Data governance: the precondition for useful AI

Most enterprise AI projects don't die at model selection. They die at step two: the assistant reads real data and returns an answer nobody dares act on.

Not because the model is weak. Because the same customer exists three times under three spellings, because the "date" column means order date in one table and delivery date in another, because "revenue" on the sales report and "revenue" in the ledger are different numbers — and nobody in the company can say which one is right.

That isn't an AI problem. It's data governance, and it predates the AI.

AI doesn't create data problems — it exposes them

People patch data gaps with memory. The chief accountant knows "ABC Company" and "ABC Co." are one entity; knows the March report needs two test rows removed. That knowledge lives in a head, not in a system.

Put an assistant in the middle and everything patched by memory surfaces at once. The assistant counts three customers where you know there is one.

This is actually good news — provided you read the signal correctly. A wrong answer is not evidence that AI isn't ready; it's evidence that the data isn't ready, and you've just acquired a tool that finds the unready parts faster than any manual review.

Four conditions, in order

1. Every entity exists exactly once

Customers, suppliers, products, employees — one record each. It sounds obvious, and almost no organisation achieves it without deliberate effort.

The work: pick a natural key (tax ID for legal entities, SKU for goods), merge duplicates, and close the path that creates new ones — usually a free-text field that should have been a picker.

That last point matters more than the first two. Merge once and leave the free-text field, and you'll be merging again in six months.

2. Every number has exactly one definition

Is "revenue" before or after discount? Recognised at invoice or at payment? Does "stock on hand" include goods in transit?

A healthy organisation does not need everyone to agree on a single definition — it needs every definition in use to be named and written down. "Recognised revenue" and "collected revenue" coexisting is normal; two people using the word "revenue" for different things without saying so is the problem.

With AI this moves from nice-to-have to mandatory. The assistant needs to know which column answers which question, in what unit, in what currency. Without that semantic layer it picks the column whose name sounds most plausible — and column names are the most misleading thing in any database.

3. Every dataset has an owner

Not "IT owns the database". Rather: who decides what the product catalogue looks like? Who approves a new supplier? Who rules that this item's base unit is "piece" and not "carton"?

Without owners, data drifts. With owners, each change gets judged — and the rules in points 1 and 2 finally have someone defending them.

4. Data lives in one place, or is genuinely joined

This is the condition that makes AI projects expensive.

If CRM is in one system, accounting in another and inventory in a spreadsheet, then "how profitable is this customer?" requires three manual joins. An assistant can answer it only if somebody has already built those joins — and keeps them from breaking.

A unified platform solves this structurally. Separate systems can solve it too, but you pay in continuous synchronisation, and every schema change in one system silently breaks a join in another.

Where to start

Don't start with a whole-estate data cleanup. Those run for months, show nothing, and get cancelled.

Start from the other end: pick one question leadership actually asks every week. "Which customers are about to churn?" "Which orders are stuck?" "What did we really earn this month?"

Then fix exactly the data that question needs. One question typically touches three to five tables, not the whole system. You get a result in weeks, and more importantly you get a live example to convince the rest of the organisation.

Then take the next question.

How you know you're ready

Not "the data is clean" — data is never entirely clean. The real signals are:

• Ask the same question twice, get the same number.
• Every number traces back to a source document.
• When someone edits data, you know who and why.
• When two reports disagree, you can explain the difference — instead of arguing over which is right.

Hit those four and AI becomes a thin layer over solid ground. Miss them and AI is just an expensive way to discover that you missed them.