AI Responses
Track and analyze how your AI chatbot performs and handles user questions
Understanding Response Categories
Every visitor message produces one AI turn. Each turn is grouped into one of three top-level buckets so you can read performance at a glance:
Successful
The AI answered from your knowledge base or a tool, connected the visitor to a human agent when they asked, or resolved the turn with an automated workflow.
Neutral
Simple interactions like greetings or small talk that don't require specific knowledge from your content.
Failed
The AI lacked enough information to answer, automatically handed off to a human, or hit a generation error.
The summary cards at the top of the page show each bucket's share of all turns — and because every turn is counted exactly once, the three percentages always add up to 100%.
Response Outcomes
Each row in the log carries a specific outcome badge that tells you exactly what happened. Use the outcome filter to drill into a single pattern.
The AI answered the question using context retrieved from your knowledge base or an API tool. (Counts as Successful)
The AI handled a basic interaction — like a greeting or simple question — without needing your content. (Counts as Neutral)
The visitor explicitly asked to speak with a human, and the AI handed off appropriately. (Counts as Successful)
The AI triggered an automated workflow — like booking an appointment or creating a ticket — that resolved the turn. (Counts as Successful)
The AI didn't have enough information in your knowledge base to answer the question. (Counts as Failed)
The AI had no relevant context — or an API tool failed — so it automatically handed the visitor off to a human. This points to a knowledge gap. (Counts as Failed)
Something went wrong during AI generation. (Counts as Failed)
Inspecting a Turn
Click the expand arrow on any row to open a full breakdown of how the AI produced that answer.
Conversation
The visitor's question and the AI's answer, shown exactly as they appeared in the chat — including any product carousel the AI added.
Details
A metadata strip summarizes the turn at a glance:
| Field | Meaning |
|---|---|
| Outcome | The result badge (see above). |
| Channel | Where the turn happened — e.g. web chat or voice. |
| Grounded | Whether the answer was based on content retrieved from your knowledge base. |
| Model | The AI model that generated the answer. |
| Tokens | Total tokens used for the turn. |
| Latency | How long the answer took to generate. |
Steps
A timeline of everything the AI did to reach its answer. Each step shows a status, and you can expand it for more detail:
- Knowledge search — the search query the AI ran against your knowledge base, plus the results it matched. Each result lists its source and a relevance score; expand it to see the exact text the AI used.
- API tool — an external API tool the AI called, with the data it sent and received.
- Workflow — an automated workflow the AI triggered.
How to Use This Data
Check Your Success Rate
Look at the percentage of successful turns to understand how well your AI is performing overall.
Identify Knowledge Gaps
Review Failed and Handoff (no context) turns to find questions your knowledge base doesn't cover yet.
Inspect the Steps
Expand a turn to see what the AI searched for and which content it retrieved — a fast way to spot when the wrong article is matching, or none at all.
Review Error Patterns
Look for recurring errors that might indicate a configuration issue.