Why most AI projects never reach production

The whole world is talking about AI, yet most companies still cannot put it to work. In July 2025, an MIT team published one of the most complete pictures of the problem. It combined an analysis of more than 300 public AI initiatives, structured interviews at 52 organisations and a survey of 153 senior executives (Challapally et al., 2025).

95% of the organisations studied saw no measurable return from generative AI in their results.

The line went round the world, often distorted into “95% of pilots fail”. What the study actually says is more useful: the problem lies mostly in how projects are chosen, designed and integrated.

Pilot purgatory

Of the AI tools designed for specific tasks, 60% of organisations evaluated them, 20% reached a pilot and only 5% reached production. Many projects were designed to produce a demo. They impress in a meeting and die quietly, because nobody planned access to the data, integration with the systems people actually use, or who would own the tool after launch.

The authors identify one central factor: most tools neither learn nor adapt to the company’s context. A tool that does not retain what the team teaches it forces people to repeat the same work every day, and ends up abandoned.

Partners and in-house teams

One of the most cited findings compares two paths. Tools developed with external partners reached real use in around 67% of cases; those built in-house, in around 33%. Employee usage was also almost double for the external tools.

The authors themselves urge caution: the data comes from 52 interviews, it is self-reported by the companies, and the gap may reflect the organisations’ capabilities and not just the path they chose. But the direction is consistent with what we see on the ground.

Three signs a project will get there

A process that hurts. The best projects start from a pain that shows up every week: reports that take days, requests handled twice, forecasts made by gut feeling.

Data that exists. It can be imperfect, as long as it exists. A surprising number of initiatives are designed around information the company intends to collect one day.

An owner on the inside. Tools survive when someone in the operation wants them to survive, which is why adoption has to be planned as a phase of the project.

Once it reaches production, the return shows up

A study by IDC commissioned by Microsoft, covering more than 4,000 business leaders, reports that every dollar invested in generative AI returns 3.7 dollars on average, and 10.3 at the most advanced organisations (IDC, 2024). It is a figure reported by the companies and funded by a vendor, so it should be read as an indication, not a guarantee. But it points in the same direction as MIT: the gap between those who gain and those who don’t lies in the discipline of building only what can be launched, and putting it in the hands of the team that will use it.

That is why we measure our work by what stays in production: your numbers moving and a system the team keeps using.