Why AI pilots stay on the shelf: what the surveys say

In 2025 a figure went around the world: 95% of AI pilots in companies fail. It comes from a preliminary report by the NANDA research group at MIT, which itself speaks of a 95% failure rate. Its data, however, shows something else: 95% of organisations so far see no measurable impact on profit and loss from generative AI. And for custom or vendor-supplied enterprise AI systems, roughly a quarter of the companies that actually launched a pilot succeeded.

What exactly does the MIT NANDA report say?

The authors tracked more than 300 publicly announced projects, interviewed 52 organisations and collected 153 survey responses. For custom or vendor-supplied enterprise AI systems the funnel looked like this: 60% of organisations evaluated them, 20% got as far as a pilot and 5% reached production. Of the companies with a pilot, roughly one in four succeeded.

The report is labelled preliminary, and its authors write that the results are indicative rather than definitive. According to a podcast by 80,000 Hours, it was not peer reviewed, relies on a small sample and counts as success only a marked and lasting benefit.

Sources: MIT NANDA, The GenAI Divide, 80,000 Hours

What do other surveys show?

So individuals use AI routinely. The difficulty is turning trials into company processes where the benefit can be measured.

Sources: BCG, September 2025, Deloitte 2026, BCG, June 2026

Why do pilots not move into production?

The MIT report gives as the main cause that most enterprise tools do not retain context or learn from feedback. People then use their own chatbots for quick tasks but still hand complex work to colleagues. Employees at more than 90% of the companies surveyed said they regularly use personal AI tools for work. Yet according to the report, only 40% of companies had bought an official subscription to a language model.

Other surveys add a lack of training and guidance (BCG) and, for AI agents, costs, unclear value and weak risk controls (Gartner). Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027.

Sources: MIT NANDA, BCG, June 2026, Gartner

What holds companies back in Europe?

Eurostat asks companies that considered AI but did not adopt it why. In the EU in 2025 the reasons given most often were a lack of expertise (70%), unclear legal consequences (54%) and data protection (53%). Data quality or availability was cited by 44%, incompatibility with existing systems by 42% and cost by 38%. Only 18% said AI was not useful to them.

Source: Eurostat, isoc_eb_ai

What does this mean for companies?

• A pilot needs a measure of success agreed in advance. Without one there is no way to tell whether it is worth continuing.
• The tool has to fit the work and the systems the company already uses. Incompatibility with existing systems was cited by 42% of EU companies that considered AI and did not adopt it (Eurostat).
• People make the difference. A lack of expertise is the most common obstacle in the EU, and employees need to know how to work with AI.

Summary

• "95% of pilots fail" is a simplification that the MIT NANDA report itself used. Its data refers to 95% of organisations with no measurable impact; of the companies that piloted a custom system, roughly a quarter succeeded.
• People use AI routinely, but companies lag behind in turning pilots into processes and measuring the benefit.
• EU companies that considered AI and did not adopt it are held back most often by a lack of expertise and unclear rules; only 38% cited cost.

If you have AI in a trial phase and need to get it into your company systems, tell us what you are working on. How we approach it is described on the AI solutions page. How Czech companies use AI is covered in the article How many Czech companies use artificial intelligence.

As of 1 November 2026.

• artificial intelligence
• AI in business
• AI pilots