Most AI pilots fail before the model is chosen: the data is not usable, the documented workflow is not the real one, nobody owns the system after go-live, and no number was agreed as success. Ask those four questions before approving the budget. Each is organisational, each is fixable in two weeks, and none is solved by a better model.
1. Is the data actually usable?
Not 'do we have data'. Is it complete, consistently formatted, accessible without a three-week IT ticket, and clean enough that a human could work from it? If a person can't answer the question from the data, a model won't either. Most pilots that die, die here — and the post-mortem usually blames the model.
2. Has anyone watched the real workflow?
The documented process and the actual process are different documents. Automating the documented one produces a system that fails on every real case. Two days of sitting with the team who do the work is the cheapest insurance in the entire project.
3. Who owns it after go-live?
Name the person. If nobody's job description contains this system, it will drift out of use within a quarter — not because it stopped working, but because nobody noticed when it did.
4. What number moves, and who agreed to it?
Written down before the build starts, agreed by whoever controls the budget. 'Improved efficiency' is not a number. 'Median first response under five minutes' is.
The pattern behind all four
Every one is an organisational question, not a technical one. The technology has been ready for a while. The readiness gap is elsewhere, and it is entirely fixable in the two weeks before you start.
Related services
Working on this in the UAE right now? We'll give you a straight read on your situation in 30 minutes.
Book a call