ReviewDesk
← Back

ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026

What is automated decision-making technology (ADMT)?

Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing

The short answer. Automated decision-making technology is technology that processes people's information and uses computation to make — or materially help make — a decision about them. A spell-checker is not ADMT in any sense that matters; a tool that ranks job applicants, prices insurance applicants or screens tenants is. ReviewDesk exists for that second category: it holds the pre-use notice, the adverse-outcome letter and the human-review path that the Colorado, California and Illinois rules attach to such tools.

What is automated decision-making technology?

Automated decision-making technology (ADMT) is any technology that processes personal information and uses computation to execute a decision, or to substantially assist or replace human discretion in making one. In ReviewDesk's world the decisions are consequential ones — employment, credit, housing, insurance, healthcare, education — and the legal consequence of the label is a set of duties: notice before use, explanation after adverse outcomes, human review on request, and records.

Where does the line between software and ADMT actually sit?

Almost every modern business process uses software; very few processes let software decide about a person. The line is discretion. A spreadsheet that totals a recruiter's interview scores is arithmetic in service of a human decision. A model that scores 800 applications overnight and discards 600 before a human opens the file has exercised discretion, whatever the vendor calls it. Ask two questions. Would the outcome plausibly differ if the tool's output changed? And does any human genuinely re-derive the decision, or merely ratify the output? If the output drives the outcome and the human ratifies, you are looking at ADMT.

Borderline cases deserve a recorded judgement rather than a shrug. ReviewDesk's scope-test worksheet exists for this: per decision type, you record whether the system materially influences or substantially replaces human decision-making, with a timestamp. That worksheet becomes part of the evidence pack, so a later reviewer sees your reasoning at the time — not a reconstruction after a complaint. The test itself is explored in the scope-test guide.

Which decisions count as consequential or significant?

The regimes use different labels — Colorado speaks of consequential decisions, California of significant decisions — but the territory overlaps heavily: getting a job, keeping a job, borrowing money, renting a home, being insured, receiving healthcare, accessing education. The common thread is that the decision changes a person's material circumstances and is hard to reverse by simply trying another provider tomorrow. A declined rental application costs weeks and fees; a declined job costs income. Contrast a film recommendation, where the cost of a bad output is two wasted hours.

Within a single business the same technology can cross the line in one use and not another. A chatbot answering opening-hours questions is not ADMT; the same vendor's module scoring support-agent applicants is. Classify uses, not products. That is why ReviewDesk organises work into workflows — one decision, one jurisdiction, one notice — rather than one blanket disclosure for a whole software estate. See setting up your first workflow for the unit of work in practice.

What forms does ADMT take in ordinary businesses?

The recognisable forms are ranking, scoring, filtering and recommending. Applicant screening ranks candidates. Tenant screening produces a score and an accept-or-review recommendation. Insurance workflows flag applications for standard or enhanced underwriting. Credit workflows sort applicants into approval bands. Less visible forms matter too: interview-scheduling tools that quietly prioritise some candidates' slots, shift-allocation systems that distribute hours by predicted reliability, and assessment platforms that grade recorded responses. If a person experiences the output as a decision about them, treat the use as in scope until your scope test says otherwise.

Vendor language obscures form. 'Insights', 'signals' and 'copilot suggestions' can all be recommendations that a busy human accepts ninety-five times in a hundred. The honest measure is the acceptance rate and the time a human actually spends per case. A reviewer given forty seconds per file is not exercising discretion; they are witnessing it. Meaningful human review — the standard the letters promise — is defined against exactly that failure mode in the human-review checklist.

Why are governments regulating ADMT now?

Three forces converged. Scale: a biased or buggy human harms dozens; a deployed model harms thousands before anyone notices a pattern. Opacity: affected people historically could not learn that a tool was involved, let alone which factors counted, so errors and proxies went uncorrected. And asymmetry: the business holds the model, the data and the logs, while the person holds only a rejection letter. Notice, explanation, correction and human-review duties rebalance that asymmetry without banning the technology — none of the three ReviewDesk regimes prohibits ADMT; all three insist it operate in the open.

The timelines reflect that settlement. California's regulations took effect 1 January 2026 with compliance from 1 January 2027; Colorado's replacement act, SB 26-189, takes effect 1 January 2027; Illinois employment duties are already live; and the European Union's high-risk duties are labelled preparatory in ReviewDesk until they apply on 2 December 2027. Businesses get lead time, but the direction is not in doubt, which is why building the workflow now — notice, inbox, letters, records — beats a scramble in the final quarter.

