Why do AI pilots fail to reach production?

What does the 5% figure actually mean?

The preliminary report The GenAI Divide, by authors associated with Project NANDA, covers January–June 2025. It describes a review of over 300 public initiatives, interviews with representatives of 52 organisations and responses from 153 leaders. The authors state that the views are their own, not those of affiliated institutions.

Its summary claims that 5% of integrated AI pilots extract millions in value. That does not mean the others fail technically. The methodology defines success as use beyond the pilot with measurable indicators, with return impact tracked six months after the pilot.

These are preliminary, largely self-reported observations, not representative statistics for every AI project. Definitions vary, selection may affect the sample and the short window may miss later benefits. Do not merge a lack of reported financial impact, failure to launch and technical error into the same event.

What was Gartner’s prediction?

In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons named were:

• Poor data quality.
• Inadequate risk controls.
• Escalating costs.
• Unclear business value.

This is a historical forecast, not a confirmed outcome for that year. It is not the same measure as the NANDA report. Do not add the percentages or turn them into a general probability of failure.

What is missing between a demonstration and daily work?

Check how the trial will work beyond selected examples: who has access, where data comes from, how exceptions are handled and what happens when the system changes. The technical work depends on the task; not every use requires a custom database or complex infrastructure.

Before launch, clarify:

• Who accepts the result and against which criterion?
• Which actions are permitted and which require a person?
• How are errors, decisions and changes recorded?
• How do you stop, restore and maintain the work when something goes wrong?

The NANDA report compares approximately two thirds reaching deployment through external partnerships with about one third for internal builds in its interview sample. The authors warn that organisational capabilities and conditions differ. This does not establish that hiring a provider causes higher success or that a model is never the problem.

Why clarify governance during the pilot?

WitnessAI, a provider of AI governance and security tools, describes how late security, legal and compliance reviews can stop a technically working trial. It advises setting approval criteria and evidence from actual operation earlier. This is provider guidance, not an independently measured frequency of causes.

In practice, record who needs to review the trial, what they need to see and which questions remain open. Match checks to data, actions and consequences. You need not turn every small experiment into a large organisational programme, but an absence of questions is not approval.

What is the practical rule?

A bounded task, owner, baseline and acceptance criterion before building. Test typical and difficult cases, human review time and stopping conditions. This recommends checkable work; it is not a proven formula for reaching production or a savings guarantee.

Questions

Sources

• The GenAI Divide — preliminary report, copy hosted by Steelcase.
• Gartner — forecast for project abandonment after proof of concept.
• WitnessAI — barriers between pilots and production.

Related reading

• How much does AI consulting cost?
• How much does AI implementation cost after assessment?
• What can you check yourself with an AI readiness checklist?