Copilot Studio Friction
Agent gives no answer and no error message
Last verified
Details & related
Assessment
- Confidence
- Confirmed officialConfirmed by Microsoft documentation or an official statement.
- Severity
- Blocking
- Typical time lost
- Hours
Identification
Verification & changes
Verified
Doc check by human: Official Responsible AI troubleshooting page, Q&A threads and MVP post re-read during seed migration; silent filtering still documented.
Change
Provisionally approved by the Product Owner; external LLM quality review pending.
Change
Initial record created from the seed dossier (migration wave 1 reference set).
Related
Are you in the right place?
- A user sends a message and the agent returns nothing — no answer, no error.
- Or the agent reports a
ContentFilterederror on harmless input, such as "hello" or a policy number. - The identical input often works on a second try.
- Nothing in the chat explains the missing answer.
- Automated or flow-triggered runs are hit especially often.
If Teams answers with outdated content instead → see Teams answers with an old version of your agent. If knowledge questions return nothing because permissions or consent are missing → see the silent knowledge record (record planned).
What's happening
Cause 1: Silent moderation. Copilot Studio checks every generative answer against Microsoft's Responsible AI content filter. When the filter blocks a response, the agent sends neither the answer nor a notice. The block is logged only if your agent sends telemetry to Application Insights, an Azure logging service. Without that connection, the failure is invisible everywhere. It works like a mail room that shreds suspicious letters without telling sender or recipient.
Cause 2: False positives, including on automation. The filter sometimes blocks harmless input. Documented cases include the policy number "PI-2824" and the word "hello". Retrying the identical input often succeeds. One Teams case ties the errors to active generative orchestration. Automated topic calls, for example from an agent flow, can look like an attack to the filter. Microsoft has not published which patterns trigger these false positives.
For technicians
Moderation events appear in telemetry only when the agent is connected to Azure Application Insights. Query the GenerativeAnswers events and filter where the result contains "Filtered" (official troubleshooting guidance). For the automation case, an MVP documented the rebuild agreed with the product team. Split into one agent per task and invoke each with "Please execute your instructions". The task description moves into each agent's instructions. Community support answers recommend switching affected paths to classic orchestration or disabling generative answers nodes.
How to fix it
Solution 1
Official workaroundMake the filtering visible with Application Insights
- Create an Application Insights resource in Azure, or pick an existing one.
- Copy the resource's connection string.
- In Copilot Studio, connect your agent to that resource in the agent settings.
- Reproduce the silent no-answer once.
- In Application Insights, search the GenerativeAnswers events for results containing "Filtered".
✅ You should now see: a logged moderation event at the time of the missing answer.
A hit confirms the filter blocked the answer — continue with the next solutions.
Solution 2
Community workaroundRetry, then reduce the generative surface on the affected path
- Send the identical input again.
✅ You should now see: in the documented cases, the identical input passes on the retry.
- If one guided path keeps failing, test it with classic orchestration instead of generative orchestration.
- Where the errors continue, disable the generative answers node on that path.
- Weigh the trade-off before you keep this state: these switches can remove the behavior your users wanted.
Solution 3
Community workaroundOne agent per automated task
Use this when autonomous triggers or agent-flow calls keep hitting the filter.
- Create one agent per automated task, instead of one agent with several topics.
- Move each task description from the calling prompt into that agent's instructions.
- Call each agent from your agent flow with the generic prompt "Please execute your instructions".
- Retest the automated runs.
✅ You should now see: the automated runs completing without ContentFiltered errors.
Check that it worked
If it didn't work
- Your change has not reached the channel yet. Publish and propagation delay is suspect number one. See Teams answers with an old version of your agent first.
- Application Insights shows nothing. Check that you connected the agent in the right environment to the right resource. Telemetry needs a reproduced case after the connection was made.
- The silence has a different cause. Missing permissions on knowledge sources cause the same silent symptom (record planned).
- Filtered again and again. Reduce the generative surface on that path, and consider the one-agent-per-task rebuild.
Prevent it next time
- Connect Application Insights before go-live; silent filtering is invisible without it.
- Test autonomous trigger architectures against the content filter early, before you build on topics.
- After every published change, verify behavior in the real channel, following the publish ritual.
Evidence
Official documentationlearn.microsoft.com
When Responsible AI moderates content, the agent returns no answer and no notice; the event is only logged via Application Insights telemetry.
Community threadlearn.microsoft.com
A maker reports ContentFiltered when a user enters the policy number 'PI-2824'; the identical input succeeds on retry.
Community threadlearn.microsoft.com
Teams users hit ContentFiltered even on harmless prompts like 'hello' when generative orchestration is active.
MVP blogsimondoy.com
Autonomous topic invocations were flagged as an attack; product-team guidance was one agent per task, invoked with 'Please execute your instructions'.