A pattern worth noticing: when an AI initiative stalls, the blocker is rarely the model. It is that nobody can produce a clean, agreed set of records to point it at.
The question that stops projects
Ask what a customer is and you often get several answers. The CRM has one record, accounts has another under a slightly different name, and the delivery spreadsheet uses a third. Each was correct for its own purpose. Together they mean the business cannot state how many customers it has without someone reconciling by hand.
No model resolves that. Asked to work across three definitions, it will produce a confident answer built on whichever one it saw most, and you will have automated the disagreement rather than settled it.
Structure before intelligence
The unglamorous work comes first, and it is the same work that pays off whether or not you ever add AI.
- Agree the entities the business actually runs on — customer, order, product, site — and what uniquely identifies each
- Decide which system owns each entity, so there is one place a change is made
- Validate at entry rather than cleaning up later
- Keep history, so you can tell what a record looked like at the time a decision was made
A business that has done this can adopt almost any tool. A business that has not will struggle with all of them.
Why this order saves money
Cleaning data for one project produces a clean extract that goes stale in a month. Fixing the model underneath produces a system that stays correct because correctness is enforced where records are created.
The second is more work up front and less work permanently. It is also the difference between a pilot that impresses in a demo and a system your team keeps using in the second year.
A practical first step
Pick the report your leadership argues about most. Trace every number on it back to where it was entered. The disagreements you find on the way are your data model problem, written down for you.
Fix those, and the question of which AI tool to buy becomes much easier to answer — and considerably less urgent than it looked.
