Assess a Data Agent for AI Readiness in Microsoft Fabric

When people ask a Fabric data agent a question, the answer they get is only as good as the data agent's configuration. That configuration keeps changing as people edit instructions, add tables and connect new sources. By assessing a data agent in BI Pixie, you get a score from 0 to 100, a list of the places where its configuration doesn't follow Microsoft's published guidance, and changes you can apply to the data agent.

Each score falls in one of three bands:

Ready to answer: 80 or higher. The data agent can answer reliably from your data.
Needs attention: 50 to 79. The data agent answers, but it is likely to get some answers wrong.
Not ready to answer: below 50. The data agent's answers can't be trusted yet.

Prerequisites

A BI Pixie item in a Microsoft Fabric workspace.
A published Fabric data agent.
The Contributor role or higher on the workspace that holds the data agent. Reading a data agent and changing it use the same Fabric permission.

Start an assessment

The Overview tab of the AI Readiness page, filtered to the BI Pixie Demo SaaS workspace, with the Your data agents card below the Your semantic models card. The Your data agents card lists two data agents, each with its AI Readiness score and change, its benchmark, the date it was last assessed, its number of assessments, and its benchmark runs.

In your BI Pixie item, select AI Readiness in the left sidebar.
In the header of the Your data agents card, select Start assessment.
Choose a Workspace. Only workspaces where you hold the Contributor role or higher are listed.
Choose one or more data agents, or select Select all to choose every data agent in the workspace.
Select Assess data agent.

When you choose one data agent, you see its result when the assessment finishes. When you choose several, they are assessed one at a time, and you can follow each one in a table:

In Readiness, State and Findings, you see the outcome.
In Status, you see Pending, Assessing, Done or Failed.
To open a finished assessment, select its row.

To assess data agents in another workspace, choose that workspace and select Assess data agent again.

Assess a data agent again

To assess a data agent again, select one of these:

New assessment, on the data agent's page or on its row in Your data agents.
Reassess, on an opened assessment.

You can run up to three assessments at once on each of your licenses, counting semantic models and data agents together. When three are already running, you see a note under the button you selected.

Note: Assessing a data agent is read-only.

BI Pixie writes nothing to the data agent until you apply a change.
No AI is used unless you select Ask your AI.
The assessment runs in BI Pixie's service, not on your Fabric capacity, and in your own Fabric region where regional processing applies to your account.

Find your data agents

You can find every data agent your account has assessed or started to assess, on any license, in the Your data agents card. The card is on the Overview, Assessments and Benchmarks tabs of AI Readiness, below Your semantic models. The columns you see depend on the tab:

Overview: the AI Readiness score and the benchmark score.
Assessments: the AI Readiness score, its trend and the number of optimizations.
Benchmarks: benchmarked data agents only.

In Your data agents:

To sort by a column, select its heading.
To open a data agent's page, select its row.
To assess a data agent whose row reads Not assessed yet, select its row. The workspace and the data agent are chosen for you. Nothing runs until you select Assess data agent.

Note: You can't assess the data agent BI Pixie installed for your account, so it isn't listed. A test data agent that BI Pixie created for one of your benchmarks is listed with the label Created by BI Pixie, so you can change it and assess it again.

Read the result

When you open an assessment, you see three parts, from top to bottom:

The score card: the score and its band, the date of the assessment with Reassess, the trend once there are two assessments, and one tile for each dimension under Dimensions. You see the costliest dimension first.
Optimize data agent: the changes BI Pixie can write to the data agent for you, after you review them.
Fix on your own: the findings you fix in Microsoft Fabric, grouped by dimension, each with where to make the change.

The AI Readiness score card of a data agent assessment, with the Data agent instructions dimension opened. That dimension reads Needs work and lost 39 points. It explains what instructions tell a data agent, quotes Microsoft's guidance on what belongs in data agent instructions with a Learn more link, and lists one Medium check, Instructions on the data agent, as Missing. The Published description, Runtime and Semantic model settings tiles read Passed, and the Score limit from semantic models tile reads At most 89 out of 100.

On the score card:

To see what a dimension measures, the Microsoft rule behind it and each check with its severity and result, select the dimension's name. You see the label Your AI on each check your AI answered.
To see the three bands, select What do these scores mean?
To see the points lost to each dimension, largest first, select Why this score?
To save space, collapse the card. You still see the score, the trend and every dimension's verdict.

To remove an assessment you no longer need, select Delete this assessment under the data agent's name.

Note: You get a score for the published configuration of the data agent, because that is what its answers come from. If you saved changes in Microsoft Fabric and haven't published them, you see a notice above the score card. To include them, publish the data agent and assess it again.

For every check and the Microsoft rules behind them, see Score dimensions.

Understand the score limit from semantic models

A data agent can't answer better than the semantic models it is grounded on. For that reason, a data agent's score is limited by the AI Readiness scores of those semantic models:

The limit starts at 70 and rises with the average score of the assessed semantic models, up to 100 when that average is 100.
With one semantic model that scored 90, the data agent can reach 97.
With one semantic model that scored 40, the data agent can reach 82.
Semantic models nobody has assessed are left out of the average. When none is assessed, there is no limit.

