Why Most AI Pilots Never Reach Daily Work in Germany

Many German companies have tried an AI tool, and many of those trials go nowhere. A study by Sopra Steria with Lünendonk, published on 2 November 2026, asked 180 IT and department managers in Germany, Austria and Switzerland about agentic AI. Of the 141 companies that answered the scaling question, 66 percent said fewer than one in four of their AI pilot projects had reached regular operation. Only 1 percent said AI agents were fully integrated in more than one core process.

That gap is the real story. The problem is usually not that the model is weak. It is that nobody fixed the process around it, nobody owned the result, and nobody checked the data it was allowed to use.

Start with one process, not one tool

A pilot that tests a tool in general will end in general praise and no decision. A pilot that tests one process has a clear before and after. Pick a task that repeats every week, has a clear owner, and has a result you can count. Examples are answering routine customer questions, sorting incoming requests, or preparing a standard offer from existing notes.

Write down the steps before you automate them. If the process is unclear on paper, an agent will only repeat the confusion faster. Our workflow automation service starts with that process map, not with a software choice.

Governance is what lets a pilot grow

The same study found a result that surprises many managers. Internal rules, data quality and compliance helped AI agents scale, rather than holding them back. That makes sense in daily work. An agent needs to know which data it may use, what it may never send to a customer, and when a person must check the result.

Keep the rules short and written down:

• Which systems and documents the agent may read
• Which actions need a human approval
• Who is responsible when an answer is wrong
• How often results are reviewed in the first month

A two page rule sheet is enough for most small companies. It also makes it easier to explain the tool to staff and customers.

Fix the data before you add more tools

Agents fail quietly when the data behind them is old, duplicated or stored in private inboxes. Before a second pilot, check the three sources the first pilot actually needs. Are customer details current? Are prices and opening hours in one place? Are standard answers approved, or does every employee use a different version?

If the answer is no, fix that source first. Clean data is not a side project. It is the part of the work that decides whether the agent gives a useful answer on a normal Tuesday.

Measure hours, errors and handover

A pilot needs three numbers, measured for at least four weeks:

• Time saved per task, compared with the old way
• Error rate, counted from real cases, not from a demo
• Handover rate, meaning how often a person has to step in

If time saved is high but errors rise, the pilot is not ready. If nobody has to step in but the task is rare, the pilot is not worth scaling yet. Choose the next process only after the first one passes all three checks. Our earlier guide to AI agents that answer customers and book appointments shows the questions to ask before buying such a tool.

A simple rollout plan

Use this order for the next pilot:

1. Name one process and one owner.
2. Write the steps, the rules and the approval points.
3. Run the pilot beside the old process for two weeks.
4. Compare time, errors and handover, then decide.
5. Only then move the task fully to the new routine.

This order is slower in week one and faster in month three, because the team does not have to rebuild trust after a failed launch.

The takeaway

AI pilots stall when they stay experiments. They scale when one process is clear, the data is clean, the rules are written, and a person stays responsible for the result. Start small, measure honestly, and let the second pilot earn its place.

If you want help turning one pilot into a routine your team trusts, talk to Küstenkurs about workflow automation and digital systems.