What records should I keep when an AI agent works in a regulated workflow?
Keep the decision trail end to end: the inputs with their check states, the agent's steps in order, every approval gate with its recorded decision, and the outputs as delivered, retained for the period your rules set. Review the receipt and actual deliverable before accepting a handover.
Scope: general information about record-keeping, not regulatory or legal advice. Which records a given regime requires, and for how long, is set by that regime and your compliance function; this page describes the record structure that regimes consistently ask for.
Regulated workflows differ from ordinary client work in one structural way: the examiner is not a counterparty but an authority, and the examination can arrive years after the work, conducted by people who were not present, against written requirements. Records built for clients, who can be phoned, fail examinations built for strangers. The decision trail is designed for the stranger.
Its emphasis falls on the joints ordinary records underweight. Inputs carry their states because regulated decisions are judged partly on what the decider could have known: a decision resting on a source that was never checked reads differently from the same decision on resolved ground, and the record must show which. Approval gates carry named deciders and their recorded decisions because "the system approved it" is the sentence examiners exist to reject; a human gate that fired is a step, with a name and a timestamp, or it did not happen in any sense an audit accepts. Outputs are kept as delivered, versioned, because the regulated question is always about the specific artefact that reached the world.
Retention follows the regime's clock, not the firm's convenience, and the exports discipline applies with extra force: records an authority may demand in year five cannot live solely inside a vendor relationship that may end in year two. Build the trail, export it on schedule, and let the compliance function map it to the regime's clauses while the mapping is still a desk exercise.
Checklist: the regulated-workflow record
1. Inputs: every material input with its check state, saved at run time. 2. Steps: the agent's actions in order, each with its saved result and receipt. 3. Gates: every approval point as a step, naming the decider, the decision, and the moment; no gate lives in email. 4. Outputs: as delivered, versioned, with the delivery record attached. 5. Access: the inventory entries for every agent and person touching the workflow. 6. Retention and export: the regime's period applied per class; exports verified from your own storage on the annual test. 7. Mapping note: which clause each record class answers, written by the compliance function while the trail is young.
Worked example
The gate layer in live form: https://zambo.dev/handover/45434a17-4afc-486a-9646-e689bcee4297 carries a client's recorded decision on the work itself, decided_by the client, with its acceptance receipt. In a regulated workflow that same pattern sits at every approval point inside the job, which is what converts "a human reviewed this" from a policy sentence into an examinable fact.
Frequently asked questions
Our regulator has not asked about AI records yet. Is this premature?
Examinations reach backwards. The question is never whether today's work will be examined under today's interest, but whether it will survive the interest of the year it is examined in. Records are cheap in the present tense and unavailable in any other.
Does the EU AI Act change what belongs in the trail?
For systems in its scope, its record-keeping articles point the same direction as this checklist: logs and records sufficient to reconstruct what the system did. The timing duties phase in under the Digital Omnibus changes (Annex III high-risk duties deferred to 2 December 2027; Article 50 transparency duties from August 2026). Your compliance function maps the clauses; the trail structure holds either way.
How do we handle records the regime requires us to delete?
Deletion duties outrank retention habits: where a regime or a data right requires deletion, delete on schedule and record the deletion event (what class, when, under which duty) without keeping the deleted content. The trail records its own pruning the same way it records everything else.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai