Why most AI projects fail

When I look at the last two years of AI projects (my own, observed, postmortemed), the failed ones look more like each other than the successful ones do. Here are the five patterns I see most often.

1.The problem is never properly defined.

“We want to do AI” is not a spec. Neither is “we want to build LLM-based search”. A proper spec says: who uses this, what was their alternative before, what has to improve measurably, what is allowed to degrade. If those sentences aren't on a single page at the start, the project fails. Usually by scope inflation, sometimes by anti-adoption.

2.It's optimised for demos, not production.

AI demos are so good that they always impress. Production is the same problem with latency, failure modes, cost per call, auditability, and three edge cases per hour the CEO has never seen. Underestimate the demo-to-production gap and you build a stopgap under pressure: one that never leaves stopgap status.

3.The data question is deferred.

Every company has worse data than it thinks. With classical software that was a tolerable nuisance. With AI it's cost factor number one, measured in engineering weeks nobody planned for. “We'll do that later with RAG” is not a plan, it's a wish.

4.The org is left behind.

An AI feature often changes how a team works. If the team doesn't understand why, or fears it'll replace their job, the feature gets quietly or openly sabotaged. I've seen projects that were technically perfect and quietly switched off after six months because nobody used them. That isn't a technical question.

5.The success metric erodes.

At the start: “we want to double engineering productivity”. Three months in: “we want to show that anything works”. Six months in: “we want to write the lessons-learnt deck”. That erosion is normal if you don't fight it. Fighting it means pulling out the original spec every month and measuring honestly.

What the successful ones share

A sharp problem. A champion with the authority to push. Someone who reads the numbers honestly. A team that knows the price. Patience of 12+ months. And a plan B in case AI doesn't help.

That's it. Projects that have worked since 1999 and projects that have worked since 2024 share that list. The tools change. What leads to success doesn't.