Evergreen field guide · updated August 5, 2026
AI change management.
A practical operating guide for turning fast-moving AI capability, policy, identity, data, and workflow change into safe, useful, and measurable adoption.
Plain-language definition
What is AI change management?
AI change management is the discipline of turning a change in AI capability, policy, data access, or workflow into an owned operating transition. It connects the decision, affected people, controls, communication, adoption evidence, and rollback path instead of treating launch as the finish line.
Governance defines the boundary. Change management makes the boundary usable. Evidence proves whether the system and the organization stayed inside it.
The operating difference
AI changes more than the interface.
A normal software release can change screens and steps. An AI release can also change capability, confidence, data use, identity, cost, judgment, and the pace at which the workflow evolves. The review has to follow the whole operating chain.
| Change signal | Operating question | Minimum evidence |
|---|---|---|
| Model or capability | What can the system do now that it could not do before? | Evaluation, access boundary, task test, and rollback owner |
| Agent or connected tool | Whose identity, credential, data, and authority move through the workflow? | Requester, agent, connection, approval, action, and disposition receipt |
| Policy or regulation | Which role, workflow, decision, disclosure, or record must change? | Named owner, effective date, implementation path, and review record |
| Data or memory | What is collected, retrieved, inferred, retained, shared, corrected, or deleted? | Purpose, minimum needed, actual data path, retention, and deletion test |
| Workflow adoption | Can each affected person perform the changed task safely and successfully? | Role readiness, observed task outcome, support signal, and retained use |
The operating framework
Six receipts before scale.
Use the sequence for a new deployment, a material model update, a connected agent, a policy change, or a workflow that is producing different outcomes than intended.
- 01
Detect the operating change
Name the model, policy, data, agent, workflow, or vendor change—and the real decision it creates for people doing the work.
Receipt: A dated signal, affected workflow, and decision deadline. - 02
Name the owner and boundary
Assign the decision owner, approved users, data boundary, tool scope, budget, and the authority required before work moves.
Receipt: A named owner and explicit allowed, denied, and approval-required actions. - 03
Redesign the workflow
Map where AI enters, what remains human-led, which handoffs change, and what happens when confidence or context is insufficient.
Receipt: A before-and-after workflow with escalation and exception paths. - 04
Prepare people to operate it
Give each role the reason, task guidance, practice, support channel, and feedback loop needed to use the change safely and well.
Receipt: Role-based guidance, practice results, support ownership, and readiness gaps. - 05
Prove adoption and control
Measure retained use and task outcomes alongside overrides, denied actions, incidents, cost, and the quality of evidence produced.
Receipt: A review that joins adoption, outcome, risk, and operating-health measures. - 06
Keep rollback real
Test how access is revoked, credentials are rotated, work is corrected, data is deleted, and the prior safe process is restored.
Receipt: A witnessed rollback or revocation test with a final disposition.
Practitioner field tools
Move from the framework to the working table.
Use these source-backed artifacts with one real workflow. Each resource makes the decision boundary, operating evidence, and next disposition visible—and can be printed or copied for a working review.
Public-sector operating playbook
Six service gates, accountable roles, and a bounded 90-day path.
Open the playbook →SCORECARDAI change-management metrics
Thirteen formulas joining adoption, outcomes, control, and rollback.
Use the scorecard →CHECKLISTSix receipts before scale
Thirty checks and one explicit scale, revise, pause, or retire decision.
Run the review →RESEARCH2026 signal report
Original analysis of 83 source-linked AI operating signals.
Read the report →Measurement
Adoption and control belong on one scorecard.
A high login count can coexist with poor outcomes, shadow use, weak review, or untested rollback. Pair behavior and value with operating-health evidence.
- 01Reach
- Eligible people who received the right access, guidance, and practice.
- 02Activation
- People who complete the first intended task—not merely open the tool.
- 03Retained use
- People who return when the workflow calls for it and use the approved path.
- 04Task outcome
- Quality, time, completion, error, or service improvement tied to real work.
- 05Human control
- Overrides, escalations, denied actions, approvals, and exception patterns.
- 06Operating health
- Incidents, cost, drift, support demand, correction time, and rollback readiness.
Run it this week
A one-hour AI change review.
Choose one live workflow. Keep the scope small enough to inspect the actual identities, data, tasks, controls, and outcomes—not a strategic diagram of the entire platform.
Signal + decision
Name what changed, the affected workflow, the deadline, and the accountable decision owner.
Identity + data
Map users, credentials, purpose, effective permissions, data movement, retention, and disclosure.
Workflow + people
Define human judgment, changed tasks, role guidance, practice, support, and escalation.
Evidence + rollback
Agree on outcome and control measures, run one boundary test, and prove how the path stops.
Applied record
Follow the framework through the archive.
AI Change Desk develops this operating model in public through source-linked episodes, complete notes, transcripts, practical checks, and explicit production disclosures.
Welcome to AI Change Desk
Introduces the 4D Desk Memo: Decision, Data, Drift, and Deployment.
Open episode file →EP003AI Governance Implementation
Moves AI policy from kickoff language into a repeatable operating loop.
Open episode file →EP037Work Agent Receipt Check
Defines the evidence needed when an AI agent says the work is complete.
Open episode file →EP038The Receipt Is the Trajectory
Connects evaluations to the real tools, identities, data, and networks they may touch.
Open episode file →EP039Whose Account Did the Agent Use?
Separates audience permission, credential capability, purpose authority, and action approval.
Open episode file →Frequently asked questions
The short answers.
What is AI change management?
AI change management is the discipline of turning a change in AI capability, policy, data access, or workflow into an owned operating transition. It connects the decision, affected people, controls, communication, adoption evidence, and rollback path.
How is AI change management different from AI governance?
AI governance defines decision rights, principles, risk boundaries, and accountability. AI change management makes those choices usable in real work through workflow design, role preparation, communication, support, measurement, and continuous review. Strong programs operate them together.
Who should own AI change management?
One accountable business owner should own the outcome, with technology, privacy, security, legal, data, communications, learning, and frontline representatives contributing where the workflow touches their responsibilities. A committee can advise; it cannot replace a named owner.
Which AI adoption metrics matter most?
Measure eligible users, activation, retained use, task success, time or quality improvement, human overrides, escalations, denied actions, incidents, cost, and time to correct or roll back. Usage alone cannot prove that the change is useful or controlled.
When should an AI change review be repeated?
Repeat the review when the model, system instructions, connected data, tools, user population, policy, vendor terms, risk classification, cost model, or operating owner changes—and whenever evidence shows the workflow is behaving differently than intended.
Method + sources
Use the guide as an operating layer.
The framework is Michael's practitioner synthesis. It is designed to complement—not replace—your organization's legal, privacy, security, records, procurement, labor, accessibility, and risk requirements.