Why do most AI projects fail to reach production?

Failure rarely announces itself. A pilot impresses, a demo circulates, and a year later nothing runs in production. A 2024 RAND Corporation report, built on interviews with 65 practitioners, cites estimates that more than 80% of AI projects fail — about twice the failure rate of IT projects that do not involve AI.

The causes are organisational before they are technical

RAND's interviewees put one cause above all others: the people commissioning the AI and the people building it did not agree on what problem it was meant to solve. The other leading causes follow the same pattern — the data needed to train an effective model did not exist or could not be reached, the organisation chased the newest technology instead of a real problem, or the infrastructure to deploy and monitor the model was never budgeted.

Notice what is not on that list: the model. Model quality is rarely the binding constraint. By the time a model underperforms, the project has usually already gone wrong upstream.

AI also fails when it is aimed at the wrong tasks

A 2023 Harvard Business School field experiment with 758 consultants found that AI assistance made them meaningfully faster and better on tasks within the technology's capability — and 19 percentage points less likely to produce a correct answer on a task outside it. The same tool, opposite results. Choosing which tasks to hand to AI is itself a strategy decision, and getting it wrong makes work worse, not just no better.

What prevents failure

Three things, done before any build starts. First, write the problem down as a number: the line that should move, the baseline it moves from, and who signs off that it moved. Second, check the data honestly — whether it exists, whether it can legally be used, and whether it actually contains the signal the use case needs. Third, name an owner who is accountable for the system's results and its running cost after go-live, because a system that belongs to everyone belongs to no one.

If a proposed use case cannot survive those three checks, cancelling it early is a success, not a failure. The money saved funds the use cases that can.