For the past eighteen months we have been walking into companies that had already tried artificial intelligence. The conversation nearly always starts the same way: “we tried it, it worked well in testing, and then it just stayed there”. When you dig, the tool is usually fine. What does not work is the place they put it in.
A pilot is not an operation
A pilot runs with three motivated people, the easy cases, and someone from IT paying attention in case anything breaks. A real operation is a different animal: forty people in a hurry, odd cases every single day and nobody watching. An assistant that gets it right 85% of the time is a triumph in the demo and a problem in production, because nobody has decided what happens to the other 15%.
That is exactly where most projects die. Not in the model, but in the absence of answers to some very untechnical questions: who reviews this, against what criteria, how quickly, and what happens when the reviewer is on holiday.
The four places it gets stuck
- Nobody owns the whole process. Each department controls its own stretch and the problem lives in the handovers. AI gets dropped into one stretch, improves that stretch, and the overall result does not move.
- The original process was bad. It gets automated as is, and now you are wrong faster and with better formatting.
- There is no baseline. Nobody measured how long it took before, so three months later the discussion about whether it improved is a discussion of opinions.
- Nobody's job changed. People were asked to use a new tool, but their tasks, targets and meetings stayed the same. With a full diary, the tool is the first thing to go.
The week-eight test
There is a quick way to tell whether an AI project will hold: look at how many people are still using it eight weeks after launch, with nobody reminding them. If the number has dropped since week one, the problem is not the model. It is that those people's work never changed.
If usage has fallen by week eight, do not buy a better model. Go back to the process.
What we do differently
We do not start by asking which tool you want. We start by walking the process end to end with the people who run it, measuring how long it takes today and writing down where it actually jams. Quite often the conclusion is that two steps are unnecessary, and removing them is worth more than any model.
When AI genuinely does help, the work is rebuilding the whole path: what the machine does, what the person does, where they hand over, what gets recorded and what happens when something goes wrong. And training whoever will run it while it is being built, not in a two-hour session once it is finished.
It is not more expensive. It usually costs less, because half the ideas fall away before anybody writes a line of code.