Conversation analysis
Every conversation, scored against your criteria
AI evaluates every completed conversation against the criteria you write: as yes/no or with a score. The conversation gets an overall score from 0 to 100. Sentiment analysis shows in which message the visitor turned negative and where they got frustrated.
Reason and source
Every score comes with its reason
AI doesn't just give a score. For each criterion it writes what its decision rests on: an answer the visitor gave, a message in the conversation or a document they uploaded. Click the message and the conversation scrolls to that line.
- Every result states its confidence: confirmed by data, partly based on what the visitor said, or no data.
- Something the visitor only said in the chat counts for less than an answer they entered in a form or a document they uploaded.
- When there isn't enough data to decide, the result is "Unknown" and it doesn't count toward the overall score.
You decide what to measure
Your own criteria
You give each criterion a name and an instruction: "Is the budget over $2,500?" or "Was the issue resolved?". It is evaluated as yes/no or against a maximum score you set. You can add your scoring guide as a document or a link.
Reads documents too
If you like, the content of the documents and images the visitor uploaded in a file step is evaluated as well. What a document says counts for more than what was said in the chat.
Notices frustration right away
When a visitor gets annoyed or asks for a person, the agent briefly apologizes and offers to pass the matter to your team. Your team gets a notification with a link to the conversation.
In three steps
Write the criteria
On the agent's "Evaluation" tab, write what you want to measure.
Publish
The evaluation is part of the agent's version; conversations that start after you publish are scored.
Read the results
Conversations are listed from the highest score. For each you see the summary, the result per criterion and its reason.
TelegramSentiment analysis
How the visitor felt through the conversation
Every message the visitor writes to the agent is read for two things: sentiment (from very negative to very positive) and frustration (from none to high). The conversation detail shows a message-by-message chart and the worst moment, and negative messages get a small badge. Sentiment analysis is on in every agent by default.
Just for you
In your dashboard, in your language, up to date
Visitors don't see it
Scores, reasons and sentiment results appear only in your dashboard.
In your language
Reasons and the summary are written in the agent's main language, even if the visitor spoke another one.
Stays up to date
If a visitor comes back to a finished conversation with new information, it is evaluated again and the new score replaces the old one.
What teams do with conversation analysis
The score shows whom to call back first when screening, and which conversations need a look in support.
Online sales
Recommends products, adds to cart, answers order questions
Customer support
Answers questions 24/7, passes on what it can't solve
Appointments and bookings
Finds a free slot and puts it in your calendar
Applications and quotes
Collects details and documents for your team
In the help center
- How are conversations evaluated by AI?Define criteria on the Evaluation tab; each is answered yes/no or scored, and the overall score is 0–100. Sentiment analysis is switched on or off on the same tab.
- How do I read agent analytics?Start, completion and success rates, the step funnel, traffic sources, the conversation list and evaluations on one tab.
- How do I set up the AI Assistant step?Build an agent that answers free questions using a system prompt, a model, the knowledge base, tools and integrations; set handoffs and limits.
- How do I set up the File upload step?Collect PDFs, images or office files from visitors; optionally read the details inside, save them to variables and let the visitor confirm them.
Frequently asked questions
What does the evaluation read?
It reads the visitor's answers to the steps and what they wrote to the agent. If you like, it also reads the content of uploaded documents. What the agent wrote is read only for context and doesn't count as the visitor's information.
How is the overall score worked out?
It's the average of the criteria on a 0–100 scale: "Success" counts as 100 and "Failure" as 0; a scored criterion is divided by its maximum (4 out of 5 counts as 80). "Unknown" results don't count toward the average.
Does it use credits?
Yes. Every evaluation and sentiment analysis uses credits; an evaluation that reads documents uses more. The spending shows as its own line in the conversation's credit breakdown.
Do visitors see the result?
No. Only your workspace's members see the results; owners and admins change the settings.