Practical guide
How to create a change plan with AI
A change plan turns an approved change into an accountable transition. It should explain what will change, who is affected, what must happen before the change, how people will be supported, which risks stop or alter the rollout, and what evidence will show whether the new way of working has become normal practice. AI can help organise source material, surface missing dependencies, compare.
A change plan turns an approved change into an accountable transition. It should explain what will change, who is affected, what must happen before the change, how people will be supported, which risks stop or alter the rollout, and what evidence will show whether the new way of working has become normal practice. AI can help organise source material, surface missing dependencies, compare stakeholder needs, draft communications, and test scenarios. It should not invent approval, assign authority, decide employment consequences, or convert uncertain outcomes into commitments.
Start with a decision packet
Collect the records that establish the change rather than asking an AI assistant to infer them from a short prompt:
The UK Government's service-management guidance describes a service change as a planned event with a change authority, change owner, change window, and back-out plan. The Government of Canada's Outcome Management Guide connects sustained outcomes with timely communications, preparation for new activities, monitoring, and reinforcement. These sources support a plan that joins operational control with adoption work.
- approved business case or decision record;
- current and proposed process maps;
- affected services, roles, systems, suppliers, and policies;
- named change authority, implementation owner, and specialist reviewers;
- timing constraints, dependencies, and operational blackout periods;
- workforce consultation, accessibility, security, privacy, and contractual requirements;
- current performance evidence and the intended outcome; and
- known objections, incidents, and previous transition lessons.
Build a CHANGE record
Use one CHANGE record as the spine of the plan.
### Charter
State the approved change, reason, boundaries, intended outcome, decision owner, and effective period. Separate facts in the approval record from assumptions that still need testing. Name what is explicitly out of scope.
### Humans
Map groups by how their work, access, decisions, or support needs will change. For each group, record the current task, future task, likely friction, required consultation, training need, and feedback route. Avoid treating a broad label such as “all users” as a useful segment.
### Actions
Sequence the smallest observable actions required before, during, and after the change. Every action needs an owner, due date, dependency, completion evidence, and escalation condition. Distinguish preparation, enablement, cutover, stabilisation, and closure.
### Navigation
Create a communication and support route for each affected group. Specify the sender, purpose, channel, timing, response path, and accessible alternative. A broadcast message is not evidence that a person can perform a changed task.
### Guardrails
Record approval gates, stop conditions, security and privacy checks, service thresholds, fallback steps, and the authority to pause or reverse. Link each guardrail to a source requirement or accountable reviewer.
### Evidence
Define leading, operating, and outcome evidence. Leading evidence may include access, training completion, or readiness checks. Operating evidence may include correct use of the new process, exceptions, and support demand. Outcome evidence relates to the intended service result. Do not treat one layer as proof of another.
Use AI at bounded steps
AI assistance is most useful when the source packet and expected output are explicit.
| Step | Suitable AI assistance | Human verification | | --- | --- | --- | | Source inventory | Extract named decisions, dates, owners, and dependencies | Trace every material field to the original record | | Stakeholder analysis | Suggest affected tasks and questions to investigate | Confirm impact with process owners and affected groups | | Scenario testing | Generate failure cases and dependency conflicts | Select credible cases and define stop authority | | Communication drafting | Produce audience-specific first drafts | Check accuracy, accessibility, tone, and approved claims | | Plan review | Find missing owners, dates, evidence, or fallback steps | Resolve gaps and obtain accountable approval |
Retain prompts and source versions when they materially shaped the plan. Mark generated suggestions that were rejected so they do not reappear as settled decisions in later drafts.
Worked example: change an AI-assisted review workflow
Consider a synthetic operations team introducing an AI-assisted first draft for monthly supplier reviews. The approved scope covers summary drafting from an authorised document set. It does not allow the assistant to approve suppliers, alter contract records, or use unrestricted data.
The first draft plan says: “Train reviewers, launch on 1 June, and monitor adoption.” The CHANGE review exposes missing decisions:
| Record | Evidence found | Repair | | --- | --- | --- | | Charter | Approval covers summaries only | Add excluded decisions and authorised source set | | Humans | Reviewers named, contract owners omitted | Add contract-owner review at exception points | | Actions | Training has no proficiency check | Add a synthetic review exercise and feedback owner | | Navigation | One launch email for all groups | Separate reviewer instructions from owner escalation guidance | | Guardrails | No response to unsupported claims | Require source trace, pause on missing evidence, retain manual workflow | | Evidence | Login count proposed as success | Track traceable claims, corrections, exceptions, and service timing separately |
The revised plan runs a two-week pilot with eight reviewers using synthetic supplier files. A reviewer must attach a source trace and flag unresolved contradictions. If more than two material claims in a review lack support, the team returns to the manual workflow and the change owner investigates. Contract owners review only escalated exceptions. At the end, the authority decides whether to expand, repair, repeat, or stop based on the retained evidence.
This example does not assume the pilot improves quality or speed. It shows how a plan makes the decision and evidence route inspectable.
Test the plan before approval
Ask a reviewer who did not write the plan to run five tests:
Record pass, repair, unresolved, or not applicable for each test. Approval should include the selected rollout, accepted exceptions, residual uncertainty, review date, and named authority.
Two Prune can assess how a participant frames a change, navigates source material, uses AI, verifies claims, weighs risks, and communicates an executable plan in a bounded simulation. It does not approve a real organisational change or establish that the change produced an operational outcome.
- Traceability: Can each scope, owner, constraint, and claim be linked to an authoritative source?
- Executability: Can every action be performed with the stated access, time, skill, and dependency?
- Impact: Are affected tasks and groups specific enough to support consultation and enablement?
- Recovery: Are pause, escalation, and back-out conditions observable and owned?
- Learning: Can monitoring distinguish preparation, correct operation, and the intended outcome?
Sources
Government of Canada, Outcome Management Guide
- UK Government, service management good practice
- NIST AI Risk Management Framework Core
Sources: [1] [2] [3]
Sources
- 1.Government of Canada, Outcome Management Guide · www.tbs-sct.canada.ca
- 2.UK Government, service management good practice · UK Government
- 3.NIST AI Risk Management Framework Core · National Institute of Standards and Technology