When a company asks us for "an AI assistant that answers questions about our business," the honest first question is: can a person answer those questions today, reliably, from your data? If the answer involves three spreadsheets and a phone call to finance, an AI system will inherit the same confusion — just faster and more confidently.
Symptoms your data isn't ready
- Two dashboards show different revenue numbers for the same month.
- Key reports are rebuilt by hand in Excel each week.
- Nobody can say when a given table was last refreshed.
- Customer records are duplicated across your CRM, billing and support tools.
The minimum viable data platform
You don't need a multi-year program. For most mid-sized businesses, six to ten weeks gets you:
- A warehouse (BigQuery, Snowflake or Postgres) with raw copies of your core systems, loaded automatically.
- Modeled tables in dbt that define customers, orders and revenue once — with tests that fail loudly when the data breaks.
- A semantic layer so "active customer" means the same thing in every report and every AI prompt.
- Freshness monitoring, so stale data raises an alert instead of a wrong answer.
Why this makes AI cheaper
With clean, documented tables, retrieval becomes simpler, evaluation becomes possible (you finally have ground truth), and hallucinations drop because the model is working from one consistent source. The AI project that follows is typically shorter and cheaper than the one you'd have attempted first.
The unglamorous work is the fast route. It just doesn't look like it in the first month.