There is a pattern to how AI projects fail in ordinary businesses, and it is almost never the model.
A company buys or builds something promising. It works beautifully in the demo, on the three examples someone picked. It goes live against the real data, and the results are subtly, persistently wrong in ways that are hard to explain and harder to fix. Eventually it is quietly stopped being used, and the conclusion drawn is "AI is not ready".
AI was ready. The data was not.
Why this happens specifically with AI
Every system is limited by its inputs, so this is not a new law. What is new is the failure mode.
A traditional report with bad data underneath produces obviously bad output. A total that is clearly wrong. A row that does not balance. You notice.
A language model given bad data produces plausible output. It reconciles the contradiction for you, silently, and hands you a confident answer. Ask it to summarise your customer base and it will not tell you that "Acme Ltd", "ACME Limited" and "Acme" are the same company — it will just give you a number that is wrong in a way nothing flags.
The model does not degrade loudly. It degrades convincingly. That is what makes data quality a different order of problem here than it was for reporting.
What "clean" actually means
Not perfect. Not a two-year programme. Four specific properties, in order of how much trouble they cause:
- One name per thing. The same customer, supplier, product or site is written the same way everywhere. This single issue causes more wrong answers than the other three combined.
- One source per fact. When two systems hold the same number, one of them is authoritative and everybody knows which. If nobody knows, both are wrong.
- Consistent shape. Dates in one format. Amounts with a currency. Empty meaning empty rather than "0", "N/A", "-" and a space.
- A definition for the ambiguous fields. What counts as an active customer, what date a job is "complete", whether a figure includes tax. Half of all reporting disputes are definitional, not numerical.
None of this is difficult. All of it is tedious, which is why it does not get done until something forces it.
The useful reframe
Data cleaning is a terrible thing to sell and a terrible thing to buy. Nobody gets excited about deduplication.
So reframe it: the cleanup is not preparation for the AI project, it is the first deliverable of it. Sorting out one name per customer improves your reporting today, your invoicing today, and every automated thing you do later. It pays for itself before anything intelligent is switched on.
That is also the honest test of whether an AI initiative is real. If the plan does not include getting the underlying data into a state worth reading, it is not a plan. It is an expectation.
The one-hour version
Take your customer list. Sort it alphabetically. Read it.
You will find duplicates within about ninety seconds, and you will find them in the same list that feeds your reporting, your invoicing, and whatever you build next. Fixing that list is not an AI project. It is the thing that decides whether your AI project works.