Ask the Warehouse

Analysts were queuing days for the data team to write SQL, and the ones who wrote their own ran it against tables nobody had vetted, so two people brought two different revenue numbers to the same meeting. The status quo was a shared request backlog plus ad-hoc SQL pasted around in chat, and it could not stay because every answer carried a hidden reconciliation tax. I took it on with one rule learned the hard way: an LLM left alone with raw schema will invent a column with total confidence, so I designed the assistant around that failure mode instead of pretending it away.

• LLMs hallucinate schema they cannot see — saw it in the first week
• Read-only access, no production writes, ever
• Untrusted queries need row and cost limits after one runaway join
• Every prompt or model change must be regression-gated

Analysts answer their own questions now, and every answer shows the exact SQL underneath so trust is checked, not assumed. The data team stopped being a ticket queue and reviews the allowlist instead of rewriting the same queries. Queries the validator does not like are refused, not executed, and the eval gate means the system gets safer with each change instead of drifting into confident nonsense.

At 100x this becomes a governed analytics gateway: per-team semantic scopes, query-cost budgets, and an audit trail of every question the business ever asked the warehouse.