The board has decided: artificial intelligence is to be used in the company. The strategy is in place, the first budgets have been approved, and there might even be a pilot project. Yet, surprisingly little happens. Or the wrong things happen.
A manager I advised in exactly this situation described it like this: “Some are experimenting wildly and uploading customer data to ChatGPT, while others completely ignore the topic. And I’m stuck in the middle, not even knowing what the next right step is.” Six months after the official AI initiative, his company had spent a lot of money on licenses, produced a lighthouse project without connection to regular operations, and had not a single established usage policy.
AI implementation rarely fails due to technology. It fails due to leadership that does not recognize it primarily needs to solve an organizational and cultural problem.
The problem manifests in three typical patterns: uncontrolled proliferation, where employees use AI tools without guardrails and upload sensitive data to publicly accessible systems. Quiet refusal, where everyone officially supports the initiative, but in practice, no one changes their behavior. And activism without foundation, where ambitious goals are announced, but no clear data strategy, no defined use cases, and no realistic expectations underpin them. “Do something with AI” is not a strategy. It’s an invitation to fail. Three levers can lead out of this trap.
In most companies today, employees are already using AI tools without the organization knowing or controlling it. This phenomenon, shadow AI, is the AI equivalent of shadow IT. The problem is not the usage itself, but the lack of control. If employees enter customer data, contract content, or internal strategy papers into public AI tools, risks arise that extend far beyond the IT department: data breaches, compliance violations, loss of trade secrets. In regulated industries, this can trigger reporting obligations and jeopardize certifications.
The answer is a clear prohibition of uncontrolled use of public AI tools with company data. But a prohibition alone is not enough. Those who only forbid without offering alternatives drive usage deeper underground. Both are needed: an unequivocal no to uncontrolled use and an equally clear yes to approved, secure tools. This includes data classification, which defines which data may enter which systems, and a usage policy that is regularly updated because the tool landscape is evolving rapidly.
| Activism (typical mistake) | Guided implementation (what works) |
|---|---|
| Release all tools simultaneously | Define approved tools, prohibit the rest |
| “Everyone decides for themselves what’s okay” | Data classification: which data into which systems |
| Prohibition without alternative | Clear policy plus approved tools |
| Regulate once, then forget | Regular review, because the tool landscape evolves |
Regulation has long been operational reality. The EU AI Act is in force, and the grace periods for essential requirements have expired. Companies operating AI systems today without documented risk classification and defined human oversight risk tangible compliance violations. Governance here is not paralysis, but a prerequisite.
Successful AI implementation follows a pattern that experienced transformation managers know: start small, learn quickly, then scale.
A division head I supported in AI implementation in her area chose summarizing meeting minutes as the first use case. Not the most ambitious use case, but the most tangible. After four weeks, her team demonstrably saved three hours per week. This sounds modest, but the real value lay elsewhere: the pragmatists in the team saw concrete benefits for the first time, the skeptics realized their concerns were heard, and the uncertain ones noticed that AI complemented rather than replaced their work.
You don’t have to become a technology expert. But you need to understand enough to ask the right questions. Ethan Mollick of the Wharton School describes a central principle in his research on organizational AI implementation: AI does not work as an isolated tool, but only when people learn to collaborate with it, experimentally, not passively trained. AI is a tool, not an end in itself. Before you ask “Where can we use AI?”, you should ask “What problems do we want to solve?”. AI changes roles, not just processes: if an AI assistant takes over routine tasks, the clerk suddenly has 60 percent of their time free. For what? Someone has to answer that. And the quality of the results depends on the quality of the data. Those who have kept data in silos for years and ignored standards will not be rewarded by AI, but punished.
The real work begins when scaling after the pilot. The biggest mistake is not being too slow. It is scaling too fast before the organization understands what it is doing.
Many companies rely on training programs for AI implementation. This is necessary, but not sufficient. managers do not need the ability to build AI models themselves, but rather basic technological competence aimed at decision-making: What are the limits? How do I evaluate result quality? How do I explain to my team why we are doing this? Specialists need practical skills in dealing with the specific tools of their daily work. And the organization needs multipliers: someone in each team who is a little further along than the others and serves as a collegial point of contact.
The crucial level here is middle management. Division and department heads are expected to drive the AI transformation, while also ensuring day-to-day business, and at the same time, they themselves are just beginning to understand what the technology means. Many resolve this dilemma by postponing engagement with AI. This is understandable, but fatal, because if this level does not actively support the implementation, it will not happen.
What middle management needs is not an additional project, but support in integrating AI into existing leadership work: protected time for their own engagement, clear expectations from above, room for decision-making for implementation, and access to expertise. If, as executive management, you leave your division heads alone with AI implementation, you have delegated responsibility, not empowerment.
How you deal with the four typical reactions—the enthusiastic, the pragmatic, the skeptical, and the uncertain—is a stress test for your leadership culture. Take the skeptics’ objections seriously, because they ask the questions that enthusiasts overlook. And create clarity for the uncertain: not with promises you cannot keep, but with honest communication about what will change and what will not. The implementation of AI says more about you as a manager than any strategy presentation.
First: Can you describe in two sentences what specific problem AI is supposed to solve in your area? If not, start there, not with tool selection.
Second: Do you know which AI tools your employees are already using and what data they are entering there? If the answer is “no,” you have a shadow AI problem that you should address this week.
Third: Ask your division heads how much time they had last week to engage with AI. If the answer is “none,” it’s not a lack of motivation, but a lack of space. Create it.
AI will change your company. The question is not if, but how: led or unled, planned or chaotic, with your employees or against their resistance.
A company that has cleanly implemented three use cases in twelve months and brought its employees along is further ahead than one that has started twenty pilot projects, none of which reached regular operation. Implementing AI without overwhelming the organization does not mean being slow. It means leading instead of being driven.
From Pilot to Production – Why 70 percent of AI pilots fail at scaling and what the other 30 percent do differently.
Technical Literacy – What digital competence managers need to maintain strategic control.
All insights can be found in the Overview.
Proven tools and models for self-application are available under Solutions.
If you want to take these thoughts further for your company, a no-obligation initial conversation is worthwhile.
