ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
What should you ask a software vendor before you rely on their tool for consequential decisions?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. Ask what the tool actually decides, which factors it uses in a form you can print in a letter, how its versions change and how you will be told, what records it keeps and exports, and how a human review overturns its output in practice. A vendor who can answer those five can support your duties; a vendor who cannot is selling you their roadmap as your compliance. The deployer's duties remain the deployer's — ReviewDesk's workflows assume a vendor who passed this interrogation.
What is vendor ADMT due diligence?
Vendor ADMT due diligence is the structured questioning of a decision-tool supplier before and during deployment: the tool's decision role, factor transparency, version and change practices, record-keeping and export, human-override mechanics, and data-correction handling. It matters because the regimes place duties on the business making decisions about people, while the facts needed to discharge those duties sit inside the vendor's product.
Why does the duty stay with you when the tool is theirs?
Because the person affected has a relationship with you. The applicant applied to you, the tenant rents from you, the borrower owes you. Notices carry your name, letters go on your letterhead, and the review is of your decision. The vendor's contract may spread costs and indemnities, but it cannot transfer the experience — or, in the frameworks' design, the responsibility — of the business that chose to decide this way. That asymmetry should shape procurement: you are not buying a score, you are buying the evidence supply chain for every decision the score touches.
It also shapes the questions. Everything in this guide is really one question in five costumes: when this decision is challenged — by the person, by your own reviewers, by an examiner — will your product hand us the truth about it, quickly, in a form we can show? Vendors who understand that question are partners in the programme described across these guides. Vendors who hear it as hostility are telling you something procurement should record.
Question one: what exactly does the tool decide?
Pin the decision role in the vendor's own words, in writing: recommends, ranks, bands, auto-declines below a threshold — and what portion of outcomes are fully automated. Compare the answer with the behaviour you observe in a pilot: acceptance rates and review times, the measurements from the scope-test guide. A gap between the described role and the measured role is the single most valuable finding due diligence can produce, because it recalibrates your notice, your scope worksheet and your review staffing before the public does it for you.
Ask also what the tool does not decide — the negative space defines your remaining human process. Vendors describe capabilities; deployers must describe the decision as a whole. The workflow you build in ReviewDesk will be drawn around the vendor's honest answers, and an answer that later proves wrong becomes a contract conversation with the pilot data attached.
Question two: which factors, in printable form?
Require the factor list at the resolution your letters need: the inputs that materially count, in terms a letter can carry to the person. 'Proprietary model features' is not a factor list. You do not need the model's mathematics; you need to be able to write, truthfully, 'these things about you counted'. Test the answer by drafting a sample adverse letter from it — the exercise in the principal-reasons guide. If the letter cannot be written, the explanation duty cannot be discharged, and no other feature compensates.
Ask how factor importance is exposed per decision: does the product return the principal factors for this applicant, or only the global list? Per-decision factors are what traces record and letters quote. A product returning only outcomes forces you to guess reasons — and guessed reasons, as the evidence guides keep repeating, are worse than an honestly recorded gap.
Question three: how do versions change, and how will we know?
Models change silently in modern software; evidence cannot. Ask for version identification on every decision, advance notice of material changes, and a changelog that describes behavioural changes — thresholds, factors, data sources — not marketing summaries. Your traces record the version per decision, per the evidence-pack guide; a vendor who cannot version their own outputs makes that record impossible and your cohort analysis fictional.
Tie change notice to your own duties in the contract: a material model change may require a notice update and a scope re-test on your side before the version goes live on your traffic. The vendor's release calendar and your compliance calendar must meet somewhere; procurement is where they meet first, or they meet later in an incident review.
Questions four and five: records, export and override
Records: what does the product log per decision, for how long, and in what export form — documented, standard, complete enough for your own engineer to verify against your traces? Export you cannot independently check is evidence you rent, per the chaining guide. Override: when your reviewer overturns an outcome, what happens inside the product — is the override recorded against the decision, does the product learn or flag, can the original output be re-presented later as if the override never happened? A tool that quietly re-asserts its original answer downstream will corrupt your decision history while looking helpful.
