AI in operations: where do you actually start?

The numbers are stark. According to RAND Corporation research, more than 80% of enterprise AI projects fail to deliver their promised value — roughly twice the failure rate of conventional IT projects. MIT's Project NANDA reaches a similar conclusion: the vast majority of generative AI pilots produce no measurable return on the P&L.

The cause is almost never purely technical. Converging studies from RAND, MIT, and Gartner point to the same factors: a poorly defined scope from the outset, insufficiently reliable data, an approach that starts from the technology rather than the business problem, and organizational sponsorship that fades once the pilot phase ends.

That changes how a first AI project should be approached entirely. Three questions help identify a realistic use case, rather than a showcase project destined to join the failure statistics.

1. Where is your team losing time on repetitive tasks?

Manually sorting complaint emails, re-entering data between two systems, generating reports that always follow the same structure — these are natural candidates, not because they're spectacular, but because the gain is immediate and measurable.

2. Do you already have good enough data?

This is the most frequently cited failure factor in Gartner's research on generative AI projects. An AI agent connected to incomplete or inconsistent data just accelerates the production of errors.

3. Who will own this use case day to day?

This exact loss of organizational sponsorship, once initial enthusiasm fades, is one of the leading failure causes RAND identifies — more so than the model's technical performance.

Once these three questions are answered, the right first project often becomes obvious — and it's rarely the most ambitious one on paper. An agent that resolves 80% of first-line support requests beats a general AI platform nobody is still using six months later.