Data Agent

Data Agent overview

The Data Agent answers business questions from your data. You ask something like “which customers placed the most orders this year?” and it works out the query, runs it against your bound database or lakehouse, and gives you a sortable grid, a chart, and the exact SQL it ran.

In the app: Agents › Data Agent

[screenshot]
The Data Agent page: the ontology and connection pickers along the top, the chat on the left, and the Results, Chart, and SQL tabs on the right.

Before you start

  • The Data Agent feature must be on for your organization. It’s included in the Trial, Business, and Enterprise plans. On other plans an organization admin turns it on under Organization Settings › Features; turning it on also turns on Bind to Database, which it depends on. When it’s off, the page shows Data Agent not enabled. See Feature flags.
  • At least one ontology must be bound to data, either to a database connection or to a Fabric lakehouse, with classes mapped to tables. If none is, the page shows No database-bound ontologies with a Go to Ontologies link. See Bind an ontology to a database or Bind an ontology to a lakehouse.
  • Anyone in the organization can ask questions. Saving, renaming, and deleting reports needs the Read/Write role.

How it works

The agent reads your ontology first: its classes, attributes, relationships, and descriptions. It matches the words in your question to that vocabulary, then uses the bindings to turn the matched concepts into tables, columns, and joins. The ontology carries the business meaning; the bindings say where the data lives.

That’s why the quality of answers depends on the ontology. Clear labels and good descriptions lead to better SQL. A class that isn’t bound, or a relationship that isn’t mapped to a join, can’t be queried, and the agent tells you so rather than guessing.

Where you can ask

  • The Data Agent page, under Agents in the sidebar. This is the full workspace.
  • The Home dashboard. Its Data Agent card lets you type a question or pick a suggested one, then hands it to the Data Agent page with the ontology already selected. See Your Home dashboard.
  • Your own AI assistant, through an ontology’s MCP server. See Expose an ontology over MCP.

What you get back

Each answer fills three tabs:

  • Results — the rows the query returned, which you can sort, filter, and download as CSV.
  • Chart — a chart the agent proposes when the data suits one, or one you build yourself.
  • SQL — the exact query the agent ran, to copy or download.

The agent also replies in the chat with a short summary. Above the grid, an Assumptions list shows anything the agent decided that you didn’t say, such as the time range or which status counts as “active”, with a Change button to redo the question another way.

Limits to know

  • It’s read-only. Every query is checked before it runs. Anything other than a single SELECT (or WITH … SELECT) is refused, as is a query that touches a table outside the ontology’s bindings. The agent can’t insert, update, delete, or change your schema.
  • Check its work. The agent writes SQL from your words, and like any AI it can misread what you meant. Read the SQL tab before you act on a number.
  • Results are capped. Each query fetches at most the organization’s row limit (10,000 rows unless an admin changes it), and the grid tells you when a result was cut off.
  • It works on one ontology and one connection at a time. It can’t reach another ontology, another database, or another organization’s data.
  • It uses your AI allowance. Each question counts toward your organization’s AI usage. See AI token allowance and usage.
  • A turn has a time limit. If a question takes too long, the agent stops and says The Data Agent took too long to answer. Try a narrower question.