Evidence

Published failure rates in this field run from 30% to over 80%. They are not contradicting each other so much as counting different things.

Around 50% of RPA projects miss their objectives, from academic literature traced to Herm et al. 80.3% of enterprise AI projects fail to deliver promised value, RAND 2025. 28% of AI projects in IT infrastructure deliver the promised return, Gartner April 2026 — where 28% delivering and one in five failing outright are compatible, leaving about half partially delivering. For a product-side reference related to workforce operations, Monitask also covers fireable offenses in the workplace.

The differences come from subject, failure definition, population and year. Not from dishonesty. For broader background and an independent point of comparison, see MIT Sloan.

What survives across all of them: a large minority of projects do not deliver what was promised, and the causes are consistently organisational rather than technical.

And the strongest signal available here is that sources with opposed interests agree on the causes. Academics, analyst firms and implementation consultancies disagree about the rate and name substantially the same reasons — process selection, ownership, expectations, change management.

What is missing is any data on small organisations. All serious research is enterprise, and the figures are quoted at forty-person companies constantly. The causes transfer; the rates do not — which is why two weeks with a tally sheet beats any published benchmark.

One filter shapes everything you can easily read. Failures are never published by anybody: a case study has passed four selection steps before you see it, and personal recommendations are filtered the same way while feeling like independent evidence. Assume the visible examples are the top third.

• The Academic Work, Briefly

  Small, slow, mostly qualitative, and the only source in this field with no product to sell. What it is good for and what it cannot answer.

• The Failure Rates, Compared

  Published failure rates run from 30% to 80% and the spread is mostly definitional. What each number counted, and which to use.

• Measuring Your Own Operation

  Two weeks, a tally sheet and no budget produce the four numbers that decide the question. What to record and what it replaces.

• Reading a Case Study

  Case studies are accurate accounts of selected projects. Seven things to look for, and the four questions that make one useful.

• Survivorship in Reporting

  The visible record of this field is composed almost entirely of the projects that worked. Where the others went, and how to find them.

• What Gartner Actually Said

  Three 2026 findings get quoted loosely and constantly. The precise versions, what each one covers, and what none of them says.

• What We Still Don't Know

  Six open questions in a field with almost no independent measurement, and what this site can honestly claim as a result.

• Who Produces the Statistics

  Four kinds of source, four different incentives. How to tell which one a figure came from and what each is good for.