Henry Hays dropped a number in a recent episode of The Fast Slow Motion Podcast that stopped the conversation: 90% of AI pilots fail. It came from an MIT study, and if you’ve been inside an AI initiative that stalled, lost momentum, or quietly got deprioritized after the first few months, that number probably doesn’t surprise you.

What’s worth understanding is why. Because the failure rate isn’t a technology problem. It’s not that the tools don’t work. It’s that the conditions required for those tools to work are almost never in place when companies decide to start building.

What the Study Is Actually Measuring

The MIT research that surfaces this statistic isn’t measuring whether AI tools are capable. It’s measuring whether organizations that attempt to deploy AI in a meaningful way actually get to a point where that deployment produces sustained value. The answer, most of the time, is no.

That framing matters because it changes what the problem actually is. If 90% of pilots fail, and the tools themselves are functional, then the failure is happening somewhere between the decision to invest and the point where the investment starts paying off. Something is breaking down in the middle, and it’s breaking down consistently enough across industries that the pattern is worth taking seriously.

The Assessment Step Nobody Takes

Hays uses a medical analogy that makes the problem concrete. A surgeon doesn’t walk into the operating room and decide what to do when they get there. The patient has been evaluated, the procedure has been planned, and the team knows exactly what they’re walking into before anyone picks up a scalpel.

Most AI pilot failures happen because companies skip the equivalent of that evaluation. They identify a use case, pick a tool, and start building without ever asking the three questions that determine whether the initiative has any real chance of working.

Is the technology ready? Not just “do we have software that could theoretically support this,” but are the systems this AI will need to connect to actually in a state where that connection is going to be reliable and useful?

Is the data ready? This one kills more pilots than any other single factor. AI doesn’t generate insight from nothing. It operates on whatever information you feed it, and if that information is incomplete, inconsistent, or siloed across systems that don’t talk to each other, the output will reflect that. Hays is blunt about this: if everything else is ready and the data isn’t, you’re burning money.

Are the people ready? This is the question that gets skipped most often, and it’s often the most important one. Do the people who will be working alongside this AI understand what it is and what it isn’t? Do they understand that the goal is to augment what they do, not replace them? If the answer is no, the best-designed pilot in the world will run into a wall the moment it touches real people doing real work.

Why Pilots Get Approved Without These Answers

Part of the reason companies skip the assessment is that assessments take time, and there’s real pressure to move. AI is on every board agenda. Competitors are talking about it. The executive team wants to see progress. So someone gets assigned to stand up a pilot, and the pilot gets stood up, and six months later it’s quietly been shelved because it never really worked.

The other reason is that no one wants to be the person who says the organization isn’t ready. That conversation requires honesty about the state of the data, the maturity of the systems, and the readiness of the team, and those aren’t comfortable conversations to have in a room full of people who are excited about AI.

But the discomfort of that conversation is far smaller than the cost of a failed pilot. Failed pilots don’t just waste money. They create organizational skepticism that makes the next attempt harder. When a team goes through an AI initiative that doesn’t deliver, they carry that experience forward. The next time someone proposes an AI project, the response in the room is more cautious, and sometimes more resistant, than it would have been otherwise.

What a Real Readiness Check Looks Like

A proper pre-pilot assessment doesn’t have to be a lengthy engagement. It’s a structured set of questions designed to surface the blockers before they surface on their own, mid-build, when fixing them is much more expensive.

On the technology side: what systems does this initiative touch, are those systems stable and well-maintained, and what integrations will be required to make this work? Are those integrations straightforward, or are they going to require significant work before the AI component can even be evaluated?

On the data side: where does the data live, who owns it, how clean is it, and how consistent is it across systems? Has anyone actually looked at the data that will be feeding this AI, or is the assumption that it’s probably fine?

On the people side: who will be working with this AI day to day, have they been involved in the planning, and do they understand what’s being asked of them? Is leadership visibly behind this, or is this an initiative being handed down from above without real buy-in at the operational level?

None of these questions are technically complex. They’re just questions that require honest answers, and honest answers sometimes reveal that the organization needs to do some foundational work before the pilot makes sense.

The Cost of Skipping This Step

The 90% failure rate isn’t inevitable. It’s the result of a predictable pattern of skipping the preparation that makes the difference between a pilot that builds momentum and one that gets quietly canceled.

Companies that do the assessment work first move faster in the long run. They don’t spend months building something that runs into a data problem they would have found in week one if they’d looked. They don’t lose organizational credibility on an initiative that was underprepared from the start. They don’t have to overcome the skepticism that a failed pilot leaves behind.

The businesses that are getting real value from AI right now aren’t the ones that moved fastest. They’re the ones that moved most deliberately, which is a different thing entirely.

Listen to the full podcast episode here.

Related Resources

• The AI Execution Problem: Why Most Growing Businesses Can’t Get Past Experimenting: The 90% failure rate is one symptom of a larger pattern. This post covers the full picture of why AI execution breaks down in growing businesses and what it takes to get through it.
• How to Evaluate Any AI Investment Before You Spend a Dollar: A practical framework for evaluating any AI investment before committing budget, built around the questions that separate well-prepared initiatives from the ones that stall.
• Three Things That Have to Be True Before an AI Investment Pays Off: Technology, data, and people readiness are the three variables that determine whether an AI pilot succeeds or fails. Here’s how to assess each one honestly before the build starts.