Why Most Businesses
Fail With AI

Here is the number that stopped a lot of boardrooms in their tracks: in a widely-cited 2025 study from MIT, around 95% of enterprise generative-AI pilots produced no measurable impact on the bottom line. Only about 5% created real value. Billions were spent; most of it returned nothing.

It's tempting to read that as 'AI doesn't work.' That's the wrong lesson. The models mostly work fine. What fails is the way businesses put them to use, and the failures follow a pattern predictable enough to avoid. (It's worth noting the study is preliminary and self-reported, but its direction matches everything else in the field.)

Failure one: no real problem

Most failed projects started with 'we should use AI,' not 'we have this specific, expensive problem.' A tool bought to look innovative has no target to hit, so it hits nothing. The 5% that succeeded did the opposite: they picked one painful, well-defined problem and aimed the whole effort at it.

Failure two: the tool never touched the work

MIT's core finding was a 'learning gap', generic tools that dazzle in a demo but don't plug into how work actually happens, don't learn from it, and don't adapt. A model answering questions in a sandbox creates no value. The same model wired into a real workflow, learning from real inputs, does. Most pilots stopped at the sandbox.

Failure three: money went where the excitement was, not where the return was

The study found budgets concentrated in sales and marketing (the visible, exciting functions) while the clearest returns showed up in dull back-office automation. Businesses spent where the hype was instead of where the payoff was, and got the ROI that mismatch deserves.

Failure four: building instead of buying

Homegrown tools underperformed. In the study's sample, externally built, customised tools reached deployment far more often than internal builds. Most companies don't have the specialised talent to build well, and burn months discovering it.

What the 5% did differently

McKinsey's research lands in the same place: adoption is nearly universal now, but value is rare, and it comes from redesigning the workflow around the tool rather than bolting the tool onto an unchanged process. The businesses that succeed share a short list of habits: pick one real problem, wire the tool into the actual work, measure the outcome, and buy proven tools instead of building from scratch.

None of that is about having better AI. It's about implementation discipline. The technology isn't the hard part. Using it on purpose is.