Two Things Change Management Never Had to Handle
Last week I argued that the change management frameworks built by Kotter, Bridges, and Prosci still hold for AI initiatives. The mechanics they identified have not changed. The pace has, and the manager's job is to run those mechanics faster than the frameworks were designed to move.
But there are two things the old frameworks did not anticipate. Both fall on the manager. Both require a kind of attention that was not in the original playbook.
The first is that AI fails in a way old quality controls were never built to catch.
In the old world, quality was a product of iterative process improvement. Automated testing provided large coverage. Manual testing caught what the automation could not like a missed edge case, a typo, a logic bug that someone could trace back to a specific assumption. Quality control was a check against human limitations, and the failure mode was known.
AI produces a different failure mode. It does not just make mistakes. It produces confident hallucinations. Output that reads as authoritative, with internal consistency, citations, and the cadence of someone who knows what they are talking about. The failure mode looks like a success.
Managers have always needed to evaluate their team and their team’s product. But AI has narrowed the gap between good and bad output so much that you can no longer just review the output. You have to review the judgment of the person who used the tool. Did they spot the places where the model was likely to drift? Did they verify the citations? Did they recognize when the answer was too clean? That kind of review takes more context, not less. The more autonomous the tool, the more expert the human oversight needs to be.
This is a real shift in our working lives with real implications for failure. The team members most at risk of producing confident hallucinations are the ones with the least experience to spot them. A junior engineer using an AI coding assistant is generating output at a senior pace, but without the senior context to know when it is wrong. The work moves faster. The review burden moves up.
The second is that the vendor relationship is no longer stable.
The frameworks of the 90s assumed a stable vendor. You bought a license, the price was the price, and you could count on your budget this year and a small increase next. The tools were a given.
AI vendors are refactoring their pricing in real time. Unlimited plans disappear overnight. Seat-based pricing gets replaced by credit-based pricing. Products are being moved in and out of your license. The cost per prompt for a workflow you depend on can be different based on when you run it during the day.
For a manager, this means reinforcement now includes governance. Reinforcement, as Prosci and Kotter framed it, was about behavior under the assumption that the tools were a given. That assumption is no longer safe. You are not just managing how your team works. You are managing a volatile supply chain that runs underneath their work. If you are not reviewing your vendor's terms regularly, you are setting your team up for a work stoppage when the budget hits a cap nobody saw coming.
What the shift looks like in practice
Old world to AI world, where the manager's attention has to go:
Urgency was "the three-year plan is failing." Now it is "the tool changed this morning."
Short-term wins used to be quarterly milestones. Now they are Friday afternoon workflow demos.
Quality control used to mean reviewing for human error. Now it means reviewing for confident hallucinations produced at machine speed.
The risk profile used to be slow erosion of market share. Now it is one bad prompt that leaks PII.
The vendor relationship used to be stable, seat-based, and predictable. Now it is volatile, credit-based, and moving.
These are not edits to the change management playbook. They are additions, and they sit on top of the work the frameworks already asked of you.
Four things to try this week
If you want to move from managing the chaos to leading the initiative, these four moves are a starting point. None of them require organizational permission.
First, name your high-governance zones. List the three places where a human in the loop is non-negotiable, like production database access, PII handling, and final client deliverables. Reducing the risk area gives the rest of the team more room to experiment without losing sleep over it. What is irreversible?
Second, run a weekly credit check. Have one person on your team check your AI usage and credits every Tuesday. They are watching for changes on the vendor's side, not on yours: a per-prompt cost that shifted, a plan structure that got refactored, a product that quietly moved out of your license. If any of that changed, you want to know before the budget hits a cap nobody saw coming.
Third, build AI usage habits as a team. Who are the power users that are still not hitting their usage limits? What are they doing differently? Isolate the best habits and share them. Train people to avoid the costly ones. Something as simple as using a deep-thinking model for a basic task or letting a context window run too long can increase costs unnecessarily. High cost without a visible win is a signal that you need to evaluate.
Fourth, run a failure demo. In your next team meeting, instead of asking who used AI successfully, ask who caught an interesting failure. Celebrate the person who spotted the hallucination. Skepticism and judgment are the new quality controls, and the team needs to see you value them.
The frameworks still work. The job is still recognizable. The pace is faster, the failure modes are new, and the vendor relationship is part of your job now. None of that changes the fact that you are leading people through something hard. That part has not changed at all.