How do I read agent analytics?

Where is it?

Open the “Analytics” tab in the flow editor. On the “Agents” page, “Conversations” in a row's three-dot menu takes you straight to this agent's conversation list. The conversations of all your agents are together on the “Conversations” page in the menu.

Every role can view analytics. Use the date range at the top to pick “All time” (the default), “Today”, “Yesterday”, “Last 7 days”, “Last 30 days”, “Last 90 days” or “Custom range”.

Next to it, the “Version” filter narrows the overview, conversations and evaluations to conversations held on the versions you choose. It appears once the agent is published. Opens, requests, actions and credits aren't split by version; while a version is selected, “Start rate” is hidden. Tests from “Preview” in the editor count when all versions are shown and are left out once you pick a version.

If nobody has talked to the agent yet, you see the “No one has talked to the agent yet” intro; once you share the link, data collects here.

What is counted?

Each time a visitor opens the agent counts as an “open”; when the first question arrives, it counts as a “start”.

“In progress” is a conversation with activity in the last 30 minutes; “Left midway” is one that has been idle for 30 minutes. “Completed” appears when the flow reaches its end or when the chat in the AI Assistant step ends (the visitor said they were done, or didn't write for 30 minutes).

Overview

At the top are the agent's headline numbers; the details are written under each number:

• “Conversation starts”: with the number of opens below.
• “Start rate”: what share of the visitors who opened the agent started a conversation. Shown only when no version is selected.
• “Completion rate”: with the number completed and, if your end steps have outcomes, the success rate below.
• “Median time”: the middle start-to-finish time of completed conversations; half took less, half took longer.
• “Avg. fit score”: if there are scored conversations in the range.
• “Frustrated visitors”: if sentiment analysis has run; the share of conversations whose frustration is “Moderate” or “High”. Frustration the agent noticed during the conversation counts too. Click the card to see those conversations.
• “Credits used”: if the agent used credits; with credits per conversation below.

The cards below are in four sections: “Results and quality”, “Where visitors get stuck”, “Traffic” and “Team and integrations”. Most cards without data don't appear; “Outcomes”, the two traffic charts and “Handed to a person” always do.

Results and quality

• “Outcomes”: conversations that ended successfully, unsuccessfully, or without an outcome set; conversations left midway (“Left midway”) and in progress (“In progress”) are shown here too. If your end steps have no outcome, owners and admins see a “Mark end steps” link.
• “How visitors felt”: how visitors felt at the end (“Very negative” … “Very positive”) and their frustration (“None” … “High”), each as a single bar split into colors; click a color or its name to see those conversations. Only conversations with sentiment analysis are counted.
• In store agents, the “Assisted sales · revenue” card shows the orders of visitors who talked to the agent.

Where visitors get stuck

• “Step funnel: where it was left midway”: how many visitors left at each step, and the median time spent on that step. Conversations still in progress are not included.
• “Answer distribution”: how often each option was picked in Selection steps.
• “Unanswered questions”: the 5 most-asked questions the agent couldn't find in the knowledge base. They add up until closed, regardless of the date range. Owners and admins can use “Answer in the knowledge base” to add the missing information.

Traffic

• The “Conversation starts over time” chart.
• The “Starts by hour and day” heatmap (the busiest slot is shown too). Times are in your browser's time zone.
• “Where starts come from”: pick “Source”, “Campaign”, “Country”, “Device” or “Language” at the top; each row shows the number of starts and the completion rate. Add ?utm_source=instagram&utm_campaign=… to the link and the source is broken down here; Meta, Google and TikTok append their own click id, so the platform is usually known even without utm.

Team and integrations

• “Handed to a person”: tickets opened from the agent.
• “Integration calls”: calls the agent made to external systems and their success rate, including the store agent's calls. Open alerts appear at the top of the card.

Conversations

In the “Conversations” view each start is a row. The columns are “Visitor” (name or contact), “Status” (“Completed”, “Left midway”, “In progress”), “Left midway at”, “Location”, “Date”, “Duration” and “Credits”; agents with a file upload step also have “Uploaded files”. Next to the status, these badges may appear: “Confirmed” / “Cancelled”, “Frustration: High”, “Length limit”.

You can search by name, email, phone, answer or conversation number. Filter by status, source, location, “Left midway at” and “Confirmation”; a filter with only one option is hidden. When the range has analyzed conversations, “Sentiment” and “Frustration” filters appear too: they narrow conversations by how the visitor felt at the end (“Very negative” … “Very positive”) and by frustration level (“None”, “Low”, “Moderate”, “High”); conversations not yet analyzed don't match them. “Clear” resets them.

Click a row to open the conversation details. On the left is the “Conversation transcript”: the whole conversation, divided into steps. Messages where the visitor wrote negatively or showed frustration get a small badge under them. While “Show system events” is checked, consents given and tool and integration calls also appear in the transcript where they happened; uncheck it to see only the conversation.

Above the details are “Previous” / “Next”, “Copy link”, “Download conversation” (a text file) and the “More actions” menu: “Copy conversation number”, “Other conversations with this person” and, for owners and admins, “Personal data request”. There's no bulk export.

On the right there are three tabs:

• “Summary”: the list at the top shows “Overall score” if the conversation was evaluated, and the “Sentiment” and “Frustration” rows if sentiment analysis ran. Below them are the visitor's name and contact details, then “Duration”, “Source”, “Device” and “Location”; credits used and orders, if any, are here too.
• “Analysis”: a chart of the sentiment in the visitor's messages with a link such as “Worst moment: message 3” (click it to scroll the transcript to that message), and the AI evaluation's criteria with their results, reasoning and “Based on”. This tab appears only if the conversation has sentiment analysis or an evaluation.
• “Data”: “Answers”, the “Uploaded files” and “Consents given” if any, then “Tool and integration calls”.

The tab you chose stays open when you move to the previous or next conversation.

If a tool or an integration action ran during the conversation (a Trello card, a Slack message, your own tool), it is listed with its time under “Tool and integration calls” on the “Data” tab; while “Show system events” is checked it also appears in the transcript. Click a row to see the result, a link to the created record and the sent fields. Sent fields of integration actions are kept for 7 days; tool calls drop off once the retention period set under Tools ends.

Evaluations

The “Evaluations” view is always there. If there are scored conversations, it shows the “Evaluated” count, the average score, the score distribution and a list sorted from the highest score; otherwise it shows “No evaluations yet” or “Evaluation is off”. For details see AI evaluation.