ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
How long must you keep ADMT decision records — and what should the audit trail hold?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. Keep the whole decision file — the notice version in force, the factors recorded, the letter sent, the review request and its outcome — for three years, on one timestamped timeline. Storage alone is not a trail: an audit trail is a sequence of events you can read back as a story. ReviewDesk builds that trail automatically as the workflow runs, and paid plans export it from the 3-year evidence vault.
What is the ADMT audit trail and three-year records?
An ADMT audit trail is the timestamped record of everything that happened around an automated decision: the workflow and notice version, the decision and its recorded factors, every letter generated, every request received, every review outcome. ReviewDesk's retention model is three years, matching the record horizon its plans are built around — the evidence vault holds the file for that period with full export.
What belongs in the trail — and what does not?
If it affected the person, it belongs in the trail. The notice they were shown and when. The decision, its date, and the factors the trace recorded. The letter, numbered, dated, as sent — not the template it came from. Requests: review, correction, opt-out, each with its arrival time. Outcomes: who decided, what changed, what was communicated. Configuration events matter too — threshold changes, factor changes, notice updates — because a decision can only be judged against the system as it stood that day.
What does not belong is bulk without meaning: raw model internals nobody can interpret, duplicated exports in five folders, personal data beyond what the decision needed. A trail stuffed with everything explains nothing. The discipline is the same as the trace discipline in the evidence-pack guide: record the elements a reviewer needs to reconstruct the decision, in the form they need them, and resist archivist maximalism.
Why timestamps are the trail's backbone
Disputes about automated decisions are usually disputes about order. Was the notice live before the tool ran? Did the explanation arrive inside thirty days? Did the opt-out precede the decision? Memory cannot answer those questions; only timestamps can, and only timestamps written at the moment of the event — not reconstructed when the question arrives. ReviewDesk stamps every audit event as it happens, which is why the trail can settle sequence questions in seconds rather than in correspondence.
Protect the stamps' credibility. A trail anyone can quietly edit is a diary, not evidence. The traces layer adds hash-chaining — each record linked to the previous, so alteration shows — which is explained in the hash-chain guide. Even without the traces engine, the operational rule holds: corrections are new entries that reference the old, never silent edits. A trail with a visible correction is more believable than a trail with no corrections at all.
How does three-year retention work in practice?
Three years per decision file, counted from the decision — not three years of whatever happens to survive in a shared drive. Practically, that means retention must be a system property, not a person's habit: files enter the vault as the workflow runs, the vault enforces the horizon, and export produces the complete file on demand. Retention also means deletion discipline at the other end — holding personal decision records indefinitely beyond their horizon is its own risk, and a vault with a defined horizon lets you honour both duties at once.
Test the export before you need it. Once a quarter, pick a closed case at random and export its file cold: notice version, trace, letter, requests, outcomes. Time how long a competent stranger takes to reconstruct the story from the export alone. If the answer is measured in phone calls rather than minutes, the trail is storage, not evidence. The export mechanics live in the evidence-vault guide.
Who should be able to read the trail?
Fewer people than can read the inbox, and every reading by someone with a reason. The trail contains the sensitive core of your decision-making: factors, reasons, applicants' disputes. Role discipline applies — reviewers see their cases, administrators see configuration, leadership sees aggregates. ReviewDesk's workspace model keeps the trail inside the account that owns the workflow, and the audit itself records administrative events, so the trail watches its own custodians.
External readers arrive in defined moments: the person themselves exercising their rights, counsel preparing a response, a regulator examining a complaint. In each case the export should be scoped to the question — one person's file, one workflow's quarter — rather than the keys to the archive. Scoped, readable exports are a feature of the vault design, not an afterthought, and they are the difference between cooperating with an enquiry and drowning in one.
What breaks trails in real organisations?
Four ordinary habits. Decisions leaking outside the system — a manager declining by phone, a recruiter rejecting inside the screening platform without the outcome reaching the desk. Letters sent from personal inboxes, unlogged. Spreadsheets maintained 'temporarily' for two years. And tool changes made directly in the vendor console with no record. None of these is malice; all of them sawtooth the trail until it cannot tell a continuous story about any single person.
The countermeasure is routing, not rules: make the desk the easiest place to do the work. When the letter generator is faster than drafting in email, letters get generated and logged. When the inbox is where requests naturally land from the notice page, requests stop arriving as stray emails. ReviewDesk is designed around that gravity — the free workflow already produces the trail, because the trail is a by-product of doing the work in the desk rather than beside it.
How does the trail serve the business, beyond defence?
It is the raw material of every improvement this programme offers. Review outcomes reveal which factors generate the most successful challenges — usually the factors with the worst data quality. Letter sampling reveals drift in reason quality. Opt-out and correction volumes reveal notice problems. The annual risk assessment is, in large part, the trail read at altitude: see the risk-assessment guide. Teams that treat the trail as a compliance cost miss that it is also the only honest dataset about their own decision quality.
It also disciplines vendors. When your trail records factor-level outcomes, renewal conversations stop being about dashboards the vendor chose and start being about overturn rates, correction rates and time-to-explanation in your own operation. That is purchasing power no contract clause provides on its own.
Your five actions for the ADMT audit trail and three-year records
- Record every person-affecting event on one timeline: notice version, decision, factors, letters, requests, outcomes.
- Timestamp at the moment of the event; correct by appending referenced entries, never by silent edits.
- Set the retention horizon to three years per decision file, enforced by the vault — not by individual habit.
- Export-test one random closed case every quarter and time how long a stranger needs to reconstruct it.
- Scope external exports to the question asked — one person, one workflow, one period.
Worked example: a complaint answered from the trail in one afternoon
A fictional care provider — Willowbrook Care — declines an applicant for a support-worker role after automated screening. Four months later a complaint arrives claiming the applicant never knew a tool was used and never received reasons. Willowbrook exports the file: notice version 3, live from January, linked in the application flow; the decision of 14 February with three recorded factors; letter RD-1187 generated the same day with those reasons; no review requested. The answer assembles itself in an afternoon, with documents rather than recollections.
The same export shows one honest gap: the applicant's correction request, sent by email in March, was answered but logged a week late when an officer backfilled it — visible as a late entry, timestamped as such. Willowbrook discloses the gap and fixes the routing so email corrections enter the inbox the same day. A trail that shows its own blemish, corrected, is worth more than a spotless story nobody believes.
Frequently asked questions
How long should ADMT records be kept?
ReviewDesk's model — and the horizon its evidence vault is built around — is three years per decision file. That covers the notice version, the decision record, letters, requests and outcomes for each person.
What is the difference between storage and an audit trail?
Storage holds files; a trail holds a timestamped sequence of events that can be read back as the story of one decision. If reconstructing what happened requires interviewing staff, you have storage, not a trail.
Can records be corrected in the trail?
Yes — by appending a new entry that references the original, never by editing history silently. ReviewDesk's decision traces go further with hash-chaining, so any alteration of a past record is detectable.
Does the free tier keep an audit trail?
Yes. The timestamped audit trail is part of every plan including free. Paid plans add the 3-year evidence vault with full export and the three-regime evidence engine.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. Retention is described as ReviewDesk's model; consult counsel for your obligations.
Keep reading: What goes into California's annual ADMT risk assessment (due 1 April 2028)? · 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? · 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.