Build the warehouse before you buy the model

Most requests that arrive as "we want AI" are, on inspection, requests for numbers people can agree on. The fastest route to a working AI feature usually runs through six weeks of unglamorous data work first.

The tell

Ask three people in the business for last quarter's revenue and see whether the answers match. If they do not, the disagreement is not a reporting problem — it is a definitions problem, and no model will resolve it. An assistant reading from disagreeing sources will confidently produce a fourth answer.

What we check before agreeing to an AI engagement

1. 01Is there a single system that is authoritative for each entity — customer, order, product — or several that partially overlap?
2. 02Are the key metrics defined once somewhere, or re-implemented in each report?
3. 03Can the data be read on a schedule without someone exporting a file by hand?
4. 04Is there a test that fails when the data is wrong, or does a human notice eventually?
5. 05Do permissions travel with the data, or are they applied only at the reporting layer?

Three or more "no" answers and we recommend a data engagement first. It is a less exciting proposal and it is usually the one that makes the AI work possible.

What that actually involves

Less than people fear. A warehouse for a mid-sized business is not a year-long programme. It is a managed database, pipelines that pull from the systems you already run, a modelling layer where each metric is defined exactly once, and tests that fail loudly when a number goes strange.

The modelling layer is the part that matters. Once "active customer" is defined in one place, every dashboard, every report and every assistant inherits that definition. That is what makes the answer the same whoever asks.

The warehouse is not a prerequisite because it is fashionable. It is a prerequisite because an assistant is only ever as trustworthy as the numbers underneath it.

The sequence we recommend

Warehouse and definitions first. Then reporting, so people can see the numbers and argue about them while they are cheap to change. Then, once the definitions have survived contact with real users, the assistant on top. Teams that invert this order spend the same money and arrive later.