Improve an incorrect answer
Turn a real customer example into a focused knowledge, instruction, or setup improvement.
When an answer is wrong, start with the conversation. Keep the customer's question, the answer they received, and the answer you expected. A specific example makes it easier to fix the right thing and check whether the fix worked.
If the customer is still waiting, help them first in Inbox. Improving the agent is a separate follow-up.
1. Identify what went wrong
| What you observed | Start by checking |
|---|---|
| An outdated or missing product fact | The source material and assigned knowledge release. |
| Correct facts, but the wrong tone or too much detail | Instructions, tone, and answer length. |
| The answer ignored the customer's plan, order, or account | The contact information and customer-context result for that conversation. |
| It promised an action that did not happen | The connected tool and its actual result. |
| It kept answering when a person should help | Handoff instructions, assignment rules, and any active workflow. |
Use these starting points alongside the surrounding messages. An earlier statement or missing clarification can help explain the answer.
For a voice call, compare the transcript and recording when available. A misunderstood spoken question needs a different fix from a correctly understood question answered with outdated facts.
2. Correct the facts at their source
Open the relevant knowledge collection. Correct or remove outdated content, fill the missing explanation, and resolve contradictions between sources.
For example, if the returns period changed from 14 to 30 days, update the policy source rather than adding “always say 30 days” to several prompts.
Publish the collection, wait for processing to finish, and apply Use latest on the intended agent's assignment when an update is available. For voice, also check that activation completed. Confirm each intended agent is using the updated release before retesting.
3. Adjust how the agent responds
For a behavior problem, edit System Prompt in General Settings for chat, or Behavior → Instructions for voice. Keep the change specific:
| Vague instruction | More useful instruction |
|---|---|
| “Be concise.” | “Answer the question in a short paragraph, then offer the next step if one is needed.” |
| “Never be wrong.” | “If the available knowledge does not confirm a policy, say what needs checking and offer human help.” |
| “Always solve the issue.” | “Explain the available options. Pass exceptions requiring approval to the team.” |
Remove conflicting instructions rather than stacking more rules on top of them. Save, then test. See Write effective instructions.
4. Check the connected behavior
If the answer depends on account information, inspect the customer-context result and available contact fields. Start a new conversation when testing a change to conversation-start context; an existing conversation keeps its original result.
If the answer depends on an action, check the tool's assignment, status, and returned result. Confirm the update in the connected service before repeating it.
For the wrong destination, review human handoff and the active workflow's Run Logs. Check the rule and information used by that run before changing the agent's prompt.
5. Repeat a small set of examples
Retest the original question after the relevant change has been published or activated. Then try:
- A different wording of the same question.
- A nearby question that should still receive the correct answer.
- A request with a key detail missing, where the agent should ask for clarification.
- A question outside the available knowledge, where it should admit the limit or offer human help.
- Another supported language or channel when the change affects it.
Change one cause at a time where practical. Record what changed and the results so the next person can tell whether the improvement held up.
6. Check the published experience
Repeat the important example through the customer-facing Widget, Page, or voice number. Confirm that you reached the intended agent and channel, then review the resulting conversation or call.
If the issue persists, share the example, agent, channel, approximate time, and changes already tried. Troubleshooting explains what to include.