ReviewDesk
← Back

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

How do you choose ADMT compliance software? (A buyer's checklist)

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

The short answer. Test the artefacts, not the pitch. Can it publish a hosted notice generated from your workflow? Generate a letter whose reasons come from a real decision record? Run one inbox where every request is docketed, clocked and answered? Keep a three-year file and export it cold? Show you a gap as a gap? Software that does those five things can carry your duties; software that demos well and does three of them will leave the missing two in your lap, at examination time.

What is choosing ADMT software?

ADMT compliance software is the tooling that operates automated-decision duties: pre-use notices, adverse-outcome explanations, correction and human-review handling, audit trails and evidence retention for regimes such as Colorado SB 26-189, the California CCPA ADMT regulations and Illinois HB 3773. The buyer's task is to evaluate such software against the duties' operational shape rather than against feature matrices.

Test one: the notice — generated, hosted, versioned

Ask to see a notice produced from a workflow you describe in the demo, then ask what happens when the tool or factors change. The right answers: the notice regenerates from the workflow record, the change is versioned with an effective date, and superseded versions remain retrievable for past applicants' files. The wrong answers involve document editors, PDF exports and the phrase 'you can customise the template'. A notice system built on documents will diverge from your operation within two vendor updates; a notice system built on the workflow record cannot diverge without leaving evidence, per the notice guide.

Check placement reality too: the page must be linkable into your application flow at a stable address, carry the request form, and render properly on a phone in the demo — not in a future release. Buy the surface applicants will actually judge, because that surface is your compliance in their eyes and in any examiner's sampling.

Test two: the letter — reasons from a record, or from a template?

Hand the vendor a fictional adverse decision with specific facts and watch the letter get made. The decisive question is where the reasons come from: entered per person from the decision record and rendered in a fixed five-part structure, or selected from a library of stock paragraphs. Stock-paragraph letters are the brush-off at industrial scale, and no jurisdiction note wrapped around them cures it. The standard is the anatomy in the examples guide — decision, disclosure, per-person reasons, jurisdiction note, routes — produced same-day, numbered and filed.

Then ask the audit question: show me where this letter lives in six months, with its number, its record and its outcome. Software that generates letters it cannot later produce has solved the easy half of the letter problem. The letter is evidence; buy the filing, not just the prose.

Test three: the inbox — one docket, real clocks, honest closure

File a test request in the demo and follow it. Does it land as a classified docket row attached to its case, or as a ticket in a general queue? Are deadlines attached with provenance and computed live, or reported weekly? Can a case close without an outcome note and a sent reply — and if it can, what does the vendor's 'answered' metric actually count? The inbox is where duties go to be lost quietly, so its demo behaviour under your probing is the most predictive ten minutes of the whole evaluation. Apply the standards of the inbox guide and the clocks guide without announcing them; watch which the product meets unprompted.

Ask also for the coverage number's definition and its data source. A vendor who computes coverage from the real docket — and shows an empty state for a new workspace — is telling you their honesty architecture in one feature. A vendor who demos with lush sample data pre-loaded is telling you the opposite, however the samples are labelled.

Test four: evidence — traces, sealing and the gap test

For platforms offering evidence packs, run the gap test from the honesty guide: file a deliberately incomplete decision record and read the output. Gaps rendered as gaps — with the pack still sealed and coherent — indicate an evidence system. Gaps silently completed indicate a storytelling system, and storytelling systems eventually write a story about you, to an examiner, that you will have to disown. Ask also how corrections are represented (appended with references, or edits), how versions are stamped per decision, and whether your own engineer can verify an export independently, per the chaining guide.

Retention is the quiet half of the test: the horizon enforced by the system (three years in the model these guides use), deletion executed on schedule, exports scoped and logged. Ask who can export, and where the export itself is recorded. Evidence about people deserves evidence about its own handling.

Test five: exit, pricing honesty and the vendor's own record

Ask the exit question early, while everyone is friendly: if we leave, what do we take, in what format, how fast, and what do you retain? The right answer is a complete export of your decision files on a defined timetable — the standard of the vault guide. Evasiveness here predicts the relationship's worst day with unusual accuracy. On pricing, map the tiers to your triggers rather than to the vendor's packaging: what does the second workflow cost, the clocks, the vault — and are definitions ('answered', 'recorded') constant across tiers, or does the cheap tier quietly mean less by them?

