Copilot Studio Friction
Production problem — but no logs anywhere
Last verified
Details & related
Assessment
- Confidence
- Confirmed officialConfirmed by Microsoft documentation or an official statement.
- Severity
- Degrading
- Typical time lost
- Hours
Identification
Verification & changes
Verified
Doc check by human: Official analytics, telemetry and transcript documentation plus CAT-team blog re-read during seed migration; latency, retention and App Insights gaps 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).
Are you in the right place?
- A user reports a problem in your published agent, and you want to see what happened.
- The analytics dashboard shows nothing useful, or nothing recent.
- You never connected Application Insights, Azure's logging service.
- You can't find the reported conversation anywhere.
- Your own test pane sessions never appear in analytics.
Many silent failures lead here: before you can diagnose them, you need somewhere to look. If the agent silently returns no answer at all → see Agent gives no answer and no error message. If transcripts exist but arrive late, incomplete, or redacted → (record planned). If the test pane itself goes silent after the first message → (record planned).
What's happening
Copilot Studio records nothing deep by default. The analytics dashboard shows session-level data with up to one hour of latency (documented as of 2026-07-08). It never shows test pane activity. Error- and event-level telemetry only exists if you connected Application Insights before the problem happened. Conversation transcripts land in Dataverse, the Power Platform database, on their own schedule. So the answer to "where are the logs?" is: three partial sources, each with its own latency and gaps. It's like asking for the security camera footage and learning the cameras were never plugged in. Logging starts when you wire it up; the past stays dark.
For technicians
App Insights needs an Azure subscription, a connection string, and KQL for queries. Even then, coverage is partial (documented as of 2026-07-08). App Insights never contains knowledge-chunk details, intent confidence, the orchestration plan, or session outcomes. Knowledge-chunk details means: which sources were searched and which chunks came back. The Dataverse transcripts are the detailed source for those. Transcripts have their own constraints (documented as of 2026-07-08). They are created after 30 minutes of inactivity and deleted by a default 30-day job. Developer environments store no transcripts at all. SharePoint-based answers appear in transcripts as REDACTED.
How to fix it
Solution 1
Official workaroundConnect Application Insights now, for the next incident
This cannot recover the past incident; it makes the next one visible.
- 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 reported problem once, in the real channel.
- Search the telemetry for the time window of your reproduction.
✅ You should now see: telemetry events for your reproduction, including errors and topic triggers.
Solution 2
Official workaroundPull the Dataverse transcript of the reported conversation
- Wait at least 30 minutes after the conversation ended. Transcripts are created after 30 minutes of inactivity (documented as of 2026-07-08).
- Open the transcripts for your agent in Copilot Studio.
- Locate the conversation by time and channel.
- Read the transcript detail — it holds the plan, knowledge detail, and outcomes that App Insights lacks.
✅ You should now see: the full conversation, including which knowledge was searched.
Developer environments store no transcripts (documented as of 2026-07-08). The default job deletes transcripts after 30 days.
Check that it worked
Run a short test conversation in the published channel, not the test pane. Expected: within an hour, the session appears in analytics. Your Application Insights resource shows matching telemetry events. After 30 minutes of inactivity, a transcript of the conversation exists.
If it didn't work
- App Insights shows nothing. Check you connected the agent in the right environment. Telemetry only covers failures after the connection was made.
- No transcript exists. Developer environments store no transcripts, and the 30-day job may have deleted older ones.
- The dashboard looks empty. Analytics data can lag up to one hour; retest later.
- You reproduced in the test pane. Analytics never shows test pane activity; reproduce in a published channel. For channel oddities, see Teams answers with an old version of your agent.
Prevent it next time
- Make Application Insights wiring a go-live gate: no production users before telemetry exists.
- Choose your environment deliberately — the environment choice pattern; developer environments cost you your transcripts.
- Decide transcript retention before launch; the default 30-day deletion is a decision made by omission.
Evidence
Official documentationlearn.microsoft.com
Deep telemetry requires connecting Azure Application Insights via a connection string; without it there is no error- or event-level logging.
Official documentationlearn.microsoft.com
Native analytics offers session-level views and shows data only up to an hour after a session ends.
Official documentationlearn.microsoft.com
Transcripts are created as Dataverse records after 30 minutes of inactivity and deleted after 30 days by a default bulk-delete job.
Microsoft statementmicrosoft.github.io
The Microsoft CAT team confirms knowledge-source details, intent confidence, the orchestration plan, and session outcomes are absent from App Insights; Dataverse transcripts are the detailed source.