AI Data Strategy Fundamentals

Most AI projects that stall do not stall on the model. They stall because nobody can say which system is authoritative for a given fact, and the answers inherit the disagreement.

A data strategy for AI is mostly four decisions, and none of them are about tooling.

One system of record per entity

For every important thing your business tracks, name the system that wins when two disagree. Customer, account, product, employee. Write it down.

This sounds bureaucratic until you build something that reads from three systems and the CRM says one thing while the core says another. Then it is the whole project. The retrieval is the easy part. The decision about which one is right is the work.

Definitions that survive a meeting

Ask your organisation how many customers you have and count the answers. The core returns open relationships. The CRM returns contact records including people who left in 2019. Digital banking returns enrolled logins. All defensible, none the same.

If "active customer" means four things, anything built on top of it means four things. Certify the handful of metrics that cause arguments and wire them into one place.

Quality with an owner

Automated checks on the tables that matter, running on every load, failing loudly. A named steward per domain, in the business rather than in IT alone. And a habit of fixing at the source, because patching the report fixes it once and fixing the intake form fixes it permanently.

Access that follows the person

A retrieval layer that can reach everything will answer using documents the person asking should not see. Permissions have to travel with the request. Retrofitting this after something is in production is a rebuild, not a change.

What order to do it in

Score honestly where you are on each of the four, then fix the lowest two before starting anything model-shaped. Skipping to the AI part on a weak foundation produces confident, expensive mistakes, and the confidence is the part that costs you.

None of this is glamorous and all of it is cheaper than the alternative, which is discovering it eighteen months in.