BI Pixie Workload in Microsoft Fabric
Generally Available

AI Provider for AI Readiness in Microsoft Fabric

Scoring never uses AI. The text BI Pixie authors for you does: proposed descriptions, AI Instructions, folder names, the phrasing of benchmark questions, and the generic AI agent a benchmark can test. That text is written by an AI provider you choose, called server-side with BI Pixie's prompts. In the Workload the default is the BI Pixie data agent already in your workspace, which runs on your Fabric capacity and needs no API key.

Connect a Provider

On the AI Readiness page, the AI provider status row reads Not connected until one is set up. Open the Set up an AI provider dialog from the prompt beside it. The same dialog opens from Use as AI provider on the AI Assistants page once a data agent exists, so that you can install the agent and connect it in one place, and you can manage the provider at any time from the Account page, on the AI tab. Once connected, the status row names the provider, and the AI Assistants icon in the item sidebar carries a status light that is spoken as "AI provider connected".

The Set up an AI provider dialog showing the connected provider with options to replace or remove it

BI Pixie data agent (recommended)

Connecting takes one selection. If the agent is not yet installed, an in-context Install data agent action installs it into your workspace. Connecting triggers a single Execute data agents consent prompt, which you can grant yourself, with no tenant administrator required. If you decline, the provider is still saved, and a Grant access banner offers the permission again the first time it is needed. This option is not exclusive to the Workload: portal accounts with a designated Fabric workspace can use the same agent, as described on the portal's AI Provider page.

Bring your own AI provider

To use your own account instead, select Bring your own AI provider and connect Azure AI Foundry, OpenAI, or Anthropic. Azure AI Foundry is the recommended option: supply the endpoint and deployment name, and authenticate with Microsoft Entra ID (recommended) or an API key. OpenAI and Anthropic each take an API key. Keys are stored encrypted, and all calls are made server-side against your own account.

What Works Without a Provider

Everything that does not use AI works with no provider connected, and keeps working when a provider call fails: scoring, the full dimension breakdown, findings, usage grounding, assessment history, AI data schema curation, field defaults, and the first pass of folder proposals. The AI-authored items wait instead. Those cards show a note that an AI provider is required, with a connect action beside it.

A benchmark can still be built and run without a provider, using fixed question templates instead of natural phrasing; business-domain detection and question variations are unavailable, and each says so where it appears. The one benchmark option that requires a provider is the generic AI agent, because the provider plays the emulated agent and is therefore what answers the questions. The BI Pixie data agent cannot play that role, so testing a generic AI agent from the Workload needs a bring-your-own provider. See AI to Test.

What the Provider Sees

When BI Pixie writes text for you, it sends your provider the part of the semantic model definition the task needs: table, column and measure names, data types and format strings, existing descriptions, measure expressions, display folder names, and the semantic model's own AI Instructions. Where the reports built over that semantic model have Pixies, it also sends the report, page and visual titles people open and the values people filter to most, so the text it drafts uses the words your own people use.

When BI Pixie phrases benchmark questions, the provider receives field names, the shape of each question, and any filter value the question has to state, such as a product category. The correct answer is never sent at that stage, so the answer key cannot leak into the test. After the tested AI has answered, BI Pixie can send your provider the question, the answer that was recorded, and the correct value it computed, so your provider can give a second opinion on whether the two match.

Your provider never receives telemetry event records or end-user identifiers, and it never receives the contents of your tables. Fields you excluded in the AI data schema contribute no values and no descriptions at any stage, and the one thing your provider sees about an excluded field is its name, and only where your existing AI Instructions already mention it, so the reference can be removed. When BI Pixie drafts text it withholds more: no values at all from a semantic model that defines row-level security, and no values from a column whose name indicates an identifier or personal data. Privacy Policy Section 3.7 states the full list.

What's Next