ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
What goes into California's annual ADMT risk assessment (due 1 April 2028)?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. The risk assessment is the annual, structured examination of your ADMT processing: what the technology decides, what information it consumes, what benefits it brings, what risks it creates for consumers, and what safeguards answer those risks. ReviewDesk's paid plans include the annual risk-assessment pack for California, due 1 April 2028 — and because the pack is generated from the workflow record, most of its content already exists if your notices, traces and inbox have been run properly all year.
What is the California ADMT risk assessment?
An ADMT risk assessment is a documented evaluation of a business's automated decision-making processing, required under California's framework for covered uses. It weighs the purpose and benefits of the processing against its risks to consumers — inaccuracy, discrimination, opacity, security — and records the safeguards in place. In ReviewDesk it is an annual pack assembled per workflow, first due 1 April 2028.
Why does California require an assessment at all?
Notices tell consumers what will happen; assessments force businesses to ask whether it should. The assessment is the regime's self-examination duty — the moment where the organisation must articulate, in writing, why the processing is justified and how its harms are controlled. A business that cannot write the benefits section without slogans, or the risks section without euphemisms, has learned something important about its own deployment. That is the point of the exercise, and why treating it as a form-filling task wastes its value.
The annual cadence matters too. ADMT systems drift: vendors ship model updates, thresholds are tuned, decision types creep. A one-off assessment fossilises the system as it was at launch. The yearly cycle forces the description, the factor lists and the safeguards to be re-verified against the operation as it actually runs — the same anti-drift discipline the notice versioning applies, at programme level.
What are the pack's working parts?
Start with the processing description per workflow: the significant decision, the technology's role, the categories of personal information, the outputs produced. Then benefits, stated concretely — processing time, consistency, capacity — with numbers from your own operation where you have them. Then risks, stated just as concretely: the plausible failure modes for this decision type. Wrong-data decisions. Proxy factors. Review bottlenecks. Vendor opacity. Each risk pairs with its safeguard: the correction route, the factor review, the SLA clocks, the scope worksheet. Finally, the record of who assessed, when, and what changed since last year.
ReviewDesk's pack draws each part from live records: workflow configuration for the description, trace volumes for scale, inbox metrics for responsiveness, overturn and correction rates for the accuracy picture. The assessment stops being a blank document and becomes an assembly task with judgement applied on top — which is why the audit-trail discipline during the year determines whether April is a week of work or a month of archaeology.
How honest must the risks section be?
Completely — it is a self-assessment, and a risks section listing no risks is a confession that nobody looked. Real risks are usually already known inside the operation: the factor whose data quality the team grumbles about, the decision type where reviews cluster, the vendor who cannot explain a score component. Writing them down with their safeguards is not an admission of failure; it is the demonstration that the programme is supervised. The packs' evidence discipline applies here as in the traces: a safeguard that is not actually operating renders as a gap, not as a paragraph.
Prioritise risks by consumer impact, not by ease of fixing. A rare failure that wrongly denies housing outweighs a frequent failure that delays a marketing offer — but only significant decisions are in this regime anyway, so weight by the decision's consequence and the failure's likelihood together. The resulting order usually surprises teams: data-correction plumbing outranks model sophistication almost every time, because most real-world ADMT harm arrives through bad inputs, not exotic algorithms.
Who should write and sign the assessment?
The people who run the decisions, reviewed by someone who does not. Operations drafts from the workflow records; a second reader — leadership, counsel, an external reviewer — challenges the benefits claims and the risk omissions. The assessor's identity and date belong in the pack, because an unsigned assessment is an opinion. In smaller organisations the same person may draft and own, but the challenge step should still happen as a distinct, recorded pass — the same separation discipline as meaningful review, applied to the programme rather than the case.
Vendors belong in the evidence base, not the authorial chair. Your assessment must be able to say what the vendor's tool does with your consumers' information; if the vendor cannot support that description, the gap itself is an assessment finding with a safeguard attached — factor documentation requested, contract terms at renewal, or a workflow redesigned to need less of the vendor's opacity. The questions to put to vendors are catalogued in the vendor guide.
How does the 1 April 2028 date organise the year?
