AI Change Management, KPIs and Operating Rhythm

In July 2024, McKinsey asked 1,491 people in 101 countries how their organizations were adopting generative AI, and then tested which management practices went with the earnings impact respondents reported. One practice stood out. Of the twelve adoption and scaling practices surveyed, tracking well-defined KPIs for generative AI solutions had the most impact on the bottom line. Fewer than one in five respondents said their organizations did it¹.

That gap is less surprising than it looks. Measuring an AI tool means agreeing what it was supposed to change, which means agreeing how the work will be done differently, by whom, and who will look at the result and act. Many organizations skip those agreements and go straight to the launch. The measures are missing because the change was never defined.

That is the problem this chapter addresses. The previous chapter settled who has the authority to decide. This one is about what happens after the decision: getting people to work differently, knowing whether they are, and running the meetings that act on what you learn.

The idea: three disciplines, one loop

AI change management is the deliberate work of helping people adopt, and keep using, a new way of working that AI makes possible. It sits inside a loop with two other disciplines. Change makes the new work possible: it shapes the tool, the workflow, the role and the manager’s behavior. Measures show whether the work has changed and whether results have moved. Rhythm is the fixed sequence of meetings in which those measures are read and decisions are taken.

The three fail in characteristic ways when one is missing. Without change, you get a tool people open in a workflow that has not moved. Without measures, you cannot tell adoption theater, which Measuring AI Business Value described, from real progress. Without rhythm, even good measures sit in a dashboard that nobody is obliged to act on. The underlying distinction between use and impact belongs to AI Strategy vs AI Adoption; this chapter is about the machinery that closes the gap.

Change is a redesign of the work, not a training plan

A common mistake is to equate change management with training and communication. Both matter. Neither changes the work. The AI Adoption Curve showed the result in Danish data: employer encouragement, tools and training nearly doubled take-up of AI chatbots, yet two years on there was no measurable effect on earnings or hours². The same chapter cited McKinsey’s finding that, of 25 organizational attributes tested, redesigning workflows had the biggest effect on bottom-line impact, while only about a fifth of organizations had done it. That is a separate ranking from the twelve adoption practices in which KPI tracking came first¹. Adoption can be bought. Value has to be built into the work.

So the change plan for any AI initiative starts from the work, not from the course catalog. Who does this work today? What will they do differently on a Tuesday morning? What step disappears, and who decides that it may? What will they stop being measured on? John Kotter called this “empowering broad-based action”: removing the structures, systems and appraisals that quietly punish people for doing the new thing³. An AI tool that saves a planner an hour, while the planner is still appraised on the old manual checks, will be used for an hour and then dropped.

Managers carry the change

Executives announce a change; managers decide whether it happens. Prosci’s long-running benchmarking finds that employees want to hear why the organization is changing from the top, but what it means for me from their immediate supervisor⁴. For AI, the second message carries most of the weight, because the worries are personal: Will this be used to judge me? Who is accountable when it is wrong? Is my job safe?

BCG’s 2025 survey of more than 10,600 employees found that regular use among frontline staff had stalled at 51 percent. Where leadership support was strong, the share of employees who felt positive about generative AI rose from 15 percent to 55 percent, yet only about a quarter of frontline employees said they had that support. Regular use was sharply higher among people with at least five hours of training, and only about a third said they had been properly trained⁵. BCG’s 2026 edition, discussed in The AI Talent Gap, found the communication problem persisting: only a third of frontline employees said leadership’s messages about AI were clear⁶. These are self-reported associations, not experiments, but they match long practitioner experience. Put managers into the change first, give them the answers to the personal questions, and give their teams real time to learn on real work.

Find the leak before you fix it

Adoption is a sequence, not a switch, and AI Strategy vs AI Adoption traced how reach narrows into regular use. What matters for execution is that each stage fails for a different reason, so each needs a different remedy. More workshops cannot fix a leak caused by a workflow that does not fit.

The diagnostic question is simple: where, exactly, do people stop? If 90 percent have been trained but only a third use the tool weekly, the leak is after training, and the remedy lies in the workflow, the quality of the output or the manager’s behavior. If use is high but the old manual steps are still being done in parallel, the leak is at embedding, and the fix is to remove the old steps formally and change what people are measured on. If the process has genuinely changed and results have not, no change effort will help; the use case was wrong, and the stop rules in Prioritizing the AI Portfolio apply.

Change capacity is the real limit

Every roadmap in this module has run into the same constraint: the organization can only absorb so much change at once. Gartner researchers reported that the average employee went through 10 planned enterprise changes in 2022, up from two in 2016⁷. AI arrives on top of restructurings, system replacements and culture programs, often for the same teams.

The practical response is to manage change load as you would manage engineering capacity. Someone should be able to say, for each frontline team, how many changes it will absorb this quarter, AI and otherwise. When that number is too high, the right decision is to sequence: delay a launch, or combine two changes to the same workflow into one. The 90-day plan in Building the 90-Day AI Plan is the place to make that trade-off explicit, and the transformation board from the previous chapter is the place to settle it when two sponsors want the same team in the same quarter.

A scorecard that triggers decisions

The opening of this chapter showed that measurement is where many AI programs are weak, and where the survey linked it most strongly to results. Baselines and leading measures have their own chapters, Baselines, Metrics and Measurement and Leading vs Lagging AI Metrics. What this chapter adds is the executive scorecard: a short list in four layers, each read at a particular meeting.

