AI Data Analyst
Ask your data warehouse questions in plain English and get the answer, the chart, and the query behind it.
Questions in, answers with proof out
Most business questions wait in an analyst's queue. Text-to-SQL alone gives confident wrong answers when it doesn't know what “active customer” means in your company. That's why we define the metrics first.
Sample data: a fictional online store. In production the analyst sees only the tables and rows your role allows.
SELECT date_trunc('month', ordered_at) AS month,
SUM(net_total) AS revenue
FROM orders
WHERE ordered_at >= DATE '2026-07-01'
AND status = 'completed'
GROUP BY 1
ORDER BY 1;Revenue grew each month: $214.3k in September, up 17.5% on July. Uses the metric “net revenue” (completed orders, after discounts).
What you get
- Semantic layer. 20–40 core metrics defined once (what counts as revenue, an active customer, or churn), so every answer uses the same definitions.
- AI analyst. Turns questions into SQL against those definitions, checks the query before running it, and answers with a chart and the query itself.
- Permissions. A read-only database role and row-level rules: people only get answers from data they are allowed to see.
- Evaluation set. 100+ real questions from your team with correct answers. We measure accuracy on it before launch and after every change.
- Where people already work. Slack, Teams, or a web app, with a usage log of every question and answer.
- Warehouse, if you need one. We can set up PostgreSQL, BigQuery, Snowflake, or ClickHouse and load data from your CRM, billing, and product databases. Priced after the Sprint.
Eight weeks to the first answers
If your warehouse is in good shape, the first two steps go faster.
- Weeks 1–2
Questions and data
Collect the questions people ask, audit the tables, list the metrics.
- Weeks 3–4
Semantic layer
Metric definitions agreed with the people who own them.
- Weeks 5–6
Analyst and evaluation
Text-to-SQL, validation, and permissions, with accuracy measured on the evaluation set.
- Weeks 7–8
Rollout
Slack, Teams, or web access, training, usage review after two weeks.
What we need from you
- Read access to the warehouse or the source databases.
- A metrics owner who can confirm what each number should mean.
- 50 real questions people asked in the last few months.
Questions
How accurate is it?
We measure it on your evaluation set before launch and show you the number. Questions that fall outside the defined metrics are flagged for a person instead of guessed.
Can it change or delete data?
No. It connects through a read-only role, and every query is checked before it runs: no writes, row limits, timeouts.
What if we don't have a data warehouse?
We set one up as part of the project and load data from your main systems. The Readiness Sprint tells you what that involves and what it costs.
Does our data go to a cloud model?
Only the question, table descriptions, and the query results needed for the answer, or nothing at all if the model runs on your servers. You choose per project.
Bring the questions your team keeps asking.
On a 30-minute call we'll look at your data and tell you which questions an AI analyst can answer reliably. We reply within one business day.