Most AI pilots fail, but not for the reason you think
I’ve heard this stat come up a lot lately: “95% of AI pilots fail.”
And honestly, I get why it’s resonating. A lot of teams did get burned in that first wave of AI adoption. The pitch looked great, the demo worked, and then… it didn’t hold up inside real workflows with real data.
But here’s the thing: they keep using that stat. That stat does not mean what they think it means.
It comes from an MIT study, and what they found was that only about 5% of AI pilots showed measurable profit-and-loss impact within six months. That’s a very specific definition of success. It’s not “AI doesn’t work”, it’s “most companies didn’t prove financial impact fast enough.”
Those are very different conclusions.
In my experience, a lot of the value from AI just doesn’t show up cleanly on a P&L that quickly. You see it in time saved, fewer errors, faster turnaround, less reliance on outside vendors—real improvements, just not always the kind that get captured in a six-month financial snapshot.
I think what the study actually highlights is something I see all the time: there’s a big gap between companies experimenting with AI and companies actually deploying it in a way that sticks.
A few patterns keep showing up:
One is that most AI tools don’t really “learn” in the way people expect. They’re great for drafts and brainstorming, but they don’t retain context, don’t improve from feedback, and don’t adapt to how your business actually works. That becomes a problem the moment you try to rely on them for anything important.
Another is the build-vs-buy decision. A lot of teams instinctively try to build their own AI systems in-house. I understand why; you want control. But in practice, it often turns into under-resourced internal projects that never quite get to production quality. The teams I’ve seen succeed are usually the ones that lean on vendors and focus their internal effort on guiding and integrating, not building from scratch.
And then there’s where the budget goes. Most of it is still flowing into sales and marketing use cases because they’re easy to demo. But the strongest ROI I’ve seen—and what the data is starting to show—is in the less visible stuff: document workflows, compliance, procurement, finance. The unglamorous areas where small improvements compound quickly.
All of this has pushed me to think about AI adoption a bit differently.
It’s not something you bolt onto the front of your business. It works best when it’s embedded into the core processes—the repetitive, document-heavy, decision-heavy workflows that actually run things day to day.
That’s also why so many pilots stall. They’re treated like experiments instead of systems. They’re measured too narrowly, or expected to deliver results without the structure to support them.
The companies getting real value aren’t necessarily using better models. They’re just deploying them better.
So when you see that “95%” number, I wouldn’t read it as a ceiling. It’s more like a snapshot of where most organizations are right now—early, still figuring it out.
There’s still a real window here for teams that get the deployment model right.
Anyway, curious how this lines up with what you’re seeing on your end. If you’re working through this stuff right now, I’d genuinely love to hear about it.
Want more reliable updates on AI? Check out the new Youtube Channel!
—John Licato, CEO of SquarePact
