Skip to main content
An insight is a finding you can trace to evidence and act on. Use existing conversation records and available dashboard measurements to identify a specific improvement, then keep the source and review method with the finding.

​
Choose evidence that exists

Public workspace usage and API-request reports currently return placeholder results. Do not use an empty report to conclude that there was no activity or that a cost is zero.

​
Produce a reviewable finding

1

Define the question

Choose one question, such as why users abandon a setup step. Record the agent, time interval, time zone, and selection rule. Follow pagination when gathering saved records.
2

Review the underlying records

Read full details when summaries omit messages. Separate what the user requested, what the agent did, and the observed outcome. Preserve unknown outcomes rather than counting them as failures or successes.
3

Record evidence and uncertainty

Keep safe conversation identifiers, the reviewed sample size, the matching cases, and any missing data. Separate a direct observation from a proposed explanation. A small sample can suggest a change without establishing a population-wide rate.
4

Assign and recheck the change

Give the proposed prompt, source, or configuration change an owner. Record the version used before and after it, then compare the same review criteria when new verified records become available.
Keep customer transcripts in access-controlled review systems. Share only the excerpts needed to explain the finding and follow the workspace’s privacy review.

​
Frequently asked questions

No. It describes an evidence review workflow using current reads. Automated topic discovery and analysis execution have separate implementation requirements.
Record the difference and account for it. A change in agent, time zone, sampling, or data availability can explain a changed metric without a change in agent quality.