Bringing people along: change management for AI

Does AI really need change management?

Yes. AI changes how a team works, and that goes well beyond opening a new tool. A speaker quoted in the IHK Karlsruhe magazine (2026) put it directly: "Wer nur die Technologie implementiert, aber die Belegschaft nicht mitnimmt, wird scheitern", whoever implements only the technology without bringing the workforce along will fail. Bitkom's 2026 survey (14 September 2026) puts a number on it: 42% of companies that use AI name a lack of employee acceptance as a barrier.

Low acceptance rarely means people reject AI as such. Many already use it on their own. According to IW Köln (19 September 2026), 29.1% of employees who use AI at work use tools their employer did not provide, and the institute warns: "Wer fördert, ohne bereitzustellen, macht die Nutzung unkontrollierbar", encouraging use without providing tools makes use uncontrollable. What people resist is a system that may be wrong in front of customers, that may be used to rate them, or that may cost them their job. Each of those fears has an answer in how the system is built and introduced, and the sections below go through them.

Will AI replace jobs at our company?

It can, and your staff know that companies expect it. In Bitkom's 2026 survey, 45% of companies expect a slight decline in staff numbers because of AI and 12% a significant one. Two thirds of German CFOs expect substantial job cuts in finance (Deloitte CFO Survey, 2 July 2026). In IT, 32% of companies expect AI to cut jobs, and 61% expect routine tasks to disappear (Bitkom, 3 September 2026). Telling people that nothing will change is not credible, and they will notice.

What you can do is answer the question for your own company, and do it before the rollout. Decide what happens to the hours the system frees: the backlog it clears, growth handled without new hiring, posts left open when people retire, new tasks, or fewer posts. Then say it, in writing where you can. A works agreement can record commitments on how saved time is used. Where AI changes work significantly, further co-determination rules can apply, and Grant Thornton (2026) names section 111 of the Works Constitution Act among them. Those are decisions for management with HR and the works council, advised by your lawyers. This guide is not legal advice.

Do not promise what the company has not decided. A sentence such as "nobody will lose their job because of this system" is only worth saying if management will sign it.

How do we handle staff fears about AI?

Answer each fear with a property of the system that people can check for themselves. Employees name errors and inaccuracies in AI-generated content and data protection as their main concerns (McKinsey HR-Monitor 2026). Other surveys show that staff want to stay in charge. In Inform's 2025 survey of logistics and supply chain companies, reported by transport-online.de on 5 November 2025, 46% preferred a model in which AI suggests and a person makes the final decision, 81% expected explanations and 61% wanted to be able to intervene.

Is it enough to inform the works council before rollout?

No. Bring it in while you are still planning. Under section 90 of the Works Constitution Act (BetrVG) the employer has to inform the works council in good time when work procedures are planned, and section 87(1) no. 6 gives it co-determination as soon as a system is objectively capable of monitoring behaviour or performance, which most AI logs are. Under section 80(3), an expert is deemed necessary when the council has to assess AI. A council that sees the design early asks its questions while they are cheap to answer. One that is shown a finished system can delay it until there is an agreement.

Bitkom's guide on AI and co-determination (early 2026) suggests what a works agreement can contain: who may access which data and how access is logged, a traffic-light process for approving new AI applications, a short profile of each system (KI-Steckbrief), test phases and rules for human involvement in decisions. Each of those is something the system can show. Grant Thornton (2026) adds a hard limit: emotion recognition at work is prohibited under the AI Act. Our guide on works council logging lists the fields an assistant, a gateway and an agent record and how to configure them so the agreement can be kept.

How do we introduce AI step by step?

Let the team see what the system would do before it does anything. In shadow mode the agent, a program that carries out steps in your systems and presents results for approval, receives the same live cases as the team and proposes what it would do, while every write is recorded instead of executed and people work as usual. The team then sees, on its own cases, where the agent is right and where it is not. Each later stage gives the system a little more room, by rule, and only after the team has seen the evidence.

1. 1

   Let the team define what correct means

   The people who do the work choose and label the test cases, including the awkward ones. That way the acceptance test is theirs.

2. 2

   Run in shadow mode

   The agent proposes on live cases and changes nothing. Its proposals are compared case by case with what people decided.

3. 3

   Show the team the comparison

   Where it was right, where it was wrong and why. Corrections from the team go into the test set and into the next version.

4. 4

   Prepare, and let people approve

   The agent prepares each case and a person approves every action before it happens, seeing the sources and the reasons.

5. 5

   Widen by rule, keep the brake

   Actions that met the agreed threshold may run with sampled review. Everything else keeps its approval step, and the team can stop the agent at any time.

Our guides on agent shadow mode and on human approval for agents describe how to make shadow mode safe and which actions need sign-off. The same sequence works for a knowledge assistant: start with a group of colleagues who check the answers and report wrong ones, then open it to the rest of the team.

What should we measure and show the team?

Measure the work, not the people, and show the team the same numbers management sees. Before go-live, record how the process runs today: cases handled per week, hours per case, how often cases come back for correction. After go-live, measure the same things at team level, plus how often people accept, edit or reject the system's proposals. Show the figures to the team regularly, including the cases the system got wrong, and say what was changed because of their feedback.

• Give people one simple way to flag a wrong answer or a bad proposal from inside the tool, and close the loop on each report.
• Report by team and by process, never by person. Put the same rule into the logging configuration, so a new dashboard cannot bypass it.
• Share the results with the works council in the same form, so both sides argue from the same facts.
• Agree up front what result would make you stop or change course, and say so when it happens.

How do we train employees on AI?

Train people on the team's real work, with the tool they will use day to day. The need is clear: in Bitkom's 2026 survey, two thirds of companies rated their employees' AI skills as low and 25% offered no training at all. Only 28% of German employees say their employer offers formal AI training (McKinsey HR-Monitor 2026). Meanwhile use keeps rising: the same McKinsey source reports that 38% of German employees use generative AI regularly, up from 19%.

• Use the team's own cases: the email types, documents and requests it handles every day, including the ones the system gets wrong.
• Teach checking as well as prompting: how to read a source, when to distrust an answer, which data must not go in, and how to stop an agent.
• Name one colleague per team who knows the system well and collects questions, so help is close to the work.
• Keep a simple record of who learned what about which tool: it is also the evidence for the AI literacy measures the AI Act asks for.

A training plan does not need to be long. Group people by what they do with AI: everyone who uses an assistant, the people who review or approve its results, and the few who run and change the systems. For each group, write down what they learn before they get access, on which of their own cases, and what triggers a refresher, such as a new model, a new tool or a changed process. That plan, kept next to the record per person, is what management, the works council and an auditor can check.

When we build a system with a client, the team learns it while it is built, on its own cases, and receives the runbooks and test set with it. What the AI Act expects in terms of literacy measures is covered in our EU AI Act checklist. To talk through a rollout in your company, talk to an engineer: a free 30-minute call that needs no access to your systems or data.

AI agents and process automationWorks councilsBusiness unit and operations leads