Ask the question. Get the number, the source, and the query.
Once a source is connected, anyone in the company can ask about it in their own words. Answers come from the governed tables, with the source named and the SQL available if you want to check it. Each answer teaches the memory, so the same kind of question is faster and cheaper next time.
Everyone who is currently waiting on a report.
Executives
Ask for the number in the meeting. Revenue by segment, claims trending, churn this quarter. Get it in a sentence, with the source.
Analysts
Skip the queue. Explore, pivot, and follow up in words, then export the query when it needs to go further.
Ops and finance teams
Question your own data without a BI licence or a dashboard sprint.
The more people ask, the less it costs.
Every question teaches the platform your schema, your language, and your definitions. Repeat and near-repeat questions are answered from memory and semantic cache instead of a fresh model call, and the cost view shows you by how much.
First ask
Question
The question goes to a model with the schema, not the data. Joins, filters, and aggregations resolve against your real tables, and the answer comes back with its source.
Similar ask
Learn
The semantic cache recognises it and serves the answer without a model call. Definitions, metrics, and the way your teams phrase things are captured once and reused everywhere.
Over time
Save
The memory holds what your tables mean and which questions people care about. Token spend per question falls, and AI FinOps shows you the saving.
Every answer respects who is asking.
The model sees the schema and the question. The query runs inside your perimeter against your tables. Results go to the person who asked, if their role allows the dataset.
No PII or PCI to a model
Only schema, meaning column names, types, and structure, is sent to a model. Raw record values never leave your perimeter.
Answers scoped to the asker
RBAC and business-unit isolation apply to every question. Two people can ask the same thing and correctly get different, entitled answers.
Every question audited
Who asked what, which data it touched, and what came back — each lands in the audit trail with a reason, ready for regulated review.
Ask it something about your data.
In a pilot, this is usually working by the afternoon of day one.