Monitor your agent and choose the next improvement

Analytics helps you find questions where your AI agent needs better information. Use metrics to select a correction and review what happened in the conversation yourself. A high confidence score or meeting a target does not prove an answer is correct.

Check advanced analytics availability in the tier table below and in your workspace. Select the correct workspace and agent and open Analytics. You need permission to view the agent’s data. Usage and message credits are a separate task: see usage and limits .

The Coverage period may be selected automatically based on traffic volume. Check the displayed period: it may differ from the Resolution view’s period. Compare results only when you know which time range and conversations they describe.

1. Choose a recurring problem

   Open Coverage or Response quality. Find an unanswered or low-confidence question relevant to the agent’s task. Read the original conversation when needed.

2. Write a verified answer

   Select Create answer for the question. Review the prefilled question and write an answer based on current information. Do not copy customers’ personal data into the knowledge base. Save the answer to the knowledge base.

3. Publish the information and test

   Open Data sources and select Update Knowledge Base. Wait for a successful update, then try the question in a new test conversation. Also check a previously successful question on the same topic.

A correction saved to the knowledge base starts as a draft. It does not change messages already sent or make earlier analytics proof of a successful answer.

If delivery-time questions go unanswered, read a few sample conversations and check your delivery terms. Correct the missing information in the source, publish the knowledge base and repeat your test questions. Later, review new conversations on the same topic over a comparable period. This is an example workflow, not a promise that a metric will improve.

• Check the agent, selected view and period. A new agent may not have data yet.
• An empty result or dash does not indicate an error-free service; the data needed for the metric may be missing.
• If a confidence score disagrees with your review, correct the actual error in sources or instructions.
• If a correction does not affect the test, check the knowledge update and conflicting sources on the same topic.

Agree on a regular team review: one finding, one focused correction and a repeatable test. See also conversation review .

Also check availability and any applicable add-on in your workspace. This table gives minimum tiers for these features; it is not a workspace usage limit or a list of every integration.