Practical guide
AI change plan verification checklist
A change plan can look complete while still hiding an unowned dependency, an unsupported claim, or a rollback step that no one can execute. This checklist is for reviewing a finished plan before approval or rollout. It does not replace the plan itself. It creates a traceable disposition for each material gap: pass, repair, escalate, or accept with a named owner and review date. Use the checklist.
A change plan can look complete while still hiding an unowned dependency, an unsupported claim, or a rollback step that no one can execute. This checklist is for reviewing a finished plan before approval or rollout. It does not replace the plan itself. It creates a traceable disposition for each material gap: pass, repair, escalate, or accept with a named owner and review date. Use the checklist with the approval record, source packet, process maps, implementation schedule, communications, training materials, risk records, and proposed monitoring evidence open. AI can help locate contradictions and missing fields, but the accountable reviewers must verify the underlying records and make the disposition.
Assemble the review packet
Freeze the plan version being reviewed and give it an identifier. Add the approved decision, current and future process, affected-role map, dependency register, control requirements, rollout schedule, support route, and evidence plan. Record the reviewer, review date, plan owner, change authority, and specialist reviewers.
The UK Government's service-management guidance describes controlled change through a change authority, owner, window, and back-out plan. The Government of Canada's Outcome Management Guide connects change with communication, preparation, monitoring, and reinforcement. NIST's AI Risk Management Framework also calls for documented roles, monitoring, and clear human oversight. Together, these sources support checking operational control, adoption work, and AI-specific traceability as separate questions.
Run eight verification gates
Mark every item pass, repair, escalate, or accepted exception. A blank item is not a pass.
### 1. Authority and boundary
### 2. Affected work
### 3. Actions and dependencies
### 4. Communication and enablement
### 5. Guardrails and recovery
### 6. AI traceability
### 7. Evidence and monitoring
### 8. Approval and retained disposition
- The plan links to an approved decision and names the authority that can approve, pause, alter, or reverse the change.
- In-scope and excluded processes, systems, decisions, data, locations, and groups are explicit.
- Claims about expected outcomes are labelled as commitments, assumptions, or hypotheses and trace to a source.
- Dates in the plan agree with the approval record and operational constraints.
- Each affected group is described through changed tasks, access, decisions, handoffs, or support needs rather than a broad label.
- The current and future state agree with the process maps.
- Consultation, accessibility, privacy, security, contractual, and workforce requirements have named reviewers.
- Groups that absorb exceptions or extra work are included, even if they do not use the changed tool.
- Every action has one accountable owner, due date, prerequisite, completion artifact, and escalation condition.
- Dependencies are ordered and agree with the schedule.
- Cutover, stabilisation, and closure work are distinct from preparation.
- Calculations for capacity, timing, cost, or coverage can be reproduced from retained inputs.
- Each affected group has a message purpose, sender, channel, timing, response route, and accessible alternative.
- Training tests the changed task rather than attendance alone.
- Support ownership and escalation hours are explicit for rollout and stabilisation.
- Communications do not present assumptions or AI-generated suggestions as approved facts.
- Approval gates, stop conditions, service thresholds, and required specialist checks are observable.
- The back-out or fallback procedure names an owner, trigger, sequence, data treatment, and verification step.
- The fallback remains usable during the planned change window.
- Residual uncertainty is visible to the authority instead of buried in narrative text.
- Material AI-assisted summaries, comparisons, calculations, or recommendations identify their source set and review status.
- Unsupported outputs, contradictions, and rejected suggestions are recorded.
- Human decision rights are explicit; the assistant does not assign authority or make employment decisions.
- Sensitive inputs and approved tool boundaries match the organisation's current controls.
- Readiness, operating, and outcome evidence are separated.
- Every measure has a definition, source, owner, collection cadence, comparison point, and response threshold.
- Missing data and adverse signals have a response, not merely a dashboard field.
- The review does not treat activity, such as logins or training attendance, as proof of a service outcome.
- All repairs have owners and due dates, and all accepted exceptions have rationale, authority, expiry, and review date.
- Contradictions are resolved in the plan, not only noted in review comments.
- The final approver receives the unresolved risks and evidence limitations.
- The signed disposition, reviewed plan hash or version, and supporting artifacts are retained together.
Worked artifact: reviewing a supplier-summary rollout
A synthetic operations team plans to introduce AI-assisted first drafts for monthly supplier summaries. The plan limits AI use to an authorised document set and keeps approval with contract owners. The reviewer applies the checklist before a two-week pilot.
| Gate | Finding | Disposition | Required evidence | | --- | --- | --- | --- | | Authority | Scope is approved, but pause authority is unnamed | Repair | Change authority added to plan and runbook | | Affected work | Reviewers are mapped; contract owners receiving exceptions are missing | Repair | Updated handoff map confirmed by both groups | | Dependencies | Training precedes access, but access lead time is absent | Repair | Identity-team lead time and named owner | | Enablement | Training records attendance only | Repair | Synthetic proficiency exercise with review rubric | | Recovery | “Return to manual” has no trigger or data-reconciliation step | Escalate | Threshold, owner, sequence, and reconciliation test | | AI traceability | Source links are required; rejected suggestions are not retained | Repair | Review log includes accepted and rejected outputs | | Monitoring | Login rate is proposed as success | Repair | Separate source-trace errors, corrections, exceptions, and service time | | Approval | Two residual risks have no expiry | Repair | Authority, rationale, expiry, and next review date |
The plan does not pass merely because seven sections exist. It is held until pause authority, fallback execution, and evidence definitions are repaired. The team then rehearses the fallback with synthetic files. The retained record links the reviewed plan version, checklist, rehearsal result, open exceptions, and final authority decision.
Decide what happens next
Approve only when every blocking item passes and accepted exceptions are within the authority's remit. Return the plan for targeted repair when the problem is resolvable without changing the approved decision. Escalate when the scope, risk appetite, legal interpretation, or decision right is unclear. Stop the review if the source packet cannot establish what was approved.
Two Prune can provide assessment evidence about how a participant frames, verifies, and communicates work in a realistic simulation. It does not approve an organisation's change plan, certify control effectiveness, or make employment decisions.
Sources
UK Government, Service management good practice
- Government of Canada, Outcome Management Guide
- NIST, AI Risk Management Framework Core
Sources: [1] [2] [3]
Sources
- 1.UK Government, Service management good practice · UK Government
- 2.Government of Canada, Outcome Management Guide · www.tbs-sct.canada.ca
- 3.NIST, AI Risk Management Framework Core · National Institute of Standards and Technology