-one
A safeguard earns its place by answering a named risk. Map them explicitly: this check catches this proxy; this review step corrects this stale field; this sampling audit measures this error rate. Unmapped safeguards — 'we take fairness seriously' — are sentiments. The hiring-exception in the opt-out rules uses the same pairing: evaluated for accuracy and discrimination and safeguards kept in place. An assessment written this way doubles as that evaluation file.
Safeguards also have owners and cadences. 'Monthly sample of 30 decisions re-checked by a person' is a safeguard; 'regular review' is not. Record the cadence in the assessment so the record can later show whether it happened — the same discipline as the notice log. The wider California duty set is in California CCPA ADMT Requirements.
Filing it: one assessment, kept with the record
The assessment is a record document: dated, versioned, and kept with the notices and decisions it covers — the kit keeps one record for all states, held for three years. When the tool, the purpose, or the data changes materially, a new version joins the file; the old one stays, because past decisions were made under it.
Start from your actual position: run the free duties checker to confirm the assessment duty attaches to your use, inventory the tool with a proper system inventory, and write the first version before the next decision run — for existing use, the January 1, 2027 compliance date is the deadline that matters.
- Write the purpose as decision + population + the tool's exact contribution — narrow enough to test
- Inventory the data category by category, with sources, and name the consequential derived fields
- Name the risks for your population — accuracy and discrimination first — including process and outage risks
- Map every safeguard to a named risk, with an owner and a cadence a record can later verify
- Date, version, and file the assessment with the decision record before first use — and re-issue it when the purpose, tool, or data changes
Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.
Related guides
- California ADMT Access Requests: Explaining the Logic, the Outcome, and the Purpose
- California CCPA ADMT Requirements: What Your Business Must Do by January 1, 2027
- How to Inventory Your Automated Decision Tools: Finding Every System That Decides
- Colorado AI Act Automated Decision Notices: What to Send
See all ADMT guides for the full set.
Frequently asked questions
What must a California ADMT risk assessment cover?
The purpose, the data involved, the risks, and the safeguards, per regulation section 7150 — completed before using ADMT for a significant decision.
When must the risk assessment be completed?
Before use. For businesses already using ADMT for a significant decision, the compliance phase begins January 1, 2027; a business starting later must comply from its first use, assessment included.
Is a vendor's model card a risk assessment?
No. The duty is the business's own assessment of its own use — its purpose, its data, its population's risks, and its safeguards. Vendor material can inform it but does not replace it.
How is this different from Connecticut's assessment?
Connecticut requires a data-protection assessment before heightened-risk profiling (section 42-520). The structure is similar; the trigger and the governing act differ. The kit keeps one record for all states.
Where does the completed assessment live?
With the decision record — dated and versioned, held for three years in the kit's single all-state record, alongside the notices and responses it covers.