GPT-5 Chat in Copilot Studio often jumps topics in Teams after publish

(2) Share

Report

Posted on by AZ-01040635-0

Hi,

I’m seeing an issue with GPT-5 Chat model in Copilot Studio. After publishing and testing in Teams, the bot often jumps to unrelated topics during a conversation.

When I switch back to GPT-4.1, this issue mostly disappears and the conversation stays more stable. However, GPT-4.1’s answers are not as precise as GPT-5 Chat.

Has anyone else seen this?

Any ideas on how to reduce topic switching with GPT-5 Chat in Teams?

Thanks!

Categories:

Building Copilot Studio chatbots in Microsoft Teams

I have the same question (0)

All responses (3)

Answers (0)

Sort by

Suggested answer

Valantis 4,803 on at

Like (1)

Report
Copy link

Link copied!

Hi @AZ-01040635-0,

This is a known behavior difference between GPT-5 Chat and GPT-4.1 in generative orchestration.

GPT-5 is more capable but also more aggressive in selecting tools and topics based on context meaning it's more likely to jump to a topic that seems semantically relevant even if you didn't intend it. GPT-4.1 is more conservative.

The fix isn't to go back to GPT-4.1 it's to give GPT-5 better signals to work with. Two things confirmed from Microsoft docs that will make the biggest difference:

1. Topic descriptions this is the primary factor GPT-5 uses to decide which topic to trigger. If your topics have no description or vague descriptions, GPT-5 will guess based on content alone and jump to the wrong one.

Open each topic in Copilot Studio, click the topic name/description area, and write a specific description that says exactly when this topic should be used AND when it should NOT be used. Example: "Use this topic only when the user asks about invoice status. Do not use for general billing questions or payment methods."

2. Agent-level instructions add explicit guardrails in the agent's Instructions on the Overview page. Microsoft docs confirm you can use instructions like: "Only use the [topic name] topic for creating tickets. For other requests related to fixing issues, use the troubleshooting topic." This tells GPT-5 exactly which topic to route to for which scenario rather than leaving it to decide.

Also worth checking: are your topics using the Redirect node to call other topics? GPT-5 can sometimes follow those redirects in unexpected ways. If topics chain together with Redirect nodes, try making each topic more self-contained.

Ref: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/generative-mode-guidance

Ref: https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-generative-actions

Best regards,

Valantis

✅ If this helped solve your issue, please Accept as Solution so others can find it quickly.

❤️ If it didn’t fully solve it but was still useful, please click “Yes” on “Was this reply helpful?” or leave a Like :).

🏷️ For follow-ups @Valantis.

📝 https://valantisond365.com/

💼 LinkedIn

▶️ YouTube

Was this reply helpful? Yes No
Suggested answer

11manish 1,922 on at

Like (1)

Report
Copy link

Link copied!

GPT-5 Chat tends to introduce topic drift due to its more generative nature.

To stabilize it in Copilot Studio (especially in Teams), you need to enforce stricter instructions, use grounding, and control context more tightly.

A hybrid approach (structured topics + GPT for responses) works best.

Was this reply helpful? Yes No
Suggested answer

Haque 2,323 on at

Like (2)

Report
Copy link

Link copied!
Hi @AZ-01040635-0,

Valantis has covered most of the areas - just my two cents:

Use Context Window Management: Limit or carefully curate the conversation history sent to the model to reduce noise and off-topic context that might trigger topic shifts.

Tune Temperature and Sampling Parameters: Lower the temperature or adjust other generation parameters to reduce creativity and encourage more focused, deterministic answers.

Hybrid Model Use: Use GPT-5 Chat selectively for complex or precision-required queries, and fallback to GPT-4.1 for general or conversational queries to balance precision and stability.

Post-Processing and Filtering: Apply post-generation filters or rerankers to detect and suppress off-topic responses before presenting them to users.

I am sure some clues I tried to give. If these clues help to resolve the issue brought you by here, please don't forget to check the box Does this answer your question? At the same time, I am pretty sure you have liked the response!

Was this reply helpful? Yes No