Using Conversation History for Questionnaire Assistance in Copilot Studio (Generative Answer)

(1) Share

Report

Posted on by CU14011854-1

I am building a Copilot Studio assistant to help users fill out questionnaires.

For this purpose, I store the answers to previous questions (per session) in a variable and pass this variable as history/context to each new Generative Answer prompt. My goal is for the model to reason over the previous answers and determine whether the current question can be answered based on information already provided earlier in the same session.

In other words, the agent should “think” about whether an answer can be derived from the previously given content instead of asking the user again.

However, this approach does not work as expected.

Observed Behavior

When I pass the conversation history via a variable into the Generative Answer node, the model consistently responds that the required information is not present in the data source.
This happens even though the information can clearly be derived from the stored history.
The same prompt works correctly only when I manually paste the entire history directly into the prompt text.
When the history is injected dynamically via a variable, the model behaves as if the content does not exist at all.
As a result, I never get a usable output when relying on variables.

Constraints and Considerations

Using a Knowledge Base is not an option:
- The agent still needs to reason over context, not just retrieve static facts.
- Each session has new, session-specific history, which cannot be pre-indexed.
The solution must work with per-session dynamic context, not global knowledge.

Core Issue

It appears that variables passed into Generative Answer prompts are not being treated as actual reasoning context. Instead, the model behaves as if only the literal prompt text (typed directly into the node) is considered.

This makes it impossible to build a reasoning-based questionnaire assistant that can infer answers from previously collected responses within the same session.

Categories:

Bot administration

I have the same question (0)

All responses (2)

Answers (0)

Sort by

Suggested answer

Giraldoj 923 Moderator on at

Like
a
(1)

Report
Copy link

Link copied!

Hi @CU14011854-1

To be honest, I’m not a big fan of fully generative answers. You don’t get granular control over the level of detail, and that makes it hard to guarantee the type of response you expect.

If you need more precise and consistent outputs, my recommendation is to use AI prompts:

https://learn.microsoft.com/en-us/ai-builder/use-a-custom-prompt-in-mcs

By creating and configuring prompts inside a tool that your agent or topic can call, you get much better control over the structure and depth of the answer. You can also keep the conversational flow exactly as you’re doing now: store the previous response in a variable, send it along with the next question as inputs, and clearly define the rules you want the agent to follow.

Hope it helps! if you want more detail on how to handle prompts feel free to ask.

Was this reply helpful? Yes No
CU14011854-1 5 on at

Like
a
(0)

Report
Copy link

Link copied!

Dear thanks for your response.

Using AI prompts was my second approach but currently I'm struggling to set it up because of access restrictions of the company, at least I think so. I get some strange data drop issues referencing, what I consider to be very strange, to another flow in which I load my questions from OneDrive. So it says that flow, which loads question must be in another data group. I get this error only if I add "AI Prompt" node to my second power automate flow and let the flow not only compose a prompt from current question and history but also try to answer the question in AI Reply.

Was this reply helpful? Yes No