Ask
An AI analyst that cannot state a number it did not query, and shows you the query.
Status
Beta Ask is built and runs inside your project's region. It is being switched on region by region as each region's model processing is approved; until then the console shows it as unavailable.
Ask answers questions about your product in plain language: "Why did Android checkout conversion drop last week?", "Which acquisition channel retains best?". The difference from a chatbot on top of your data is that every number in an answer is bound to the query that produced it, and you can open that query.
How an answer is built
- The analyst plans, then calls tools: it searches your catalog and schema, reads metric definitions and
annotations, and runs queries through the same query engine as the console (or SQL, if your token has
sql.raw). At most 12 steps per question. - It writes prose with placeholders, never numbers:
{{e1.conversion|pct}}rather than "34%". The placeholders are filled from the real query results. - A grounding validator rejects any number or quantity word in the prose that is not bound to evidence, and an independent audit re-checks the finished answer.
- Confidence is computed by rules, never self-reported: sample size, data freshness, sampling, whether the events are verified in your catalog, whether a certified metric was used, and stated assumptions.
Each answer has an outcome: answered, insufficient_data, refused or unverified. Each claim links to its
evidence: the query, the compiled SQL, when it ran, the rows, ingest lag and sampling. Queries run with the
answer id attached, so every answer is traceable in the query log.
Writes are proposals
Ask can propose an insight or an alert. It never saves or changes anything by itself: you review and accept.
Residency and models
The analyst runs in your project's region, next to the data, and conversations are stored in that region.
Each region has an LLM policy: disabled, in_region (a model endpoint inside the region) or cross_border
(the organisation opts in to processing outside the region). Models are Anthropic's Claude family, pinned in
configuration.
Limits
Each organisation has a monthly AI budget that acts as a circuit breaker. Plans also include a number of AI questions per month (see Pricing); Scale and Enterprise can bring their own model key.
From the API
POST /ask on the region API streams Server-Sent Events (conversation, thinking, tool_call,
tool_result, prose, then one answer or error, then done). It needs the ai:read scope; see
Region API.