Practical guide
Exception management policy for AI-enabled work
An exception is a time-limited, explicitly authorised departure from an approved rule. It is not informal permission, a permanent alternative route, or a way to relabel noncompliance. This policy template defines who may request, assess, approve, monitor, renew, and close exceptions involving AI tools, data, access, review, procurement, or workflow controls. Adapt it to the organisation's.
An exception is a time-limited, explicitly authorised departure from an approved rule. It is not informal permission, a permanent alternative route, or a way to relabel noncompliance. This policy template defines who may request, assess, approve, monitor, renew, and close exceptions involving AI tools, data, access, review, procurement, or workflow controls. Adapt it to the organisation's obligations and authority structure. Adoption does not establish compliance or prove that the process operates effectively.
1. Purpose and scope
The policy objective is to permit a necessary departure only when the need is documented, alternatives have been considered, risk is understood, authority is valid, compensating measures are workable, and the departure is constrained in scope and time.
The policy applies to employees, contractors, systems, vendors, data, and business processes subject to approved AI-related rules. It covers planned, urgent, renewed, denied, revoked, and expired requests. It does not override law, contract, safety restrictions, or obligations that the organisation has no authority to waive.
NIST's AI Risk Management Framework calls for transparent policies, clear roles, documented risk tolerance, monitoring, and periodic review. NIST SP 800-53 Rev. 5 provides adaptable control, assessment, and continuous-monitoring concepts. The 2025 GAO Green Book describes documented risk assessment, response, preventive controls, responsibility, and residual risk. They inform this template but do not supply local approval authority.
2. Roles and decision rights
Delegations must state subject, risk threshold, duration, and escalation route. Conflicts of interest are disclosed and resolved before approval.
- Requester: explains the need, affected rule, scope, duration, alternatives, and proposed safeguards. The requester may not approve their own request.
- Rule owner: confirms the rule and whether an approved route already meets the need.
- Risk owner: accepts or rejects residual risk within delegated authority and remains accountable during the exception.
- Specialist reviewers: provide required security, privacy, legal, procurement, accessibility, safety, records, or workforce advice. Advice does not replace the accountable decision unless local authority says otherwise.
- Exception administrator: checks completeness, maintains the register, sends reminders, and coordinates expiry. This role does not silently convert an incomplete request into an approval.
- Control owner: implements and evidences compensating measures.
- Independent reviewer: tests the process and operating evidence without approving the original request.
3. Required request record
Every request must identify:
An incomplete request is returned, not provisionally approved through silence.
- the rule and exact departure;
- business need and consequence of denial;
- affected systems, models, tools, vendors, data, people, locations, and decisions;
- requested start, end, and maximum scope;
- alternatives considered and why approved routes are insufficient;
- relevant obligations, dependencies, and prior related exceptions;
- risk analysis, uncertainty, and proposed compensating measures;
- monitoring evidence, owners, cadence, and failure response;
- rollback, deletion, access removal, and closure plan; and
- required reviewers and decision authority.
4. Assessment and decision criteria
The rule owner and reviewers determine whether the departure is actually required and within scope. The risk owner then considers necessity, proportionality, affected people and data, reversibility, duration, detectability, dependencies, cumulative exposure, precedent, and residual risk after compensating measures.
Approval requires all of the following:
The written decision records approve, approve with conditions, deny, or escalate; rationale; scope; conditions; approver; specialist advice; effective time; expiry; and revocation triggers. Approval applies only to the stated version and purpose.
- a defined need that an approved alternative cannot reasonably meet;
- authority to permit the departure;
- bounded users, data, systems, purpose, and time;
- feasible compensating measures with named owners;
- monitoring capable of detecting breach or control failure;
- an enforceable end state; and
- residual risk within delegated tolerance.
5. Compensating measures
Conditions must be observable. Examples include narrower access, synthetic or de-identified data, blocked export, manual review, additional logging, transaction limits, dual approval, isolated environments, or more frequent sampling.
