Review conversations and improve responses
Find where your agent helped and where an answer needs correction, then verify the change with a new test.
Open copy menu
Conversations show what customers ask your AI agent and how it responds. Use these observations to improve sources and instructions. A confidence level alone does not prove that an answer is correct or incorrect.
Before you start
Section titled “Before you start”Select the correct workspace and agent in Oma Aihio and open Conversations. You need permission to view those conversations. Preview conversations appear on the Playground tab and do not represent real customer traffic.
Conversations may contain personal information. Do not copy it into new sources or support requests unless it is necessary for the task and you are allowed to process it.
Review customer and test conversations
Section titled “Review customer and test conversations”The agent’s Conversations view shows that agent’s conversations. The shared workspace conversation view, available from Pro, brings agents’ conversations together. Check which view you are using and whether the open conversation is a customer conversation or a test.
Tags, available from Pro, help classify conversations by topic or follow-up, for example. Agree their meaning with your team. You can also export an individual conversation as a text file; store any personal data it contains only where the task requires.
Identify what needs correcting
Section titled “Identify what needs correcting”- Open the conversation and read the messages preceding the question.
- Compare the answer with current source information.
- Decide whether information is missing or the issue concerns instructions or response style.
| Observation | Correct first |
|---|---|
| Outdated price or policy | The original data source. |
| Missing answer to a recurring question | A source or Q&A entry. |
| An overly long answer or unsuitable tone | The agent’s instructions. |
| An unclear customer question | Instructions for asking a clarifying question. |
Save a corrected answer
Section titled “Save a corrected answer”Improving an answer directly from a conversation requires Pro or higher. If the action is unavailable, correct the information directly in Data sources and update the knowledge base.
Open the answer correction
Select Improve on the answer you want to correct. In Improve answer, check the original question and write a factual response supported by your sources.
Save to the knowledge base
Select Save to knowledge base and wait for confirmation. The correction is saved as Q&A content. Also check that conflicting older information does not remain in the knowledge base.
Update the knowledge base and test
Open Data sources and publish the changes with Update knowledge base. Wait for completion and ask the same question in a new test conversation. The correction does not alter a response already sent to a customer.
If the correction does not help
Section titled “If the correction does not help”First check that the knowledge update finished. Look for conflicting sources on the same topic and make sure the instructions do not request different behavior. If the new answer is incorrect, edit that source and update the knowledge base again.
Continue checking previously successful questions too. A useful correction helps with the affected situation without making other answers worse.
Visitor feedback and data requests
Section titled “Visitor feedback and data requests”Visitor thumbs feedback helps identify replies to review. Read the question and source before drawing a conclusion: negative feedback alone does not establish a factual error, and positive feedback does not guarantee correctness. Receiving feedback does not itself publish a correction to the knowledge base.
The workspace-wide conversation view has a privacy action that removes a conversation from ordinary use. It marks the conversation as deleted; it does not promise immediate physical destruction of every copy. Verify the conversation and the reason for processing before confirming. If the customer’s request concerns broader deletion or export, follow the data-request process . Disappearing from the interface alone does not establish that the entire data request has been completed.