Productivity
Jun 18, 2026
Most AI projects fail before anyone writes code
The expensive mistake is almost never the build. It is the decision, made months earlier, about what to build.
There is a familiar shape to a disappointing AI project. The technology worked. The demo was impressive. The team delivered on time. And eighteen months later the thing is switched off, or worse, still running while everyone quietly routes around it.
When that happens the post-mortem usually examines the build. It rarely examines the decision that preceded it, which is where the failure actually occurred. Somebody chose a problem that was not worth solving, or not solvable in the way they assumed, and every subsequent decision was made competently in service of that error.
How the wrong problem gets chosen
Three routes lead to the same place, and none of them involve anyone behaving unreasonably.
The first is vendor-led selection. A supplier demonstrates a capability, the capability is genuinely impressive, and the business works backwards to a problem it might address. The demo is optimised for demonstrability, not for resemblance to your operation. The gap between the two is where the budget goes.
The second is visibility bias. The processes that get automated are the ones leadership can see. Those are frequently not the ones consuming the most capacity. The genuinely expensive work in most businesses is diffuse — fifteen minutes here, a re-keyed record there, a question asked of a colleague who was mid-task. It never appears on a process map because nobody has ever written it down.
The third is enthusiasm. Someone in the business becomes genuinely capable with AI tooling and starts building. This is a good thing and should be encouraged. It becomes a problem only when local enthusiasm sets the organisation's priorities by default, because whatever that person finds interesting becomes what the business invests in.
The question that is actually hard
Ask a leadership team where AI could help and you will get a list. Ask which item on that list would still be worth doing if it took three times longer than expected, and the list gets considerably shorter. That second question is the one worth spending time on, because it is a proxy for whether the value is real.
The businesses that get the most from AI are not the ones that adopt the most of it. They are the ones that were most rigorous about what not to do.
Rigour here is not a matter of building a scoring matrix, though a matrix can help. It is a matter of insisting on specificity. 'Improve customer service with AI' is not an opportunity; it is a category. 'Draft the first response to the four enquiry types that make up most of our inbound volume, using live order data, for an agent to review' is an opportunity. You can estimate the second one. You can test whether it worked. You can tell, in advance, whether the data it needs actually exists.
Four tests worth applying before committing
Can you describe the work in a sentence? If the description requires a paragraph and two caveats, the scope is not yet understood well enough to build against.
Does the information already exist, in a form something can read? A great many opportunities are really data projects wearing an AI costume. That is fine, but it changes the cost and the timeline.
What happens when it is wrong? Not whether it will be wrong — it will be. Whether a wrong answer is caught, tolerable and recoverable, or silent and expensive.
Who loses if this works? Every automation redistributes work. If nobody has thought about whose job changes, adoption will answer the question for you.
The cost of choosing well
Proper opportunity identification takes weeks, not days. It requires talking to people who do the work rather than people who describe it, and it produces conclusions that are sometimes unwelcome — including, reasonably often, that the best available option is not an AI project at all.
That is a difficult thing to sell and an easy thing to skip. It is also the single highest-return activity in the whole endeavour, because it is the only stage where the cost of changing your mind is close to zero.
The build is the easy part. It has always been the easy part. Choosing what to build is the work.
