Records retention for AI-enabled knowledge work

Inventory materials before assigning periods

Map one bounded workflow from intake to final action. List the source records supplied to the system, prompt or instruction, retrieved context, intermediate output, final output, human changes, approval record, exception, audit event, system configuration, and downstream copy. Identify where each item is stored, whether it contains personal or confidential information, and which team owns it.

Do not assume that a prompt is disposable or that every log is a permanent record. The significance of a material depends on what business action it documents, the applicable authority, and whether another system holds the authoritative version. Preserve legal holds and other suspensions through the organization's qualified records and legal processes.

The US National Archives' 2026 guidance on applying the Federal Records Act to AI materials covers inputs, outputs, data, audit trails, software, and other AI materials for US federal agencies. It states that federal AI records require an approved records schedule before disposal. That jurisdiction-specific guidance is useful for classification discipline but is not a retention rule for every organization.

Sources: [1]

Build the AI material disposition map

Create one row per material class with: workflow and purpose; example; system of record; record status; authority or schedule; retention trigger; minimum or maximum period where applicable; privacy category; legal-hold behavior; deletion or transfer method; owner; verification evidence; and exception route.

Keep the trigger explicit. “Seven years” is incomplete without saying whether the clock starts at creation, final approval, contract expiry, account closure, or another event. Where a generated draft is merely transitory, distinguish it from the approved decision record. Where a prompt contains substantive instructions or evidence used to reach a consequential result, document why it is or is not part of the retained record.

The UK Information Commissioner's storage limitation guidance recommends standard retention periods for categories of personal data, review at the end of those periods, and erasure or anonymisation unless longer retention is justified. It is UK data-protection guidance, not a universal schedule, and organizations need jurisdiction-specific advice.

Sources: [2]

Worked artifact: a synthetic policy-answer workflow

Consider a synthetic internal assistant that answers staff questions from approved policies. Its material map classifies the published policy and final approved correction notice under existing corporate schedules. Routine cached passages and abandoned drafts are transitory under the organization's approved rule. Access logs follow a security-log schedule. A user-submitted medical detail is restricted, separated from the general transcript, and routed for privacy review.

A deletion test finds that removing the conversation from the user interface leaves a copy in an analytics store. The workflow therefore fails its disposition check. The owner pauses automated deletion claims, maps the hidden copy, updates the deletion job, reruns the test using a unique marker, and retains the test evidence. The scenario is synthetic and does not establish legal compliance or control effectiveness.

Verify disposition end to end

Select samples that have crossed each trigger. Trace a unique test marker through the interface, primary store, search index, analytics copy, backup process, export, and downstream integration. Confirm that deletion, anonymisation, transfer, or continued retention matches the approved disposition and that an active hold blocks destruction where required.

Check both false deletion and false retention: did the system destroy evidence it was required to keep, or retain material past its justified period? Inspect failed jobs, orphaned accounts, model-training or evaluation reuse, vendor copies, and restoration from backup. Record the date, scope, systems checked, result, exception, correction, retest, and reviewer.

Revisit the map when a new tool, data source, model provider, workflow purpose, jurisdiction, or downstream integration is introduced. Two Prune can observe how a participant handles retention evidence and uncertainty in a bounded assessment. It does not operate customer retention systems, provide legal advice, or certify compliance.