Power Platform Community Forum Thread Details

Hello Community,

I am currently working on an AI Agent in Microsoft Copilot Studio to automatically generate an Estimation Excel file based on a BRD, and I am facing an issue related to how knowledge files are being consumed by the agent.

Background / Scenario

I have created an AI Agent using Copilot Studio.
In the Knowledge tab, I have added two files:
- BRD document (Word) – contains business and functional requirements
- Reference Estimation Sheet (Excel) – contains standard estimation structure, effort breakup, and columns
In the Agent Overview instructions, I clearly mention that the AI Agent must:
- First refer and understand the reference estimation sheet
- Then analyze the BRD document
- Generate the estimation using the same format, structure, and estimation approach as the reference file
In Topics, I am using:
- Trigger + Create Generative Answer
- A Power Automate flow to create the final Estimation Excel file in a specific SharePoint library

Issue Observed

When I test or run the AI Agent:

The response seems to be generated only based on the BRD document
The reference estimation Excel file is not considered
As a result:
- Column structure differs
- Estimation breakup does not match the reference
- The output Excel file does not follow the standard estimation template

This happens even though:

The reference Excel file is available and marked as Ready in Knowledge
The instructions explicitly mention to use it as the base reference

Expected Behavior

The expected behavior is:

AI Agent should first read and understand the reference estimation Excel
Treat it as:
- A template
- Estimation logic guideline
- Standard structure reference
Then generate estimation details based on the BRD content while maintaining the reference format

Questions / Clarifications Needed

Is there any configuration in Copilot Studio to prioritize a specific knowledge source or file?
Are Excel files handled differently compared to Word/PDF in Generative Answers?
Should the reference Excel be converted or extracted (e.g., via Power Automate) before being used by the AI?
Is this behavior a known limitation of Copilot Studio AI Agents?
What is the recommended architecture for reference-based document generation in such scenarios?

Any suggestions, documented limitations, or workaround approaches would be greatly appreciated.

Thank you in advance for your support.

Categories:

Autonomous agents

When an agent references an Excel file from its knowledge, it is not a deterministic way to control output formatting. The excel file is flattened into text during ingestion, so rows, columns, formulas, and layout are not preserved. As a result, the agent may approximate the structure, but it cannot reliably follow a specific template or column order every time. The model may imitate structure, but it’s probabilistic, not enforced. If exact formatting is required, Excel should not be used as a knowledge source for this purpose.

If I’m understanding your goal correctly, I would think of this as two separate problems:

Understanding the requirements (from the BRD), and
Producing a correctly formatted Excel file (based on the Reference Estimation Sheet).

For problem #1, if a user is providing a BRD and expecting the agent to generate estimation details, neither the BRD nor the Reference Estimation Sheet should be added to your agent as Knowledge. The BRD is a dynamic input, so it should be provided by the user at the beginning of the interaction with the agent. The agent can then use a Prompt that clearly instructs what to extract and infer from the BRD, with the goal being a deterministic structured output (for example, a JSON object) that represents the estimation details.

For problem #2, producing a correctly formatted Excel file, the agent should then call an action or flow that takes that structured output and applies it to a predefined Excel template that aligns with your Reference Estimation Sheet. This is what will provide consistent structure, columns, and effort breakdown.

Matt Devaney has a great blog that walks through this pattern: Copilot Studio Autonomous Agent To Read & Write Excel Files.
While it covers both reading and writing Excel, your scenario would focus primarily on writing, since the agent has already produced a structured object from the BRD.

Hope this helps, let me know if I misinterpreted your use case or if you have any questions.