Answered

Dynamics 365 ERP MCP - Results are wrong

(1) Share

Report

Posted on by Adis

Hello everyone,

I am new to Copilot studio, so forgive me if I ve overseen obvious things. My main goal is to get the Dynamics 365 ERP MCP running and check what is possible and how good the results are.

I copied the Instructions from MS Learn.

Build an agent with Dynamics 365 ERP MCP - Finance & Operations | Dynamics 365 | Microsoft Learn

The only thing I changed was this part. In learn its 25 rows, I changed it to 100.

"# Extraction Instructions
- A tool call response can include up to 100 rows of data as form state. Generate a warning if the form state contains 100 rows of data."

The model is GPT-5 - we are not allowed to use anthropic models yet.

The answer of 75 active projects is simply wrong. I checked the environments and legal entity, all good.

My question is, am I missing a configuration or is this due to the GPT-5 model as I cant test claude.

Thanks in advance, Adis

Categories:

Model context protocol

I have the same question (0)

All responses (4)

Answers (1)

Sort by

Suggested answer

André Arnaud de Cal... 629 on at

Like (1)

Report
Copy link

Link copied!

Hi Adis,

Nice to meet you on this part of the community as well. There is a limitation in the number of records on pages which is 25. Changing the instructions to 100 is not helping here.

What version of Dynamics 365 F&O are you using? With the introduction of 10.0.47, there are also dynamics data events which can do this task without the form navigation. Then the count should be correct.

Was this reply helpful? Yes No
Adis 18 on at

Like (2)

Report
Copy link

Link copied!

Hi @André,

Nice to meet you! I wasnt sure where should I post this question, however, since you are everywhere it doesnt really matter 😊

We are on 10.0.46 (10.0.2428.114).

The example with the counting was just one example. My question is more towards, does GPT-5 is worth using or is a Anthropic model the way to go. For me, GPT-5 doesnt work at all, however, I dont know if it is up to the model, or due to a configuration I am missing.

I asked to provide me a vendor name and look what the result was:

This is a very simple task, so I was expecting to get a correct result, i.e. vendor name.

Kind regards, Adis

If this helped, please mark it as "Verified" for others facing the same issue

Keep in mind that it is possible to mark more than one answer as verified

Was this reply helpful? Yes No
Verified answer

André Arnaud de Cal... 629 on at

Like (1)

Report
Copy link

Link copied!
Hi Adis,

As you are on 10.0.46 and you copied the latest instructions as example, this is confusing the MCP server. The first part is telling the agent that there are three tools to chose from, including the data tools. There is no support for that in 10.0.46.

You can use the previous version which I pasted below. That should give better results. From my testing, the Anthropic model gives way better results, but a lot of companies are not allowing this in their tenants. First, check out the results with GPT-5 after you changed the agent instructions.

# Role Act as an autonomous data retrieval agent for inventory queries from finance and operations apps. There are two types of tools for interacting with finance and operations apps: form tools and API tools. - Form tools enable interaction with finance and operations forms in the same way a user would through the UI. - API tools allow calling actions defined in custom X++ logic. # Objective Your objective is to respond to tasks provided by the user. First execute each step of the provided task workflow using your MCP tools. Check if you have achieved your objective after each tool call. If you have not achieved your objective then continue to execute the next step in the task workflow. # Form tool usage instructions - Typical flow for record creation operations is to find a menu item, open a form, click the New button, find and set values for relevant controls, and save the form. - You can use grid filtering to find relevant records for update, delete, or inquiry scenarios. - Do not ever ask for menu item types. The find_menu tool groups menu items by their type. # Form tool parameter filling instructions - Omit optional parameters if no value is provided as input. - For menu items, grid columns, controls, and tabs, use the object name (and not labels) when filling the object name in tool calls. - (lessThanDate(x:int)) is a valid value for a grid date column filter. # Extraction instructions - A tool call response can include up to 25 rows of data as form state. Generate a warning if the form state contains 25 rows of data. # Action tool usage instructions - Use the find_actions and invoke_action tools when prompted to use actions. Otherwise use form tools. - Omit optional parameters if no value is provided as input. # Reasoning instructions - Before each tool call, plan the action. - Do not stop reasoning until all tasks are complete or an error is observed that prevents continuation. - Do not stop reasoning to ask a user questions or ask for user input. - When answering questions about data DO NOT rely on your general knowledge. Use tools to find accurate and precise data.

Was this reply helpful? Yes No
Adis 18 on at

Like (1)

Report
Copy link

Link copied!

Thanks @André!

I will try it out. We are also not allowed to use the Anthropic models yet, but this will change. Meanwhile, I can use GPT-5 since my instructions where not correct. I have to dig more into this.

Thanks a lot for the clarification.

Kind regards, Adis

If this helped, please mark it as "Verified" for others facing the same issue

Keep in mind that it is possible to mark more than one answer as verified

Was this reply helpful? Yes No