You can find the limit on the score card as the Score limit from semantic models dimension:

At most 97 out of 100 means the limit is above the score and costs nothing.
Held at 82 out of 100 means the limit is what holds the score down. You see the points it cost on the tile.

If a semantic model isn't assessed yet, you see its name on the score card, or the number of unassessed semantic models, and whether you have access to assess them. Assessing them can move the data agent's score either way.

Optimize the data agent

To raise the score, apply the changes under Optimize data agent. You can:

Rewrite the data agent's instructions.
Align its grounding with the AI data schema.
Write its published description.
Switch it to the preview runtime.

You review every change before BI Pixie writes it. After you apply a change, BI Pixie publishes the data agent and assesses it again, so the score stays current. For each change, step by step, see Optimize a data agent.

Check descriptions against real usage

If a source of the data agent is read by a Direct Lake semantic model, and reports over that semantic model have Pixies, BI Pixie also bases the assessment on what report viewers actually do:

BI Pixie checks that the columns report viewers use most have a description.
Under Fix on your own, you find the selected tables no report viewer uses. These aren't scored, because the data agent may still need them.

On the assessment, you see one of three usage states for each source:

The usage was read.
The usage is still being collected.
No tracked report reads the source through a semantic model yet.

For a data agent grounded only on semantic models, see the usage on each semantic model's own assessment. To see what a Pixie records, go to Permissions and Data Access.

Assess data agents grounded on other sources

You can assess a data agent grounded on any of these sources, alongside or instead of semantic models:

Lakehouse
Warehouse
KQL database
SQL database
Mirrored database

Each source is checked against Microsoft's guidance for that kind of source. To read the tables in some of these sources, BI Pixie needs a permission that you grant when you choose. For what is read from each source and how to grant access, see Data sources and access.

Assess on a schedule

On the Enterprise plan, you can assess a data agent on a schedule you choose, so the score stays current without starting each run yourself. To set up a schedule:

Open the data agent's page.
Select More actions > Assessment schedules.

You find the results of scheduled runs in the same history as the runs you start. BI Pixie never asks your AI during a scheduled run, because asking your AI needs you signed in. For the full procedure, see Schedules.

Review a data agent's history

To open a data agent's page, select its row in Your data agents. At the top, you see the data agent's workspace, current score and band, its trend once it has two assessments, and the date it was last assessed.

The page for one data agent, headed Marketing Campaigns 2 (B 2), with the New assessment, New benchmark and Open data agent buttons level with the heading. Under the heading the data agent reads BI Pixie Demo SaaS workspace, 85, Ready to answer, with a trend line and the date it was last assessed. Below it sit the Data sources card, which names the semantic model the data agent is grounded on and its AI Readiness score of 62, and the Assessments section, which lists the five newest of 19 records above Show all 19 assessments. The Benchmarks section, with 3 records, is collapsed underneath.

On the data agent's page:

To assess the data agent again and follow the progress on the page, select New assessment.
To benchmark the data agent, select New benchmark. To repeat a finished benchmark, select Run again, and to finish one that stopped, select Resume. If a benchmark can't start yet, you see the reason under the data agent's name.
To open the data agent in Microsoft Fabric, select Open data agent.
To schedule assessments (Enterprise plan), select More actions > Assessment schedules.
To remove every stored assessment of this data agent and take it off Your data agents, select More actions > Delete assessments. The data agent itself, its benchmark runs and its A/B tests stay.
To start an A/B test on a semantic model the data agent has answered a benchmark on, select More actions > Start A/B test on semantic model name. You see one entry for each such semantic model.

Lower on the page, you find four sections. In each heading, you see the count and the date of the newest record, and you can collapse each list:

Data sources: one card for each source, with its type and workspace.
- For a semantic model, you see its AI Readiness score, or Not assessed yet with Assess.
- For a lakehouse, warehouse, SQL database or mirrored database, you see how many of its tables the data agent uses.
- For a source BI Pixie couldn't open, you see why, and Grant access when a permission is missing.
Assessments: every stored assessment, newest first, with its date, score, change, findings and what was applied. To open a record, select it. You can also delete a single record.
Benchmarks: every benchmark run the data agent answered, with its score. To open a run's result, select it.
A/B tests: every A/B test in which the data agent answers one side, with the other side named.

The Assessments section of a data agent's page with every record shown, headed Assessments, 19 records, last 9/24/2026. Each row carries the date and time it ran, its score, the change against the record before it, how many sources and findings it covered, what BI Pixie found, such as Instructions or Recommended format, and what BI Pixie applied, such as AI Instructions added or AI Instructions updated, or No optimization applied.

Anyone on your BI Pixie account can read a stored assessment, including someone who has since lost write access to the workspace. If you turned on OneLake data residency, results are written to your own lakehouse instead of being kept by BI Pixie.

Assessments and your plan

Your first assessment of a data agent doesn't count toward your plan.
When you assess the data agent a second time, or apply its first change, the data agent becomes one of your plan's tracked items. BI Pixie tells you before the work starts.
Using a data agent in a benchmark or an A/B test never makes it a tracked item.

For the numbers on each plan, see Plan allowances.