Practical guide
What evidence shows exception management is operating?
An exception policy says when a rule may be departed from. Operating evidence shows whether real requests were identified, assessed, authorised, monitored, expired, and closed as designed. A neat register is not enough if requests bypass it, approvers lack authority, compensating measures are not performed, or expired exceptions remain active. This guide helps an organisation test its own.
An exception policy says when a rule may be departed from. Operating evidence shows whether real requests were identified, assessed, authorised, monitored, expired, and closed as designed. A neat register is not enough if requests bypass it, approvers lack authority, compensating measures are not performed, or expired exceptions remain active. This guide helps an organisation test its own approved exception process for AI-enabled work. It does not determine legal compliance or certify control effectiveness.
Define the control objective and population
Write a testable objective, such as: “Departures from approved AI data, tool, access, review, or decision rules are used only for a stated need, authorised by the right owner, constrained in scope and time, monitored, and removed or renewed through a fresh decision.”
Define the complete population for the review period. Reconcile the exception register with help-desk requests, access changes, security and privacy reviews, procurement records, incident logs, model or tool configuration, and owner attestations. A sample drawn only from the register cannot reveal unregistered bypasses.
NIST's AI Risk Management Framework calls for transparent, documented risk processes, clear roles, periodic review, monitoring, and response. NIST SP 800-53 Rev. 5 provides a flexible control catalog with assessment and monitoring families. The 2025 GAO Green Book emphasises documented risk assessment and response, preventive controls, and responsibility across organisational levels. These are authoritative design references, not automatic evidence that a local process works.
Build the evidence chain
For each exception, connect seven records.
### Request
Retain the requester, date, business need, affected rule, systems, data, users, workflow, and requested duration. The request should describe why the approved route does not work and which alternatives were considered.
### Risk assessment
Record affected risks, people, obligations, dependencies, likelihood and impact reasoning, specialist input, and uncertainty. A copied generic risk paragraph is not page-specific evidence.
### Decision
Retain the approver, decision right, date, disposition, rationale, scope, conditions, and conflicts of interest. Confirm the requester did not approve their own departure and that required security, privacy, legal, procurement, or workforce reviews occurred.
### Compensating measures
Translate approval conditions into observable actions with owners, cadence, evidence, and failure response. Examples include manual review, narrower access, additional logging, synthetic data, a blocked export, or a temporary fallback.
### Monitoring
Keep proof that measures actually ran: access records, review artifacts, alerts, sampled transactions, attestations supported by records, incident links, and unresolved findings. Evidence should cover the exception's full active period, not only its approval day.
### Expiry or renewal
Show reminders, expiry enforcement, access removal, configuration reversal, data disposition, and owner confirmation. Renewal requires a fresh need, current risk assessment, performance of prior conditions, and new authority. Repeated renewal is a signal to repair the standard route or reconsider the rule.
### Closure
Record whether the exception ended, became a separately approved standard, or was revoked after failure. Link residual actions and the next control review.
Test existence, completeness, accuracy, and operation
Use four evidence tests:
Select both representative and risk-based samples. Include denied requests, urgent requests, renewals, expired items, high-impact data, third-party tools, and incidents. Trace in both directions: register to operating systems and operating-system deviations back to the register.
- Existence: the sampled event occurred and the linked records are authentic.
- Completeness: all required fields, reviews, conditions, and periods are represented.
- Accuracy: dates, identities, scope, and status agree across independent systems.
- Operation: approval conditions, monitoring, expiry, and closure happened in practice.
Worked artifact: review an external AI workspace exception
A synthetic research team receives a 30-day exception to use an external AI workspace with de-identified interview excerpts. Approval requires a named five-person access group, field-level redaction, weekly sample review, export disabled, and deletion at expiry.
The reviewer assembles this test record:
| Evidence point | Expected | Observed | Disposition | | --- | --- | --- | --- | | Request and scope | Five named users, de-identified excerpts | Matches approval | Pass | | Specialist review | Privacy and security decisions before access | Both retained | Pass | | Access | Five users for 30 days | Sixth user added on day 12 | Fail and investigate | | Redaction | Pre-upload field check | 18 samples retained; one defect corrected | Pass with incident link | | Export | Disabled throughout | Configuration logs agree | Pass | | Weekly review | Four review records | Week three missing | Fail | | Expiry | Access removed and workspace deleted | Access removed; deletion confirmation absent | Open repair |
The register says “expired,” but the evidence does not support clean closure. The reviewer opens a finding for the extra user, missing weekly review, and absent deletion confirmation. The accountable owner must obtain deletion evidence, investigate access provisioning, and decide whether the exception process or the standard tool route needs repair. Until then, the control is not reported as fully operating for this sample.
Report results without overstating them
Report the population sources, reconciliation differences, sampling method, period, tests performed, exceptions found, unresolved items, owners, and retest dates. Separate design gaps from operating failures. A failure in one sample does not automatically describe every exception, while a clean sample does not prove that no bypass exists.
Useful indicators include unregistered deviations, self-approval, missing specialist review, overdue conditions, expired access, repeated renewals, closure without evidence, and time to resolve findings. Counts need denominators and definitions. Trend changes may reflect better detection rather than more or less underlying risk.
Failure conditions and follow-up
Escalate when the population cannot be reconciled, authority is unclear, high-impact exceptions lack specialist review, compensating measures are repeatedly missed, expiry cannot be enforced, or records conflict. Retest targeted repairs with new operating evidence. Review the control when the same exception recurs, because repeated departures may reveal an unusable rule or an unaddressed workflow need.
Two Prune can assess how a participant identifies constraints, evaluates evidence, verifies records, and communicates a control finding in a simulation. It does not audit an organisation, certify a control, or make employment decisions.
Sources
NIST, AI Risk Management Framework Core
- NIST SP 800-53 Revision 5, Security and Privacy Controls for Information Systems and Organizations
- U.S. Government Accountability Office, 2025 Standards for Internal Control in the Federal Government
Sources: [1] [2] [3]
Sources
- 1.NIST, AI Risk Management Framework Core · National Institute of Standards and Technology
- 2.NIST SP 800-53 Revision 5, Security and Privacy Controls for Information Systems and Organizations · National Institute of Standards and Technology
- 3.U.S. Government Accountability Office, 2025 Standards for Internal Control in the Federal Government · U.S. Government Accountability Office