Most organisations don’t have a shortage of data, they have sales data, finance data, operational data, customer data and increasingly sophisticated reporting platforms bringing it all together.
But how does someone get an answer from that data when the question they want to ask is not already covered by a report? Traditionally, the answer has been to find the right dashboard, export some data to Excel or ask an analyst but Microsoft Fabric Data Agents offer another option.
Instead of requiring users to understand the underlying data model, write SQL or navigate through multiple reports, they can ask questions using natural language and receive answers grounded in your organisational data.
What is a Fabric Data Agent?
A Fabric Data Agent provides a conversational way to interact with data through Microsoft Fabric. You can connect an agent to supported Fabric data sources including Lakehouses, Warehouses, Power BI semantic models and KQL databases, as well as other supported Fabric sources, and allow users to ask questions in natural language.
For example, instead of opening a sales report and filtering through several pages, a user might ask:
“Which five customers had the largest reduction in sales compared with last quarter?”
Or:
“Show me revenue by product category for the last six months.”
Behind the scenes, the Data Agent interprets the question, uses information about the available data and its configuration, and generates the appropriate query against the underlying source.
Depending on the source, this can include generating SQL, DAX or KQL. The important distinction is that the Large Language Model is not inventing an answer. It is being used to interpret the user’s question and retrieve an answer from organisational data.
Going beyond predefined reports
Dashboards are not going away. For monitoring established KPIs, distributing regular management information and presenting carefully designed analysis, Power BI reports remain an excellent way of consuming data. But dashboards are necessarily designed around questions that somebody anticipated in advance. A user might be able to see that revenue has fallen, for example, but their next questions could be:
- Which customers contributed most to the change?
- Was the decline concentrated in particular regions?
- Which products performed differently from the same period last year?
- What happened when we exclude a particular customer or business unit?
At that point, users often start changing filters, exporting data or asking the analytics team for help. A Data Agent introduces a more exploratory way of interacting with the same trusted data.
A user can ask an initial question, receive an answer and then continue the conversation with follow-up questions. That can make analytics significantly more accessible to people who understand the business but do not necessarily understand the underlying technical platform.
Why is this particularly interesting in Fabric?
Natural language querying isn’t new but what makes Fabric Data Agents particularly interesting is where they sit within the wider Microsoft data and AI ecosystem. Organisations building their analytical platform in Microsoft Fabric may already have governed business data, semantic models, security and business logic in place. A Data Agent can work with those existing Fabric data assets rather than requiring a separate analytical dataset purely for the conversational experience. Fabric Data Agents can be surfaced through wider Microsoft Copilot experiences. This includes integration with Microsoft 365 Copilot, currently in preview, and using a Fabric Data Agent as a tool within Copilot Studio.
That starts to change where users interact with analytics.
Rather than thinking:
“I need some data, so I need to open Power BI.”
The experience could increasingly become:
“I have a question, so I will ask Copilot.”
The appropriate Fabric Data Agent can then provide access to the governed business data needed to help answer that question.
What could this look like in practice?
Consider a sales organisation with a well-established Fabric platform and Power BI semantic model.
A regional director could ask:
“How are we performing against target this month?”
Then:
“Which regions are furthest behind?”
Followed by:
“Show me the products contributing most to the gap in the South East.”
The user does not need to know which tables contain the information, how those tables relate to one another or which DAX measures calculate performance against target. That complexity can remain within the data platform / semantic models. The same pattern could be applied across different areas of an organisation:
- Finance teams exploring revenue, cost and budget performance.
- Operations teams investigating service levels and operational KPIs.
- Sales teams understanding customers, products and pipeline performance.
- Executives asking follow-up questions from established management information.
- Support teams accessing operational data without raising another request with the analytics team.
Making trusted organisational data easier to interrogate.
Why governance, modelling, definitions and permissions are still critical
There is an important caveat. Adding AI does not fix a poorly designed data platform or semantic model. If two departments calculate revenue differently, adding a conversational interface does not suddenly resolve that disagreement. Likewise, unclear column names, inconsistent business definitions and badly structured models can make it harder for an agent to reliably interpret what users are asking. In fact, conversational analytics can make good data modelling and governance even more important. For an agent to be genuinely useful, organisations need to think about:
- Which data should the agent have access to?
- Are business definitions clear and consistent?
- Is the underlying semantic model designed appropriately?
- What instructions and example questions should be provided to the agent?
- Which users should be able to access particular data?
- How will responses be tested and validated?
Permissions are also fundamental. A Data Agent does not give users unrestricted access to organisational data. It can only access and return information that the user is authorised to see through the underlying data sources and security model. Microsoft provides ways to configure Data Agents with instructions, descriptions and example queries to give them additional business context and help them interpret organisational data more effectively. The technology is important, but the quality of the experience still depends heavily on the data foundation beneath it.
What Fabric Data Agents are not
It is also important to understand the current boundaries. A Fabric Data Agent is primarily designed to answer questions by retrieving and processing data from the sources it has been given. It is not automatically a replacement for advanced analytics, machine learning or detailed root cause analysis.
Ask: “Which products experienced the largest decline in sales?”
and an agent can query the available data to find out.
Ask: “Why did our sales decline?”
and the answer may require information and reasoning that does not exist within the underlying dataset.
Understanding that distinction is important when choosing the first use cases.
The strongest opportunities tend to involve well understood business data, clearly defined measures and questions that currently require users to manually interrogate reports or repeatedly go back to an analytics team.
Start with a question, not an agent
As with most AI projects, the best starting point is the business problem. Look for areas where people frequently ask the same kinds of follow-up questions about well governed data. If your analytics team regularly receives requests such as:
“Can you break this down by customer?”
“Can you show me the same numbers for last year?”
“Which locations are driving this change?”
you may already have a good candidate. Start with a focused domain, a known audience and trusted data. Then test the agent against the questions those users actually ask. That is much more likely to generate value than creating a general purpose agent and hoping people find something useful to do with it.
See Copilot in action
On 14 October at 10:00 AM, we’re joining Microsoft to host a virtual session exploring how organisations can bring their business data into Copilot using Microsoft Fabric. We will look at what Fabric Data Agents are, where they can add value and how they can be used alongside Microsoft Copilot. We will also demonstrate the technology in action and discuss some of the practical considerations around data, permissions, governance and selecting your first use case.
If your organisation is already investing in Microsoft Fabric or exploring how Microsoft Copilot can work with your own business data, the session should give you a practical view of what is possible and where to start. Join us live on 14 October, or register to receive the recording after the event. – Bring your business data into Copilot with Microsoft Fabric – Coeo