OptAI®
EN

Writing Applications

Why most AI pilots go nowhere

MIT found that 95% of corporate generative AI pilots make no measurable difference to profit. The fault lies in the last mile, where the software meets the company.

A small white model aeroplane tethered to a test stand on a yellow background

In the summer of 2025 a team at MIT published a report with an unkind number in it. Of the generative AI pilots it studied, about 95% had no measurable effect on profit and loss. The researchers interviewed 150 executives, surveyed 350 employees and looked at 300 public deployments. Companies spent the money and saw no return.

The finding travelled fast, partly because it seemed to say that the technology does not work. It says nothing of the kind. The authors blamed a "learning gap". The tools did not fit the way the companies worked, did not remember what they were told and did not get better with use. A chatbot that impresses in a demo and has forgotten everything by Monday is a toy with a budget.

McKinsey's annual survey points the same way from the other side. In 2025, 88% of the companies it asked used AI in at least one part of the business. Only about 6% said AI had brought them significant value and 5% or more of their operating profit. The 2026 edition, published in August, found the same 6%. So most companies use AI, and very few earn from it.

Where the money went

MIT also looked at where companies spent. More than half of generative AI budgets went to sales and marketing tools, which are easy to buy and easy to show the board. The best returns came from the back office, where AI replaced outsourced processes and agency work. In practice that means invoices, contracts, claims and procurement. It is unglamorous work with clear rules, clean inputs and a cost everyone can measure. That makes it the right place to start.

The second finding is less comfortable for large IT departments. Tools bought from specialist vendors and built with partners succeeded about 67% of the time. Systems built entirely in-house succeeded only a third as often.

A pilot is designed to be stopped

Most pilots fail because they are built as pilots. They run on a clean sample of data, beside the real systems rather than inside them, with a few keen users and nobody accountable for the result. When the pilot ends, someone has to connect it to the ERP, get it past data protection, train the other 300 people and answer for its mistakes. That work was never in the plan, so it never happens.

Production starts with different questions. Which job costs the most hours? Where does the data sit, and who may see it? What does a wrong answer cost, and who catches it? What will this save per month, in francs? A project that cannot answer them in the first week will not answer them in the sixth month.

How we work

At OptAI® we build the first version on the customer's real data, inside the systems they already use. It is slower in week one and much faster by week eight, because nothing has to be rebuilt. We measure the job before we automate it, so the saving is a number rather than a feeling. And we agree at the start what would make us stop.

The 5% in MIT's study had no better models than anyone else. They picked one painful job, connected it properly and kept at it until it paid. That is less exciting than a demo. It is also the only part that shows up in the accounts.

Sources

  1. Fortune: MIT report, 95% of generative AI pilots at companies are failing, 18 August 2025
  2. Computing: MIT report, 95% of corporate generative AI pilots fail to deliver returns
  3. Virtualization Review: MIT report finds most AI business investments fail, 19 August 2025
  4. McKinsey: The state of AI in 2025, agents, innovation and transformation, 5 November 2025
  5. McKinsey: The state of AI in 2026, on the road to ROI, 25 August 2026

More writing