Research

The 95% Problem: Why Corporate AI Fails

Based on MIT NANDA, State of AI in Business, 2025

In 2025 an MIT research group examined hundreds of corporate AI deployments and published a number most vendors would rather you never saw: 95 percent of enterprise generative AI pilots produced no measurable profit impact. The companies spent the money. The returns did not show up.

The reason matters more than the number. The failures traced back to how organisations used the tools: pilots that never fit real workflows, staff who never learned what the models can and cannot do, projects measured on demos instead of output. The underlying technology was rarely the blocker.

What the failures had in common

Failed pilots shared a shape. A tool arrived from the top down. Nobody mapped it to the tasks people actually do all day. Training amounted to a login and a memo. Within months the tool sat unused, and the project was written off as an AI failure when it was a skills and process failure.

What firms are actually doing about it

This lines up with what the Federal Reserve Bank of New York found in 2026 when it asked firms directly: most companies using AI are retraining existing staff rather than replacing them, and job posting data shows no distinct AI-driven drop in hiring demand. The measured story is a shortage of people who can use the tools well, not a wave of cuts.

What this means for you

For an individual, the 95 percent number reads as an opening. Companies have proven they will spend on AI. They have also proven, at scale, that the spending fails without people who know how to work the tools. The person who can close that gap for their own role holds something the measured data says is scarce.

Where to start depends on your actual tasks. Based on what you tell us about your week, the quiz shows which parts of your work sit closest to that gap.