I have watched more than a dozen organisations try to put AI into a process. The failures have almost nothing in common with each other except one thing: they started at the interesting end.
The pattern
Someone identifies a workflow that is expensive and manual. They are right about that. Then they go looking for a model, a vendor, a pilot. They build a demo. The demo is impressive. The demo works on the three examples they chose.
Then it meets production, and the first question is: where does the input come from?
And nobody knows, because the input lives in four inboxes, a WhatsApp group and a spreadsheet that one person maintains and everyone pretends not to depend on.
What the boring version looks like
The boring version starts by making the input exist in one place, in one shape, with one definition of each field. No model. No demo. It takes six weeks and nobody wants to fund it, because it looks like tidying.
Then the interesting version takes four days instead of four months, because the model has something coherent to read.
I have never seen this fail. I have seen it be skipped constantly.
Why it gets skipped
Because tidying is invisible. A demo is a slide. A schema is a meeting where you argue about whether a customer is a person or an account.
The incentive is to show something. The boring thing shows nothing until it is finished, and then it shows everything.
A test I use now
Before any AI project starts, I ask one question: if we replaced the model with an intern, could the intern do this?
If the answer is no — because the intern would not know where to look, or what the fields mean, or which of the three conflicting numbers is the real one — then the model cannot do it either. The problem is not intelligence. It is structure.
Fix the structure. Then the model is a multiplier instead of a very confident random number generator.
The uncomfortable part
This advice is unpopular with people selling AI and with people who want to buy it, because it reframes the project as something less exciting: data hygiene, ownership, definitions, and the willingness to say "this number is wrong" out loud in a meeting.
But it is the difference between a pilot that ends in a case study and one that ends in quietly not being mentioned again.