Two disciplines make the scorecard useful. First, every measure has an owner, a threshold and an agreed action if it crosses the threshold, decided before the data arrives, as Baselines, Metrics and Measurement recommends. A measure with no action attached is a report. Second, the scorecard stays short. Ten measures that each trigger a decision beat a hundred that trigger discussion. Teams optimize what is counted. Count prompts and you will get prompts; count the share of work done the new way, and the result it produces, and you have a chance at value.

The operating rhythm

An operating rhythm is the fixed calendar of meetings in which the organization reviews evidence and decides. Robert Kaplan and David Norton, who spent decades studying how companies execute strategy, argued that operational and strategic reviews need separate meetings, with different frequencies, attendees and agendas. Where they are combined, operational crises tend to crowd out strategy⁸. AI programs suffer from exactly this: the steering meeting spends its hour on a delayed integration and never reaches the question of whether the portfolio is right.

This is the single cadence for the whole course. The weekly review stays close to delivery and adoption, and it is where the escalation clock from the previous chapter starts. The monthly review reads the scorecard and adjusts. The quarterly review is where money moves: it closes one 90-day plan and approves the next. The annual test asks whether the assumptions under the multi-year roadmap still hold, the scheduled rebuild described in Building the Multi-Year AI Roadmap. The financial cost reviews in Cost Optimization and AI FinOps feed into the monthly and quarterly meetings; they do not need a calendar of their own.

Every review should end with the same four questions: what changed, what did we learn, what must be decided, and who decides. Kotter warned that transformations fail when leaders declare victory too soon, before the new way of working is anchored³. A rhythm guards against that. Once the launch is over, the meetings keep asking whether the work has changed, until the new way is simply how the work is done.

Consult before you change the work

AI that changes how people work is not only a management question. In Europe, it is also a legal one, and the duties fall on the timeline of the change plan, not after it. The AI Act sets the EU-wide floor, amended in 2026 by the Digital Omnibus [@eu-ai-act-oj-2024; @eu-ai-omnibus-oj-2026].

National rules often bite first. In Germany, the Federal Labour Court reads the works council’s co-determination right as covering any technical system that is objectively suitable for monitoring behavior or performance, whatever the employer intends, which brings many AI tools that log who did what within scope⁹. The practical consequence for the change plan is straightforward: consultation goes on the 90-day plan as a long-pole task, its dates go into the operating rhythm, and the literacy measures are designed by role, as AI and the Future of Work recommends, rather than as one course for everyone.

Story: the housekeeping board that would not go away

This is an illustrative composite, not a real company. It compares two regions of one hotel group that introduced the same AI tool in the same year.

The group operated about forty hotels in Europe. Its new tool forecast checkouts and arrivals and assigned rooms to room attendants each morning, so that the rooms guests needed first were ready first. The measure that mattered was the share of rooms ready by two in the afternoon.

The German region moved fast. It launched in fourteen hotels in six weeks, with an online course that 94 percent of staff completed. Its dashboard showed training completion and logins, and both looked healthy. Twelve weeks later, the picture underneath was different. Housekeeping supervisors opened the tool at the morning briefing, then assigned rooms on the paper board as they always had; fewer than a third of assignments went through the tool. Rooms ready by two had not moved. Then the works councils in two hotels objected that the tool recorded how long each attendant took per room, and no agreement covered it. The rollout stopped.

The Benelux region started with three hotels. Before the pilot, it agreed with its works council that room times would be reported only at team level and never used in individual appraisal. Supervisors redesigned the morning briefing with the project team, and in week four they retired the paper board themselves. The weekly review tracked the share of assignments made through the tool and the supervisors’ reasons for overriding it; two of those reasons led to changes in the tool. The monthly review compared rooms ready by two against the pre-launch baseline and three similar hotels without the tool. By week ten, eight in ten assignments went through the tool, and rooms ready by two had risen from 71 to 83 percent, while the comparison hotels were flat. At the quarterly review, the board extended the tool to nine more hotels.

The technology was identical. The German region did more training than the Benelux region and reported better activity figures throughout. What it lacked was all three disciplines: a change that removed the paper board, measures that looked past logins to the work and the result, and a rhythm that would have caught the leak in week three rather than week twelve. The group’s chief operating officer did not cancel the German rollout. She restarted it the Benelux way, with consultation first.

What this means for leaders

Transformation is complete when the work has changed and the results have moved, not when the system is live. That shifts an executive’s attention from launches to three questions asked every month: is the work being done the new way, where are people dropping out, and what did the last review decide?

It also changes what you fund. A change budget that buys only courses and communications buys adoption at best. The money that buys value goes to redesigning the workflow, freeing managers to lead it, removing the old steps and measures, and consulting the people affected early enough that their concerns shape the design.

Check yourself

1. If people complete the training, adoption will follow.MythFact
2. The practice most linked to bottom-line impact from generative AI is also one of the most widely used.MythFact
3. Employees prefer to hear what a change means for them from the CEO.MythFact
4. Every frontline team should be able to say how many changes it is absorbing this quarter.MythFact
5. Strategy and operational issues are best handled in the same monthly meeting.MythFact
6. Consulting a works council can wait until the pilot has shown results.MythFact

Reflection: where would you find the paper board?

What comes next

Each part of execution now has an owner, a measure and a meeting. What remains is how they fit together. The final chapter, Module 09 Synthesis — From Strategy to Execution, puts them together into one system and shows how the parts depend on each other.