Exception management for AI-enabled knowledge work

Separate exceptions from incidents and defects

An exception is known before or during continued operation and receives an authorised disposition. An incident is an event that occurred and may require containment, communication, investigation, or recovery. A defect is a failure in a system or process that may be repaired without authorising continued deviation. Keep the records linked but distinct so approval does not erase an incident and a bug ticket does not silently authorise risk.

State the exact rule, control, or operating condition that is unmet. Identify the workflow, data, users, actions, geography, environment, and time window covered. If the scope cannot be stated precisely, the exception is not ready for decision.

Create an exception decision record

Record the requestor, accountable business owner, control owner, reviewers, approving authority, rationale, alternatives considered, affected parties, plausible consequences, existing evidence, compensating measures, monitoring signal, stop condition, start date, expiry date, and closure evidence. Link the source policy and any relevant test, incident, vendor, or change record.

NIST's AI RMF Core describes Manage outcomes that include prioritising documented risks, selecting responses, documenting residual risk, assigning responsibilities, and maintaining response, recovery, incident, and change plans (NIST AI RMF Core). The framework is voluntary and does not approve a specific exception. It supports making the risk treatment and continuing obligations visible.

Sources: [1]

Evaluate compensating measures as claims

Do not list a compensating measure without showing how it addresses the defined exposure. For each measure, state the owner, operating step, evidence produced, test method, coverage limitation, and failure response. A requirement for human review is incomplete unless the reviewer has the necessary authority, information, time, and escalation path and the review leaves a record.

Separate design from operation. A drafted procedure shows intended behavior. A dated sample, log, or observation can show what occurred within a bounded period. Missing evidence should be labelled unknown, not converted into a passing result.

Worked artifact: a synthetic retrieval exception

Consider a synthetic internal research assistant that normally may use only an approved document repository. A team requests a 30-day exception to include a legacy archive while migration is completed. The archive contains useful historical decisions but has inconsistent access labels and duplicate policy versions.

The exception record limits use to five named analysts, blocks external publication, requires every cited archive item to be checked against the current repository, and logs each retrieval. The stop condition is any restricted record returned to an unauthorised analyst. Weekly sampling checks 20 retrievals for permission, currency, and citation traceability. The first sample finds two superseded policies, so those folders are removed and the affected briefs are re-reviewed.

The approver renews the exception for only seven days, not another 30, because migration is nearly complete and the evidence shows a currency problem. Closure requires the archive connector to be disabled, the access log retained, open briefs checked, and the approved repository confirmed as the only source. This is an original synthetic example, not evidence that the controls would work in another environment.

Make expiry and escalation executable

An expiry date should change system or workflow state. Before expiry, route the record to close, renew with new evidence, narrow, or stop. Do not auto-renew. A renewal is a new decision that must consider incidents, monitoring results, changed dependencies, and whether the original rationale still holds.

The GAO 2025 Green Book describes internal control as an ongoing management process supporting operations, reporting, and compliance objectives and includes documentation expectations across its principles (GAO Green Book). It governs U.S. federal internal control and is not a universal exception standard. Its emphasis on designed, implemented, operated, and monitored controls supports retaining evidence that compensating measures actually operated.

Escalate when scope expands, a stop condition occurs, monitoring is missed, a measure fails, an incident is linked, the owner changes, or evidence becomes stale. Preserve the original request, all decisions, and the final closure or termination reason.

Sources: [2]

Verify closure and preserve boundaries

Before marking an exception closed, confirm that the deviation ended or the underlying rule changed through the authorised process. Test removal of temporary access, connectors, prompts, permissions, or manual workarounds. Reconcile open incidents and affected outputs. Store enough evidence for a later reviewer to reconstruct the decision without relying on the approver's memory.

Exception management can improve accountability and prevent silent permanence. It does not establish legal compliance, certify control effectiveness, or establish that operation is safe. Two Prune can observe how a participant handles a bounded exception scenario in a synthetic assessment. It does not approve customer exceptions, operate production controls, or make employment decisions.