Agents that answer from your governed data — and can show their work.
Fabric data agents let people ask questions of the lakehouse in plain language. Built on governed data with the right guardrails, they give answers you can trace. Built on sprawl, they scale the sprawl. We build the first kind.
A working agent, scoped to a domain that matters
We don't deploy a chat box over everything. Each agent is scoped to one governed domain — finance, operations, sales — where the semantic model is solid and the answers can be verified.
Grounded in the lakehouse
The agent reads from your gold layer and semantic model — the same definitions your certified reports use. One version of the truth, whether a human or an agent asks.
Instructed with your context
Fiscal calendars, business rules, terminology — written into the agent's instructions so it interprets questions the way your analysts would.
Permissions it can't exceed
The agent answers with the asker's own data permissions. If a user can't see payroll, neither can their agent — enforced by the platform, not by a prompt.
Tested before trusted
A question bank with known-correct answers, run against the agent before rollout and after every change. Accuracy is measured, not assumed.
What keeps an agent honest
The agent knows which tables it may use and says "I don't know" outside them — rather than improvising an answer from the wrong data.
Every answer cites the datasets it drew from. A number without a source is a number nobody should act on — agents included.
Capacity limits and monitoring per agent, so usage growing is visible as a budget line — before it's a surprise on the invoice.
Wrong answers get logged, diagnosed — data gap, instruction gap, or model gap — and fixed at the source, so accuracy compounds.
The data foundation is what makes agents affordable.
AI cost is calculated in tokens, not tasks. An agent digging through sprawl burns context on every question; an agent reading a clean gold layer doesn't. The same governance that makes answers right also makes them cheap.
Pick the domain. We'll build the agent — and the tests that keep it honest.
If your lakehouse isn't ready yet, we'll tell you that first, and what to fix before an agent makes sense.