ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
How do you run the scope test: does your system materially influence or substantially replace human decision-making?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. Look at what actually happens, not what the process chart says. If the tool's output usually determines the outcome — high acceptance, thin human review, no realistic path for the human to reach a different answer — the system materially influences the decision, and may substantially replace the human in it. ReviewDesk's scope-test worksheet records that determination per decision type and regime, timestamped, inside the evidence pack — your judgement, evidenced, not the tool's certification.
What is the ADMT scope test?
The scope test is the determination of whether an automated system falls within an ADMT regime's reach for a given decision type. The worksheet language ReviewDesk uses asks whether the system materially influences or substantially replaces human decision-making. It is completed per decision type — the same product can be in scope for screening and out of scope for scheduling — and stored with the evidence pack.
Why is the test about practice, not architecture?
Because a system's legal character is set by its effect on decisions, and effects live in operations. The process chart says a recruiter reviews every recommendation; the timestamps say the recruiter averages forty seconds per file and accepts ninety-seven per cent of recommendations. Both statements are drawn from the same organisation, and only one of them describes the decision-making. The scope test exists to force the operational description onto the record before a complainant's lawyer does it for you.
That is also why the worksheet is per decision type and revisited on change. A tool introduced as a drafting aid for job descriptions drifts, release by release, into screening. Scope follows the drift only if someone re-runs the test — the worksheet's timestamp makes the last honest look visible, and staleness is itself readable by a reviewer. The worksheet method belongs to the evidence engine described in the evidence-pack guide.
What evidence feeds the determination?
Four measurements, all available from ordinary operations. Acceptance rate: how often the human outcome matches the tool's recommendation. Review depth: median human time per case, and whether the reviewer sees underlying data or only the output. Reversal realism: whether a human can reach a different outcome without exceptional effort or escalation. Consequence routing: whether any path exists by which a case bypasses the tool entirely. Together they sketch the tool's true discretion more honestly than any vendor description.
Gather them before writing the determination, and keep the underlying numbers with the worksheet. A determination that cites an eighty-nine per cent acceptance rate and a three-minute median review is a judgement with a basis; a determination that merely asserts 'human in the loop' is a slogan with a date. The numbers also give the annual risk assessment its factual spine — see the risk-assessment guide for how the same measurements reappear at programme level.
Where is the line between influence and replacement?
Influence means the human decision is genuinely made but materially shaped — the tool's ranking determines who gets serious attention, its flags determine which files are read closely. Replacement means the human's role has collapsed into ratification: the outcome is the tool's output in all but pathological cases. Both are within the regimes' concern — the worksheet asks the question in the alternative precisely because influence alone is enough to matter. Teams sometimes argue themselves out of scope by proving the tool 'only' influences; that argument misunderstands the threshold it is trying to duck under.
Edge cases resolve by consequence. If the influenced decision is consequential for the person — job, home, credit, cover — treat the influence as material unless the human process demonstrably re-derives outcomes. The burden of the demonstration sits on the worksheet, with measurements. When in genuine doubt, the conservative answer costs a notice and an inbox; the aggressive answer costs credibility if a reviewer measures what you declined to.
Who should complete the worksheet — and who should challenge it?
The owner of the decision process completes it, because they hold the operational facts. Someone outside the process challenges it, because process owners systematically overestimate their own review depth — not from dishonesty, but because they remember the exceptional cases where they dug deep, not the median case where they nodded. Pull the acceptance and timing numbers first, and let the challenger interrogate the numbers rather than the owner's intentions.
Legal counsel reviews the determination's framing; they do not manufacture the facts. The worksheet is explicit that the determination is the customer's own — ReviewDesk records and evidences it, and does not make legal determinations. That boundary keeps the worksheet honest in both directions: the platform cannot be used as a scope oracle, and the customer cannot later claim the platform certified their position.
What triggers a re-test?
Any change that could move the measurements. A vendor model update, especially one marketed as more autonomous. A threshold change that alters acceptance volumes. A staffing change that halves review time. A new decision type folded into the existing workflow. A volume surge that turns review into triage. Each is a scope event; the worksheet should gain a new dated determination rather than a quiet edit, following the append-only discipline of the decision-record chain.
Calendar discipline helps too: re-test annually alongside the risk assessment even if nothing obvious changed, because drift is defined by changes nobody announced. The re-test is cheap when the measurements are already collected — one more argument for instrumenting acceptance and review time as routine telemetry rather than assessment-season archaeology.
What happens after an in-scope determination?
The workflow obligations switch on for that decision type: the pre-use notice, the explanation machinery, the human path, the traces. In ReviewDesk terms, the decision type becomes a workflow with its jurisdiction configuration, and the worksheet travels with its evidence pack as the record of why. An out-of-scope determination is not the end of paperwork either — it is filed, with measurements, so that a future question meets a dated, evidenced judgement instead of a shrug.
Sequence the obligations by the person's experience: notice first, because everything else references it; the inbox next, because promises need a destination; letters alongside, because adverse outcomes will not wait for your programme plan. The step-by-step build is in the workflow setup guide. Scope is the gate; the workflow is what stands behind it.
Your five actions for the ADMT scope test
- Measure before you judge: acceptance rate, median human review time, what the reviewer actually sees, whether bypass paths exist.
- Complete the worksheet per decision type and per regime — never once per product.
- Have the determination challenged by someone outside the process, arguing from the numbers.
- File out-of-scope determinations with their evidence too; a dated judgement beats a shrug.
- Re-test after every model, threshold, staffing or volume change — and annually regardless.
Worked example: the 'advisory' tool that was deciding
A fictional insurer — Beacon & Co — introduces a tool that 'advises' on underwriting referrals. The worksheet exercise pulls the numbers: advisers' recommendations are followed in 96 of 100 cases, the median human touch is under two minutes, and reviewers see the recommendation and a summary, not the underlying file. The determination writes itself: the system materially influences underwriting decisions and, for straightforward cases, substantially replaces the human judgement it nominally advises. Beacon files the determination, stands up the workflow, and — the genuinely useful part — redesigns review for the segment where human judgement adds most.
Contrast the same company's scheduling tool, which proposes diary slots that staff rearrange freely, with acceptance under half and no consequence for the people scheduled beyond timing. Out of scope, filed with the measurements. One afternoon of arithmetic separated the two honestly; without it, both would have carried the same vague 'AI-assisted' label, and the underwriting exposure would have stayed invisible until it mattered.
Frequently asked questions
What is the scope test?
The determination, per decision type, of whether an automated system materially influences or substantially replaces human decision-making — the threshold question for ADMT duties. ReviewDesk's worksheet records your determination, timestamped, inside the evidence pack.
What numbers best evidence the determination?
The tool's acceptance rate, median human review time per case, whether reviewers see the underlying data or only the output, and whether any path bypasses the tool. Measured practice, not process charts.
Does 'materially influence' alone bring a system into scope?
The worksheet asks the question in the alternative because influence is enough to matter: if the output materially shapes consequential decisions, the duties should be treated as engaged unless a measured, documented judgement says otherwise.
Who makes the determination?
You do — the decision-process owner, challenged by a second reader and framed with counsel's input. ReviewDesk records and evidences the determination; it does not make legal determinations.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. The test is described as ReviewDesk's worksheet frames it; consult counsel for your scope position.
Keep reading: 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? · How do you write principal reasons for an adverse decision that a person can actually use? · What must you tell job applicants when a hiring algorithm screens them? · How do ADMT notice and review duties apply to automated credit decisions? · 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.