Analytics overview
The Analytics page shows how well your agent answers and what it brings in, from real customer conversations with the published agent. Use it to see whether the agent resolves conversations on its own, what customers ask about, which conversations went wrong, and what the agent costs per conversation and per lead.
Open an agent and click Analytics in the sidebar. Everyone in the workspace can read it; editors and above change its settings.
What counts
- Only live customer conversations. Test conversations, Copilot runs and automations' hidden conversations are never counted.
- Only conversations where the customer wrote something.
- A conversation is counted once it ends: when someone resolves it, or after it has been quiet for a while. Until then, the page tells you it is waiting (see Waiting conversations).
- Days are your workspace's days, in its time zone.
Period and channel
The page header has the filters, plus Settings and Copilot.
| Filter | Options | Default |
|---|---|---|
| Period | 7 days, 30 days, 90 days | 30 days |
| Channel | All channels, Web widget, Telegram, API | All channels |
The period ends today and includes today. Each card compares with the previous period of the same length: for 30 days, the 30 days before those. The filters are kept in the page address, so you can share or bookmark a view.
Cards
Six cards sit at the top. Each shows its value and the change "vs previous N days".
| Card | How it's computed |
|---|---|
| Conversations | Customer conversations that started in the period (counted on the day of the customer's first message). |
| Resolved without a person | Conversations that ended without a person ÷ conversations that ended. "Ended" means resolved, or quiet for 24 hours. A conversation in which the agent asked for a person, or a teammate took over, never counts as resolved without a person, even after it is reopened. |
| Satisfaction | 👍 ÷ (👍 + 👎) from customers' ratings. Under it: AI says answered, the share of evaluated conversations the AI judged answered. |
| Leads captured | Records created in the period's conversations (not test, not deleted), across all tables. Hover for the count per table. Records from automations or imports don't count here. |
| Time saved | Conversations resolved without a person × Minutes per conversation (a setting, default 4). Shown in minutes under 2 hours, then in hours. |
| Cost | The model cost of the period's conversations, in dollars, including replies, the records fill and the AI evaluation. Under it: cost per conversation (cost ÷ conversations), cost per lead (cost ÷ leads, when the published agent has tables) and the evaluation's share ("Evaluation: …"). |
How the change is shown:
- For rates (resolved, satisfaction), the change is in percentage points: +5 pts.
- For counts and cost, the change is in percent: +12%.
- No change is shown when the previous period had nothing to compare with.
- For Cost, a rise is shown as bad news and a drop as good news.
A rate shows — when there is nothing to divide by, for example no ratings yet.
The Cost card helps you compare periods and channels. What your workspace is charged is counted in credits: see Usage and Plans and credits.
Trend
Trend is a daily chart of the period: bars for Conversations and a line for Resolved without a person (the day's resolved ÷ ended). Hover a day to see "N conversations · N% resolved". The line has gaps on days when no conversation ended.
Topics
Topics lists what customers ask about, most frequent first, with each topic's number of conversations and its Resolved rate (with a bar).
- Each evaluated conversation gets one topic from the AI evaluation. The agent keeps a list of up to 50 topics. Similar names are merged.
- Topics appear only when evaluation is on. Until then: "No topics yet: they appear once conversations are evaluated."
- Up to 3 weak topics get a A skill for this would help button: topics with at least 3 conversations and a resolved rate below 60%, weakest first. The button opens the Copilot with a prepared prompt that asks it to read the analytics and the worst conversations on that topic and propose a skill. You press send. See Skills.
Problem conversations
Problem conversations lists up to 50 conversations of the period, newest first, that have at least one of these:
| Reason | Meaning |
|---|---|
| 👎 | The customer rated it down. |
| Made something up | The AI evaluation found a claim not backed by the agent's instructions, Knowledge or tools. |
| Not answered | The AI evaluation judged that the question wasn't answered. |
| Handed off | The agent asked for a person, or a teammate took over. |
Each row shows the reasons, the topic, the channel, how long ago, the AI summary of the conversation, and the customer's comment if they left one (personal data masked). Click a row to open the conversation in the Inbox.
