Copilot agents on your BI: four spots in your model and the RLS check before you start

9 July 2026

🎯 Hands-on⏱ 20 minutes✍️ 2 exercises

What you'll learn here

Prepare the four spots in your semantic model that a Data Agent reads
Run the RLS check before you switch an agent on
Know how to verify what the agent returns in the first weeks

Data Agents in Microsoft 365 Copilot went generally available this month. Business users can now query governed Fabric data sources, including Power BI semantic models, directly through their Copilot environment. That is a real step forward, and it asks something of your model in return. A Data Agent does not read your wiki, your report tooltips, or the post-it on your screen: it reads only what is in the semantic model itself. That makes four spots in that model decisive for what it gives back, and after those four there is one check you cannot skip before you switch it on.

The four spots in your model

Walk through the four spots on your own model and tick off what is in order.

1

Naming

Done

A table called dbo_FTS_INV_LN with columns like AMT_1, AMT_2, and DT is just as opaque to a Data Agent as it is to a colleague who started last week. Rename that table to Invoices and those columns to Amount excl VAT, Amount incl VAT, and Invoice date, and the agent has something to work with. The rule of thumb: open your field list in Power BI Desktop and ask yourself whether someone who has never seen this model would understand what is in front of them. If the answer is no, that is exactly where the agent will come up empty.

2

Synonyms

Done

In Power BI Desktop you can add synonyms per field and per measure, so a user who types “revenue” or “sales figure” lands at your Revenue measure rather than getting no result at all. This tends to get skipped, even though it is one of the few spots where you can capture your users’ natural language variation directly inside the model.

3

Measures

Done

A name that matches the question someone would actually ask, plus one unambiguous definition behind it. Revenue_v2_final_correct is an archaeological record, not a measure. If you cannot remember after a week which version was which, the agent cannot either, and it will hand your users the wrong number at exactly the moment they are starting to trust it.

4

The description field

Done

The field on tables, columns, and measures that spent years empty is now exactly what an agent reads when a name alone is not enough. You find it by right-clicking a table, measure, or column in Power BI Desktop, going to the General tab, and filling in the Description field. Pick three measures at random and check whether the field is empty: if it is, you know where to start.

📌

Takeaway

Four spots, one principle: what a new colleague cannot understand about your model, the agent cannot understand either. And what you have not configured in the model itself, the agent cannot respect.

The RLS check before you switch it on

Before you publish a Data Agent, run this test once.

✋

Try it yourself

Put a test user directly on your dataset, not through an app or a workspace role but directly on the dataset. Then ask through the agent a question that user should not be able to answer: “what is team X’s revenue”, “what does department Y cost”. If you get a number back, your row-level security is not correctly configured. Permissions that lived only as filters inside a report do not stop an agent querying the model directly.

This check takes ten minutes. The alternative, an incident report because the agent passed someone’s sales figures to people who should not have seen them, takes considerably more. And it is the conversation nobody enjoys having afterwards.

First weeks: verify before you share

Even with the four spots in order and the RLS check passing green, a Data Agent deserves verification in the first few weeks after you switch it on. Ask it things you already know the answer to, and see if it agrees. If it returns a number twice that you cannot quickly trace back through your model, you know where the model is not finished yet. This is the same check you run on a new report before you share it with the organisation, only now you do it with questions rather than with your eyes.

Getting Copilot agents on your BI is a practical task today: a session with your BI team on the four spots and the RLS check, then a few verification questions once it is live.

Want to run through those checks together, or find out how Copilot-ready your model already is? More at trainerbjorn.nl.