What duties follow once a tool is ADMT?

Four duties recur across the regimes, in different proportions. Notice before use, so the person knows a tool is involved before it processes them. Explanation after an adverse outcome — in Colorado, within 30 days, with principal reasons — so the person understands what counted against them. A human path: correction of wrong data, California's opt-out or appeal, and meaningful human review of the decision itself. And records: the audit trail that proves the first three happened, kept for three years in the ReviewDesk model. A tool purchase that cannot support all four is an incomplete purchase.

ReviewDesk maps those duties to product parts so nothing depends on memory. The hosted notice carries the first duty, the letter generator the second, the inbox the third, and the timestamped audit trail plus evidence vault the fourth. Paid plans add the machinery with deadlines attached — SLA clocks, opt-out handling, the annual risk-assessment pack — because duties with dates need clocks, not intentions.

How should a small business decide where it stands?

Start with an inventory, not a legal opinion. Walk each customer- and applicant-facing process and ask where software scores, ranks, filters or recommends about an identifiable person. Include the features bundled inside platforms you already pay for; ADMT rarely arrives as a product labelled ADMT. For each candidate use, run the two discretion questions, record the scope determination, and only then decide what notices you owe. Most small businesses find one or two genuine ADMT uses — typically hiring screening and, where relevant, tenant or credit screening — which is precisely the size of problem the free tier was built for.

Then publish one honest notice and wire one inbox before expanding. A single workflow run properly teaches the organisation more than a policy paper covering twelve hypothetical uses. When the second workflow arrives — a new jurisdiction, a second decision type — the pattern is already proven. That is the adoption path described in the small-business guide: inventory, scope, notice, review, records, in that order.

Your five actions for automated decision-making technology

  1. List every process where software scores, ranks, filters or recommends about an identifiable person — including bundled platform features.
  2. Apply the discretion test to each: does the output drive the outcome, and does a human genuinely re-derive the decision?
  3. Record each scope determination with a timestamp — a worksheet entry, not a verbal understanding.
  4. Map each in-scope use to the four duties: notice, explanation, human path, records.
  5. Stand up one workflow end-to-end — notice, inbox, letters — before expanding to the next decision or jurisdiction.

Worked example: one company, three tools, only two of them ADMT

Consider a fictional property group — Harbor & Vine Rentals — using three systems. A maintenance portal sorts repair tickets by urgency keywords; tenants experience faster or slower repairs, but no decision about a person is made, and the group records it as out of scope. A tenant-screening service scores applications and recommends accept or decline; property managers accept the recommendation in roughly nine cases out of ten. That is ADMT by any honest test, and it gets a workflow, a pre-use notice and a review path.

The third system grades recorded video answers from applicants for on-site caretaker roles. Managers insist they 'only use it as one input', but the numbers tell the truth: of 156 applicants last quarter, nobody scoring in the bottom half was interviewed, whatever their other merits. The scope worksheet records the system as materially influencing hiring, and it joins the notice regime. The exercise took one afternoon; the alternative — discovering the classification during a complaint — would have taken considerably longer and cost considerably more.

Frequently asked questions

What is automated decision-making technology in simple terms?

Software that processes a person's information and makes, or materially helps make, a decision about them — such as screening a job application, scoring a tenant or banding a credit application. Ordinary office software that merely stores or totals human judgements is not ADMT.

Does ADMT have to be fully automatic to count?

No. A tool that substantially assists or replaces human discretion counts even when a person formally signs off the outcome. The test is whether the human genuinely re-derives the decision or mostly ratifies the output.

Which decisions are covered?

Consequential or significant decisions: employment, credit, housing, insurance, healthcare and education in the Colorado, California and Illinois frameworks ReviewDesk tracks. The classification applies to each use of a tool, not to the product as a whole.

What must a business do if it uses ADMT?

Give notice before use, explain adverse outcomes (in Colorado within 30 days), offer correction and meaningful human review — plus California opt-out or appeal routes — and keep records. ReviewDesk's free tier holds one workflow with all four parts.

Is this guide legal advice?

No. ReviewDesk is templates and workflow tooling, not legal advice. The definition is summarised from the regimes covered in our research; consult counsel for your situation.

Keep reading: How do you explain an adverse automated decision within Colorado's 30 days? · What makes a human review of an automated decision 'meaningful'? · How does the California ADMT opt-out work for consumers and businesses? · California ADMT: should you offer the opt-out, the human appeal, or both? · How long must you keep ADMT decision records — and what should the audit trail hold? · 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.