Work backwards. The pack is due 1 April 2028; the assessment should be substantially drafted in January from the prior year's records, challenged in February, and finalised in March. That calendar has a hard implication for 2027: the records the assessment consumes are being created from the moment your workflows go live. A business that starts its workflows in December 2027 will write its first assessment from six weeks of thin data and a great deal of reconstruction — possible, but the weakest version of the document.
The calendar view across all three regimes — Colorado's 30-day clocks running continuously, Illinois already live, the EU horizon in December 2027 — is set out in the deadlines calendar. The assessment is the only annual, set-piece obligation among them, which makes it the natural anchor for a yearly programme review: notices re-read, factors re-challenged, overturn rates examined, vendor answers refreshed.
What does a strong first assessment look like?
Specific, evidenced and a little uncomfortable. It names the decisions and the volumes. It quantifies responsiveness — median days to explanation and to review outcome — from the inbox record. It reports the overturn and correction rates without cosmetic rounding, interprets them, and attaches actions: the factor being re-validated, the notice paragraph being rewritten, the vendor question escalated. It closes with the changes made since launch and the changes planned, each with an owner. A reader finishing it should believe the programme is run by people who look at it.
ReviewDesk cannot supply the judgement, but it removes every excuse about the evidence: the workflow record, the traces, the inbox metrics and the vault export are the pack's skeleton. Generate the structure from the desk, then spend your human hours where they belong — on the risks section's honesty and the actions' realism. You can see the plan tiers that include the pack on pricing: it sits in Pro and Team, alongside the evidence engine.
Your five actions for the California ADMT risk assessment
- Describe each workflow's processing from the live workflow record — decision, role, information categories, outputs.
- State benefits with your own operating numbers, and risks with the failure modes your team actually worries about.
- Pair every risk with a safeguard that is operating today — or record it as a gap with an action and an owner.
- Have the draft challenged by a reader who did not write it, and record assessor, date and changes since last year.
- Draft in January from the prior year's records; never let the assessment year be the year the records start.
Worked example: a first assessment assembled from a year of records
A fictional rental group — Cascade Homes — runs tenant screening in California through one workflow. Its first assessment, drafted in January 2028, takes the screening volumes from the trace record: 3,940 applications, 612 adverse outcomes, explanations issued in a median of two days. The inbox supplies 88 review requests, 19 outcomes changed — 14 of them through data correction. The risks section names the pattern plainly: namesake credit-file confusion is the single largest source of overturned decisions.
The safeguard recorded against that risk is operational, not aspirational: the correction route was moved into the explanation letter's first paragraph in September, and correction turnaround is tracked weekly. The action for the coming year is equally concrete — a pre-decision identity-verification step for files with merged-record indicators, owner named, due in June. The assessment runs to eleven pages, nine of them assembled from records. It reads like supervision because it is supervision, written down.
Frequently asked questions
When is the California risk assessment due?
ReviewDesk's annual risk-assessment pack for California is due 1 April 2028, as listed on the paid plans. It is an annual obligation — draft from the prior year's records early in the year rather than reconstructing at the deadline.
What does the assessment examine?
The processing per workflow: the significant decision, the technology's role, the information categories, the benefits, the risks to consumers and the safeguards answering each risk — plus who assessed it, when, and what changed.
Which ReviewDesk plans include the pack?
Pro and Team include the annual risk-assessment pack (CA, due 1 Apr 2028), alongside SLA clocks, California opt-out handling, the evidence vault and the three-regime evidence engine. See pricing.
What if our records only start partway through the year?
Then the first assessment says so, uses what exists, and records the gap with its fix. An honest partial record beats a reconstructed full one — the same 'not recorded' discipline the evidence engine applies to traces.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. The assessment is described as ReviewDesk's pack; consult counsel for your obligations.
Keep reading: How does one decision trace become Colorado, California and EU evidence packs? · What is a hash-chained decision record — and why does immutability matter as evidence? · How do you run the scope test: does your system materially influence or substantially replace human decision-making? · How should you prepare EU Annex III evidence for automated decisions — before the duties apply? · How do you let someone correct wrong data after an adverse automated decision? · All ReviewDesk guides · ReviewDesk home · Try the free desk · Free letter generator · Free notice builder · Pricing
ReviewDesk is built by Ingenious AI. Templates and workflow tools — not legal advice.