Annotations

Flag good and bad agent responses, mark factual errors, and capture transcript and Knowledge gaps during review.

Use annotations to give per-turn feedback during conversation review. Annotations create a running log of what’s working and what isn’t, so you can track issues, share them with your PolyAI representative, and measure improvement over time.

Annotations are available on both voice and webchat conversations from the Conversation Review side panel.

Agent turn annotations

Hover over any agent turn in the transcript to reveal a four-icon feedback toolbar:

Icon	Annotation	What it captures
👍	Good response	The agent handled the turn well. Use this for spot-checking and to mark positive examples.
👎	Bad response	The agent’s response was poor. Opens a reason picker so you can categorize the failure.
⚠️	Wrong information	The agent gave a factually incorrect response. Use this alongside or instead of Bad response when the issue is content accuracy rather than response shape.
📋	Copy turn text	Copies the turn’s text to your clipboard — handy when sharing examples in Slack or tickets.

Bad response reasons

When you click 👎 a searchable reason picker opens. Pick the category that best matches the failure:

• Response too general — the answer was vague when it should have been specific.
• Response too specific — the answer over-committed to detail the agent shouldn’t have had.
• Incorrect handoff text — the agent transferred or routed with the wrong wording or to the wrong destination.
• Long response — the answer was too long, especially relevant for voice where verbosity hurts.
• Hallucination — the agent stated something it had no basis for.

Use the search field to filter the list when you know the category name.

Wrong information

Click ⚠️ to mark a turn as containing factually wrong content. Wrong information is a single toggle — there are no sub-reasons.

You can apply Wrong information in combination with Bad response — they describe different aspects (content correctness vs. response quality).

Caller turn annotations (voice only)

For voice conversations, caller turns can also be annotated to capture ASR and Knowledge issues:

• Wrong transcription — the ASR (automatic speech recognition) produced an incorrect transcript. Use this when the agent misheard the caller. Feeds back into ASR tuning and keyphrase boosting.
• Missing topic — the caller asked something the agent could not answer because no matching topic exists in the Knowledge area. Feeds the prioritization of new topics.

These annotations are available on the caller side of voice transcripts. Webchat sessions don’t have ASR or audio, so they don’t surface these types.

What happens after you annotate

• Annotations are saved to the conversation and visible to anyone with access to Conversation Review.
• Share annotations with your PolyAI representative to drive ASR accuracy improvements, Knowledge updates, or agent-behavior changes.
• Bad response and Wrong information annotations build a queue of agent quality issues to address.
• Missing topic annotations directly prioritize your Knowledge backlog.
• Over time, annotations create a log of recurring issues that helps you track improvement.

Best practices

• Annotate consistently — Review a sample of conversations regularly and annotate issues as you find them.
• Pick the most specific reason — “Hallucination” is more actionable than a generic Bad response without a reason.
• Combine when useful — A turn can be both Bad response: Response too general and Wrong information if both apply.
• Act on annotations — Use them as input when updating your Knowledge area, ASR keyphrase boosting, transcript corrections, or agent behavior configuration.

Related pages

Conversation review

Browse and inspect conversations to find issues for annotation.

Conversation diagnosis

Debug agent behavior at each turn after reviewing transcripts.