Corrections complete the set: when source data is wrong, how is it corrected — in the product, at the source, with what propagation time — and how will you know the correction took effect for the next decision about that person? The correction experience your letters promise is built from the vendor's true answers here, per the correction guide. Promise only the correction speed the vendor conversation substantiated.
How should the answers be captured?
In the contract and the worksheet, not in a slide deck. The scope worksheet records the decision-role determination with its measurements. The contract carries versioning, change notice, export format, retention and correction service levels as obligations, with the diligence answers annexed where they are factual claims. Renewal is a re-interrogation, lighter but real: versions shipped, changes notified as agreed, export tested once more, correction statistics reviewed — your own corrections log gives you the agenda, since it has been quietly recording the vendor's data quality all year.
Where a vendor fails the interrogation but the tool is otherwise right, scope your deployment to what the answers support: human-decides mode with the tool as one documented input, narrower decision types, or a contractual remediation plan with dates. What you must not do is deploy the full automation and borrow the missing evidence from optimism. The duties do not accept optimism as a record format — and neither, in the end, will the first person who asks to see their file.
Your five actions for vendor ADMT due diligence
- Get the tool's decision role in writing and verify it against pilot acceptance rates and review times.
- Require per-decision principal factors at letter-printable resolution — draft a sample adverse letter during diligence.
- Contract for per-decision versioning, advance notice of material changes and behavioural changelogs.
- Test the record export yourself, and confirm overrides and corrections persist inside the product.
- Annex the factual answers, re-interrogate at renewal, and scope deployment to what the answers support.
Worked example: the demo that failed the letter test
A fictional employer — Halden Group — shortlists two screening vendors. Vendor A's demo is dazzling; asked for the per-decision factors behind a sample decline, it returns a confidence score and a white paper. Vendor B's demo is plainer; its API returns the decision, the model version and the four principal factors per applicant, and its changelog describes threshold changes in plain terms. Halden drafts sample adverse letters from both: B's writes itself in ten minutes; A's cannot be written truthfully at all.
Halden signs with B, annexes the factor and versioning commitments, and six months later its first evidence pack assembles from traces without a single vendor support ticket. The procurement lesson is now written into Halden's playbook: the demo evaluates the tool's intelligence, but the letter test evaluates its evidence — and in a consequential-decision programme, the evidence is the product. Vendor A, informed of the reason, ships per-decision factors the following year. Diligence, done in the open, improves markets as well as purchases.
Frequently asked questions
Who carries the ADMT duties — deployer or vendor?
The business making decisions about people carries the duties the person experiences: notice, explanation, review. Vendors supply the facts those duties consume, which is why factor transparency, versioning and export belong in procurement and contract.
What is the single most revealing diligence test?
Draft a sample adverse letter from the vendor's per-decision output. If the principal reasons cannot be written truthfully from what the product returns, the explanation duty cannot be discharged with that product as configured.
What should the contract cover?
Per-decision version identification, advance notice of material model changes with behavioural changelogs, record retention and documented export formats, correction handling and propagation, and override persistence inside the product.
What if a good tool fails part of the interrogation?
Scope the deployment to what the answers support — human-decides mode, narrower decision types, or a dated remediation plan. Never deploy full automation on borrowed optimism about the missing evidence.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. The questions are a diligence framework from ReviewDesk's operating model; consult counsel for your contracts.
Keep reading: ADMT duties explained for software vendors and the businesses that deploy their tools · How do you set up your first ADMT workflow, step by step? · What makes a hosted pre-use notice page work better than a PDF or a policy link? · How should the review-request form on a notice page be designed? · How do you write an ADMT notice in plain language that people actually understand? · 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.