The real reason behind the 95% failure rate

Most people quoting the 95% AI failure stat have never read the report it came from.

It's MIT NANDA's GenAI Divide study from last year, and it went round every boardroom as proof AI doesn't work.

Read the method and the number looks different. A pilot only counted as a success if it showed measurable return in the profit and loss about six months after the pilot ended. Most changes to how people work take longer than that to show up in the accounts.

The more useful finding is further in. The report puts most of the blame on how the tools were fitted into the work, and very little on the models themselves. Tools that sat beside the job got used when someone remembered and dropped when things got busy.

Picture a small accountancy practice that pays for an AI tool and tells everyone to try it. Six months later, two people use it to draft emails.

So when someone quotes 95% at you, read it as a warning about how you roll AI out.

Before you buy your next AI tool, pick one process and decide exactly which step AI takes over.