95% of enterprise AI pilots deliver no measurable business value. This figure from an MIT study in the summer of 2025 has reached the boardrooms – and has been quoted up and down ever since. Sometimes as proof of the great AI bubble, sometimes as a wake-up call, sometimes simply as a hook for the next consulting offer.

The number is punchy. But as with many punchy numbers, the actual insight does not lie in the headline. It lies in the question it raises: Why do these pilots fail – and what does that tell us about the state of digital transformation in organizations?

What the Study Actually Found

The research by the MIT NANDA initiative is based on 150 interviews, 350 employee surveys, and 300 public deployments. The main finding: only 5% of pilots lead to measurable revenue or efficiency effects. The remaining 95% remain stuck in a kind of perpetual pilot mode.

More interesting than the ratio is the "why". Lead author Aditya Challapally puts it clearly: it is not the quality of the AI models that makes the difference, but the learning gap between tools and organizations.

This is a remarkable statement. It shifts the focus from a technology problem to an organizational problem. And it aligns precisely with what transformation leaders have known for years – but which is often enough ignored.

Pilots Don't Die From the Model

Anyone who has accompanied an AI pilot in a larger company knows the pattern: technical feasibility is demonstrated, initial use cases work, a PowerPoint receives a standing ovation. And then – nothing happens. Or rather: the pilot is repeated, expanded, rebranded, but never scaled.

This rarely has to do with the model. It has to do with the things no one put on the roadmap:

  • Process integration – the output of the AI has to feed into an existing workflow somewhere. If that workflow is not adjusted, the value evaporates.
  • Data foundation – the quality of the answers depends on the quality of the inputs. Organizations that don't have their data under control will not get reliable results from the next model either.
  • Role clarity – who still decides, who only prepares, who controls? These questions are usually left unresolved.
  • Enablement – employees are left alone with the tool. Training takes place, but it rarely conveys the critical competency: recognizing when the AI's output is wrong.

The MIT study confirms this at a pointed moment: externally purchased tools have a success rate of roughly two-thirds, while internally developed solutions succeed only about half as often. Not because internal teams are worse – but because internal builds are usually accompanied by underestimating the organizational side of the work.

The Real Bottleneck: Transformation, Not Technology

This brings us back to a topic I have explored elsewhere: digitalization is not transformation. AI pilots are often digitalization projects in a new guise. An existing process is equipped with a new tool – the underlying logic remains untouched.

Effective use of AI begins where organizations are willing to raise the question one level higher:

  • Why does this process even exist in its current form?
  • Which decisions do we want to delegate to a system – and which deliberately not?
  • Where is our real bottleneck: in doing, in deciding, or in understanding?
  • What changes about leadership when routine decisions are automated?

These questions cannot be answered within a pilot project. They are change management questions. And they determine whether a pilot becomes a production system or not.

Beware of the Counter-Thesis

There is a loud counter-movement accusing the MIT study of methodological weaknesses. Some critics argue that the 95% figure is too broad and does not measure actual value. That is partly true – the study primarily captures short-term P&L effects and possibly underestimates the qualitative value of some pilots.

But this is precisely why the real criticism is not "the number is too high", but rather: "the number captures only part of the problem". Even the 5% of successful pilots are, in many cases, only small efficiency gains – far removed from the transformation that the term "AI" promises in board presentations.

What Transformation Leaders Should Take Away

In my view, three practical consequences stand out:

First: don't treat the pilot as the goal. A successful pilot is only worth something if, at the time of launch, a plan for scaling is already in place. Those who skip this step produce showcases without follow-through – snapshots without staying power.

Second: take the back office seriously. The MIT study shows that the highest ROI is not found in marketing and sales, where most budgets are spent, but in back-office automation. It is not glamorous. It is effective.

Third: think change management from day one. Not as a supporting measure after the kick-off, but as an integral part of the design. Which roles change? What fears emerge? What new skills are needed? Asking these questions in parallel with the technology is the only reliable way to turn a pilot into a system.

Conclusion

The 95% is not the real problem. The problem is that this number is read as a scandal rather than as a diagnosis. AI is being tested on organizations that are simply not prepared for the change AI enables. Technology without organizational transformation is not sustainable – no matter how powerful the model.

The quality of the AI does not determine the pilot's success. The quality of the organization does.

Those who internalize this will not aim for a 5% success rate, but will instead create the conditions under which the question of the rate no longer matters.