Talk to Your Database
“Let analysts ask questions in plain English” was the 2023 pitch. The products that shipped were mostly text-to-SQL demos that fell over as soon as schemas got messy. We build the 2026 version: a decision agent that actually understands your schema, your business definitions, and its own limits.
What it is
A deployed agent, scoped to your warehouse (BigQuery, Snowflake, Databricks, Postgres — all supported over MCP), that turns questions into sourced answers. It’s not a SQL generator you hope for the best with.
What makes it work
Schema-aware, not schema-blind. We load your table definitions, column semantics, and business glossary into the agent’s context up front. It knows what revenue means in your data model. It doesn’t guess.
Citations are mandatory. Every numeric answer links back to the SQL that produced it and the row count it’s based on. Non-technical stakeholders get the plain-English answer; whoever wants to verify gets the query, one click away.
Refuses to fabricate. If the data isn’t there, the agent says “the schema doesn’t currently capture X” — it does not invent a plausible-looking number. This is the single most important behavior and the hardest to reliably get.
Read-only by default, write with approvals. Analytical questions run without ceremony. Anything that would modify data pauses for human approval, with the proposed change rendered in plain terms.
Regression-tested on real questions. Your team’s actual questions become the eval suite. When we update the model or schema, we re-run the suite before shipping.
Where it fits
- Revenue, product, marketing analytics — teams that want numbers without queueing behind data engineering.
- Ops and customer success — lookup-heavy work (“show me the last 10 tickets from this account”) that shouldn’t require a JIRA to data.
- Exec-level questions that combine 4 tables and want a paragraph summary, not a table.
Pair this with DataConnect for the MCP data layer and Insights for the broader decision-agent program.