In short: Most companies have no trouble launching AI pilots. Their problem is that hardly any of them make it into daily operations and deliver measurable results. That state has a name: pilot purgatory. It rarely comes from bad technology. It comes from the fact that pilots hurt no one, while redesigning a workflow touches responsibilities, budgets and roles. Getting out of purgatory takes fewer projects and more decisions.
How big is the gap between investment and impact?
Few management topics show such a wide gap between ambition and results right now. For its report “Superagency in the Workplace,” McKinsey surveyed 3,613 employees and 238 senior executives. 92 percent of companies planned to increase their investment in generative AI over the next three years. Only 1 percent of leaders described their company as mature in its use. IBM’s 2025 CEO study paints a similar picture: only a quarter of AI initiatives delivered the expected return, and only 16 percent were rolled out across the enterprise.
Germany shows the same pattern at a lower level. According to Bitkom, 36 percent of companies with at least 20 employees use AI. The German Economic Institute (IW) finds 37 percent among 1,038 companies surveyed and notes that use is mostly limited to individual areas. The Federal Network Agency puts the share at around 30 percent among 808 companies, most of them still in a trial phase. Lots of experimenting, very little redesign.
Why do so many AI pilots get stuck?
I see three reasons, and none of them is about technology.
First, activity feels like progress. Fourteen pilots look good on a slide for the board. Cutting that number means explaining why less should be more, and in many leadership cultures that is harder than the opposite.
Second, pilots hurt no one. As long as an agent runs as a pilot, nobody has to change their role, no approval limit gets touched and no department head loses responsibility. The uncomfortable questions only come up when a workflow is redesigned for real.
Third, it’s about where initiatives come from. AI ideas are often collected bottom-up and implemented by IT. That’s a good way to find ideas. It isn’t enough to change workflows that run across several departments, because none of those departments has the authority to reorganize the whole workflow.
What separates a pilot from a redesigned workflow?
| Feature | Pilot | Redesigned workflow |
|---|---|---|
| Ownership | project team or IT | workflow owner with a mandate |
| Baseline | usually not measured | measured before launch |
| Goal | prove feasibility | outcome in one sentence with a number |
| Roles and approvals | unchanged | reassigned |
| End point | open | fixed date to scale, change or stop |
The table shows why moving into operations is a leadership job. Every line in the right-hand column requires a decision only senior management can make.
How do AI pilots get out of pilot purgatory?
In the AI-first operating model, the way out starts with an honest inventory: which initiatives are running, who owns them, what they have delivered, and which have been stuck in pilot mode for months? That review takes one to two weeks and is usually sobering. That’s a good thing.
Then comes the harder part. The leadership team picks three to five workflows and redesigns them from the outcome backward. Each workflow gets an accountable owner, a goal with a number and a baseline measured in advance. Every other idea goes on a list for later. Cutting fourteen AI pilots down to four costs you a nice slide and gives you a real shot at impact.
What should leadership teams review now?
- One question: Which of our AI pilots have been running for more than six months, and who would have to decide to move them into operations?
- One metric: the share of AI initiatives running in daily operations with a measured baseline.
- One conversation: within the leadership team about which projects to stop so the rest get enough attention.
Frequently asked questions
What is pilot purgatory?
A state in which many pilots get launched but hardly any move into daily operations. New experiments keep starting because starting is easy and scaling is hard.
How many AI initiatives should a company pursue at once?
For a genuine redesign, I recommend starting with three to five workflows. The World Economic Forum names the same range. More initiatives spread attention and budget so thin that none of them makes the jump into operations.
Do AI pilots fail because of the technology?
Rarely. The causes usually lie in workflows, data, decision rights and missing ownership. Those are questions about how the company is built.
Further reading: The AI-first operating model · The AI tool zoo · Evaluating AI investments
Sources: Mayer et al. (2025): Superagency in the Workplace, McKinsey · IBM Institute for Business Value (2025): CEO Study · Bitkom (2026): Künstliche Intelligenz in Deutschland · Engels et al. (2025): IW-Report No. 33 · Bundesnetzagentur (2025): KI in Unternehmen · Nicole Lontzek (2026): Die neue Betriebslogik, Chapters 2 and 13