Team
When other agents use this agent as a team member, a Team section appears at the bottom: "Questions other agents asked this one. Not part of the customer numbers above." Open team calls goes to the Inbox's Team calls tab.
| Card | How it's computed |
|---|---|
| Team calls | Calls other agents made to this agent in the period. |
| Average time | Average duration of a call, in seconds. |
| Cost | Model cost of the calls. |
| Asked for details | Share of calls where it asked the caller for a missing detail. |
| Errors | Share of calls that ended in an error. |
| Answered the brief (AI) | Share of evaluated calls the AI judged as answering the brief. Shown from 5 evaluated calls. |
| Made something up (AI) | Share of evaluated calls where the AI found something made up. Shown from 5 evaluated calls. |
| Caller conversations resolved | Of the callers' conversations it took part in that ended, the share resolved without a person, with their 👍 rate. |
Asked by lists the calling agents with their number of calls. See Agent teams.
How the numbers are made
- Signals for each conversation are rebuilt from scratch whenever it changes, so a reopened or re-rated conversation always counts as it is now.
- AI evaluation: when a conversation has been quiet for 30 minutes or is resolved, a small model reads it, with personal data hidden, and records whether the question was answered (yes, partly, no), whether something was made up (with evidence), the topic, the customer's mood and a short summary. A conversation that continues later is evaluated again.
- Daily totals are updated right after each conversation is counted, with a safety run every 15 minutes.
Waiting conversations
Recent conversations that aren't counted yet show a note at the top: "N recent conversations are not in the numbers yet. A conversation is counted once it is resolved or quiet for 30 minutes; this page updates by itself."
A new agent with no conversations shows "No conversations yet. Analytics fills in once customers talk to your published agent. Test conversations are not counted."
Customer ratings
The 👍/👎 for Satisfaction come from the customer. The last rating wins.
| Channel | How customers rate |
|---|---|
| Web widget | 👍/👎 after the agent's reply once the conversation is resolved or quiet for 2 minutes, with an optional comment (up to 500 characters). |
| Telegram | Buttons under "Was this conversation helpful?" when the conversation is resolved, and the bot's /feedback command. |
| API | POST /sessions/:id/feedback with {"rating": "up"} or {"rating": "down"}. See the API overview. |
Settings
Click Settings in the header. Viewers see the settings read-only: "Only editors can change these settings."
| Setting | What it does | Default |
|---|---|---|
| Evaluation | On (every conversation), 1 in 5, 1 in 10 or Off. "A small model reads finished conversations (personal data hidden) to judge answers and topics." | On |
| Minutes per conversation | "Time a person would spend on one conversation; used for Time saved." A whole number from 0 to 240. | 4 |
| Weekly summary → By email | Mondays 09:00, to the workspace owners. | On |
| Weekly summary → On Telegram | To the team chats chosen under Inbox → Approvals → Notify. | Off |
With sampling, the same conversation is always either in or out of the sample, so re-evaluating it never changes that. With evaluation off or sampled out, AI says answered, topics and the AI reasons in problem conversations cover only evaluated conversations.
Evaluation uses credits. Its cost shows on the Cost card ("Evaluation: …") and under Settings → Usage → Agents. See Usage.
Weekly summary
Every Monday at 09:00 in your workspace's time zone, each published agent that had conversations in the past week (Monday to Sunday) sends a summary:
- Conversations, Resolved without a person, Satisfaction and Leads captured, each with the change vs the week before,
- the top 3 topics with their volume and resolved rate,
- the 3 worst conversations with their reasons, summaries and links to the Inbox,
- a link to the Analytics page.
It goes by email to every active owner of the workspace, and, when On Telegram is on, to the team chats you chose under Inbox → Approvals → Notify. "Never to the chats automations notify," since customers can join those. Without team chats, nothing is sent on Telegram: "No team chats yet: the summary isn't sent on Telegram."
The summary is plain text in your workspace's language (English or Turkish). Feedback comments are never included. Each agent sends at most one summary per week. An agent with no conversations that week sends none.
Export
The Analytics page has no export. To work with the data elsewhere, export your records from Records, or read conversations through the API.
Ask the Copilot
The Copilot can read the same report. Ask "why is the agent doing badly?" or "what should I improve?": it reads the cards, topics and problem conversations, opens the worst ones and proposes a skill or an instruction fix. See Working with the Copilot.