There's a Statistic for Every Side. AI Is No Exception.

There’s a study proving that AI in companies is a failure. And there’s another proving it works wonderfully. Both are real, both come from serious sources, and both are available to anyone who wants to cite them depending on the point they need to defend this week.

The MIT study and the 95%

The most cited one in recent months was one from MIT. The headline said that 95% of AI pilots in companies generate no return at all. It was repeated in Forbes, in Harvard Business Review, in media across half the world, and it even contributed to a drop in the Nasdaq. It became the favorite line of everyone who already suspected AI was overrated.

When I read it, I’ll admit it suited me. It fit with something I’ve been saying for a while: that the problem with AI is almost never the tool, but the way organizations decide to adopt it. I had my statistic. The one that proved me right.

What the report actually says

So I went to check it, as anyone should with a number that suits them a little too well.

What I found was uncomfortable, but not in the direction I expected. The 95% doesn’t mean what everyone repeats. Read carefully, the report shows that about 80% of the companies surveyed had never run an AI pilot. Not one. And of those that did, around a quarter succeeded. Saying 95% failed is like saying 95% of users of a dating app are in unhappy marriages, when 80% haven’t even been on a first date.

The report itself warned about this: its figures are directionally correct, based on individual interviews and not on companies’ official reports. It describes itself as a preliminary finding. The warning was there from the start. We skipped it because the headline confirmed what we already wanted to believe.

I’m not going to play the trick in reverse either and declare that AI is actually going great. It isn’t. The study’s direction holds: most companies are stuck in trials that go nowhere. The number is wrong. The problem is real.

When statistics become ammunition

And there’s the real point. When there’s a statistic for every position, both stop being evidence and become ammunition. You pick the one that serves you and get on with your day. The debate settles itself, because nobody was looking for the truth, only for confirmation.

The two numbers that matter

So if the numbers cancel each other out, what’s left? Two figures from the same report, side by side, that say more than the headline. Only 5% of custom-built AI tools made it into production. And more than 90% of employees at those same companies were already using tools like ChatGPT on their own, every day.

Read that again. The same people who “fail” with AI at the organizational level are using it without any problem at the individual level. The company spent on an official pilot that returned nothing measurable, while real adoption was happening underneath, without permission and without a record, in its own people’s personal accounts. The tool was never the variable. Learning to use it was.

The Excel lesson

It’s the same pattern we saw with Excel. Excel never failed or succeeded as a corporate initiative. People learned it, or they didn’t. Those who didn’t rarely said “I learned Excel badly.” They said “Excel doesn’t work for what we do,” and fell behind while someone else built in twenty minutes what took them a full day.

The question that does matter

The question everyone is arguing about, “does AI work?”, doesn’t make sense to me, because a new study will always appear for each side. It’s a comfortable question precisely because it demands nothing from whoever asks it.

The question that does matter is a different one, and it’s more uncomfortable because the answer depends on you: is your organization going to learn to use this, or is it going to keep waiting for the numbers to agree?

Because the numbers aren’t going to agree. And while your company waits for the definitive study, someone in your industry has already learned to ask the right questions.