Finally, run the vendor's own artefacts through their own standards. Do their guides state their prices and plan contents exactly, the way the tiers guide does for ReviewDesk, or do capabilities blur upward in the telling? A vendor's honesty about their own product is the prior for their software's honesty about your records. Full disclosure in the spirit of this test: this guide is published by ReviewDesk, which is one of the products you might evaluate. Apply every test in it to ReviewDesk first — the free tier exists so you can, without a sales call — and keep the results.

The decision rule

Buy the system whose failures you have seen. Every product in this category will show you its strengths unbidden; the evaluation's value lies in the weaknesses you surfaced yourself — the gap test's behaviour, the closure rule's strictness, the exit answer's clarity. Weight those above feature counts, because the duties themselves are unforgiving about exactly those properties: completeness that admits its edges, closure that requires substance, records that leave with their owner. Software is a promise about your future conduct, made by engineers you will never meet. Choose the promise whose makers demonstrably fear the same failures you do.

And keep the evaluation's records. The tests you ran, the answers you received, the gaps you noted — filed with the procurement decision — are themselves the first entry in your programme's evidence: the diligence that chose the desk, conducted before the desk existed. Programmes that can show how their tooling was chosen, against which standards, start their examinations one chapter ahead. Begin as you mean to continue: on the record.

Your five actions for choosing ADMT software

  1. Generate a real notice and a real letter in the evaluation, from your facts — never judge from the vendor's samples.
  2. File a test request and probe closure: can a case close without an outcome note and a sent reply? If yes, discount every metric.
  3. Run the gap test on evidence features: file an incomplete record and read what the output does with the hole.
  4. Get the exit in writing before entry: complete export, defined format and timetable, retention stated.
  5. Apply every test to the vendor's own published claims — including, for this guide's publisher, the free tier, without a sales call.

Worked example: two demos, one gap test

A fictional buyer — a multi-state employer — evaluates two platforms. Platform One demos beautifully: animated dashboards, a letter library with forty templates, sample data showing 98% coverage. Platform Two's demo is plainer: a workflow described live produces its notice; a letter generates from facts the buyer dictates; the workspace is empty, so the coverage gauge honestly shows nothing yet. The buyer's gap test decides it: the incomplete test record renders on Platform One as a complete pack with the missing factor helpfully summarised from 'similar decisions'; on Platform Two it renders as 'not recorded', twice, with the pack sealed around the gaps.

The buyer chooses Platform Two, and its evaluation memo — kept with the procurement record — says why in one sentence: 'We will be examined on our worst decisions, and we choose the system that shows us our worst decisions accurately.' A year later, its first examination opens with that memo. The examiners read it, then the packs, in that order — and the order, the buyer reflects, was the whole game from the beginning.

Frequently asked questions

What is the most revealing test of ADMT software?

The gap test: file a deliberately incomplete decision record and read the evidence output. Systems that render the gap honestly can be trusted with your records; systems that complete it plausibly will eventually fabricate under examination conditions.

What should the letter test check?

That reasons are entered per person from the decision record and rendered in a fixed structure — decision, disclosure, reasons, jurisdiction note, routes — and that the generated letter is filed, numbered and retrievable months later with its case.

Why does the exit matter at purchase time?

Because your decision files must leave with you — completely, in a defined format, on a defined timetable. A vendor evasive about exit while courting you will not become generous during a dispute.

Should we trust demo dashboards?

Only as far as their data source. Coverage computed from a live docket, with an honest empty state for new workspaces, is a metric. Lush sample data in a demo is a story about a metric — evaluate the definition and the closure rules instead.

Is this guide legal advice?

No — and it is published by a vendor in this category, ReviewDesk, which is why it ends by inviting you to apply its tests to ReviewDesk's free tier first. It is a buyer's framework, not legal advice; consult counsel for procurement obligations.

Keep reading: How should you handle a formal complaint about an automated decision? · How do you prepare for future automated-decision rules without knowing what they will say? · What must a Colorado pre-use notice for an automated decision tool say? · What must a California ADMT pre-use notice include? · What does Illinois HB 3773 require when AI is used in employment 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.