Each condition states who performs it, when, what evidence is retained, the success criterion, and what happens on failure. A generic instruction to “use caution” is not a compensating measure.
The exception does not start until required conditions are implemented and verified, unless the emergency procedure explicitly authorises a shorter sequence.
6. Duration, expiry, and renewal
Set the shortest practical duration and a fixed expiry. The system or administrator sends reminders early enough for closure or a new decision. At expiry, access, configurations, integrations, data, and temporary procedures revert or are removed as specified. Status does not remain active merely because review is late.
Renewal is a fresh decision. It requires current need, updated risk analysis, evidence that prior conditions operated, incidents and changes, a closure alternative, new dates, and current authority. Repeated renewals trigger review of the underlying rule or standard route. They do not automatically justify permanence.
7. Emergency exceptions
An emergency route may be used only for defined urgent conditions, such as preventing immediate material harm or restoring an essential service. It requires a named emergency authority, minimal scope, enhanced logging, and a short fixed expiry. Required retrospective review occurs within the policy's stated interval.
If retrospective review is denied or incomplete, the departure ends and the owner initiates incident, rollback, and remediation procedures. “Urgent” is not a substitute for foreseeable planning.
8. Monitoring, breach, and revocation
The control owner records performance of every condition for the full active period. The risk owner reviews monitoring at the approved cadence. The exception administrator tracks overdue evidence, scope changes, incidents, and expiry.
Escalate and consider suspension or revocation when scope expands, a condition fails, an incident occurs, evidence is missing, authority changes, the underlying risk increases, or an approved alternative becomes available. Unauthorised departures follow the organisation's incident and accountability processes; they are not backdated into exceptions.
9. Closure and records
Closure records the disposition, rollback and access removal, data handling, unresolved actions, incidents, lessons, and owner confirmation. An exception may end, be revoked, or be replaced by a separately reviewed standard change. The original approval alone is not closure evidence.
Retain requests, advice, decisions, conditions, monitoring, renewals, incidents, and closure records under the applicable records schedule. Access to sensitive records should be limited without making oversight impossible.
10. Oversight and policy review
Oversight reports should use defined denominators and distinguish requests, approvals, denials, active items, emergency items, renewals, expired access, missed conditions, incidents, and unregistered departures. Trends may reflect detection changes rather than underlying risk changes.
Review this policy on a defined cadence and after material incidents, repeated exception patterns, major legal or technology change, or evidence that the standard route is unusable. Independent testing should reconcile the register to operating systems and sample actual performance. That testing is separate from this policy's design.
Worked artifact: approve a bounded research exception
A synthetic research team asks to use an external AI workspace for 30 days because the approved tool cannot process a required file format. The request covers five named researchers and de-identified interview excerpts.
| Policy field | Decision record | | --- | --- | | Rule departure | Temporary use of a non-standard external workspace | | Alternatives | Conversion failed required structure; manual processing would miss the deadline; approved local tool remains the default | | Scope | Five users, one project, de-identified excerpts, no direct identifiers, 30 days | | Conditions | Field-level redaction before upload, named access group, export disabled, weekly sample review, enhanced logs | | Evidence | Redaction checklist, access report, configuration capture, weekly review record, deletion confirmation | | Authority | Risk owner approves after privacy, security, and procurement advice | | Revocation triggers | Identifier exposure, extra user, export enabled, missed weekly review, or approved alternative becoming available | | Expiry and closure | Access automatically removed on day 30; workspace content deleted and independently confirmed |
The team may begin only after the access group, export block, and logging are verified. A sixth user request is not absorbed into the approval; it requires a scope amendment and current decision. If week-three review is missed, the risk owner decides whether to suspend the exception. Renewal requires evidence from the first 30 days and a new analysis. The example demonstrates policy application, not proof that the controls operated.
Two Prune can assess how a participant interprets constraints, evaluates an exception request, verifies conditions, and communicates residual uncertainty in a simulation. It does not approve organisational exceptions, certify control effectiveness, 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