Satisfaction and usage

Satisfaction (CSAT)

Your satisfaction score — CSAT — comes from the widget’s own rating prompt. When a conversation ends, a visitor who took part in it is asked for a 0–5 star rating and an optional comment. Your dashboard turns those into a score.

Over the last 30 days you get the average, the spread across all six star values, the share who were happy (4–5), the count who were unhappy (0–2), the response rate, and the five most recent comments.

Three details that decide whether the number means anything:

• Zero is a real answer. A visitor who rates a conversation 0 said it was terrible. That is not the same as a conversation nobody rated, and the two are never mixed — unrated filters never swallow a zero.
• The response rate is measured against conversations that actually finished. A chat still open was never asked for a rating, so counting it as an unanswered survey would make your score look worse every time somebody leaves a tab open.
• The card only appears once something has been rated. An empty score reads like a bad one.

Every bar in the distribution links into the chat list filtered to that rating, so a number is a way in rather than a fact to admire. You can also filter the list to rated, unrated, happy or unhappy directly.

Automatic review

Every conversation that ends is also reviewed automatically, and the answers sit on the transcript:

• Was the issue resolved? Did the visitor get what they came for.
• Was the bot factually consistent? Did it tell the visitor anything its own tool results don’t support.
• Were your rules followed? Your preferred wording and bot instructions.
• Does a person need to follow up? An unkept promise, an unresolved need, an angry visitor, a request for a human nobody answered. These land in a review queue.
• Is this a documentation gap? The bot couldn’t answer and the missing information is something you could write.
• Sentiment at the close, and a predicted star rating.

The predicted rating matters because most conversations are never rated. It’s on the same 0–5 scale as the real one, and the transcript marks it as predicted so it’s never confused with something a visitor actually said.

Every answer comes with a confidence, and only a confident one creates work for you.

What your AI usage costs

Settings → Model calls lists every call your workspace has made — the automatic decisions and the bot’s own replies, successes and failures together.

Above the list is a chart. Pick a time span (24 hours, 7 days, 14 by default, 30, or 90), what to group by (what it was for, model, provider, outcome), and what to measure (calls, cost, tokens, or latency). Totals for the window sit alongside it: calls, tokens, spend, average latency, and how many failed.

Each row names what asked for the call rather than only which model served it — “Label an open chat”, “Grade an ended chat”, “Answer a visitor” — and links to what it was about, so a cost is traceable to a conversation. The model shown is the version that actually answered.

Filters live under their own column headings, and the gear in the corner hides columns you don’t care about. That choice is remembered in your browser.

Writing numbers your way

Counts throughout the app follow a personal setting at My settings → Profile → Number style: compact (1.3k), compact European (1,3k), in full (1,337), or in full European (1.337). It follows you across every workspace, and across the web app and the mobile app.