Topics & QA
How the AI did on the conversations it handled alone, what customers asked about each week, sudden spikes in one kind of request, and a weekly insights email.
Insights (洞察) → Topics & QA (话题与质检) is for owners and admins. It needs the Growth or Scale plan, like quality review and the leaderboard (see Plan and billing). The page has four parts.
AI conversation QA
Every conversation the AI handled with no person involved is reviewed once it has been quiet for an hour. The review is done by jev, the same independent judge that checks each reply before it is sent, and gives a score from 0 to 100:
| Part | Points | Question |
|---|---|---|
| Answered | 60 | By the end, did the AI answer the question or resolve the request? |
| Tone | 25 | Did it sound right for this customer? |
| Handed over when it should | 15 | Did the customer need a person (asked for one, was clearly upset, or wanted something only staff can do) while the AI kept answering? |
The page shows, for the last 7 or 30 days:
- how many conversations were reviewed;
- the average score;
- how many were resolved;
- how many should have gone to a person.
Under Worth a look are the conversations scoring below 50 or missing a needed hand-over, lowest score first, with the customer's first question and what went wrong. Click one to open it in the inbox.
If the customer writes again later, the conversation is reviewed again. Conversations where a person took part are reviewed by the agents' quality review instead (see the leaderboard).
Spike alerts
purser watches each kind of request, meaning the topic the AI judged (such as damaged items or discount codes), overall and in each market. It sends an alert when, in the last 24 hours, that kind of request has at least 5 conversations and at least 3 times its usual day (the average of the 7 days before).
Alerts go to the notification center and by email to owners and admins, at most once a day per kind. They are meant to catch problems in bulk early, for example a batch of parcels arriving damaged in one country, or a discount code that stopped working.
Topic trends
Every Monday, the first question of each conversation from the week before is grouped by meaning. Each group of at least 3 questions becomes a topic, up to 8 a week. Each topic has:
- a short name;
- how many conversations it had;
- the change from the week before, or New if it wasn't there;
- the most common intent;
- up to three of the customers' own questions.
The page shows the last four weeks. A week needs at least 10 conversations for topics to be built.
Suggested knowledge
This links to the knowledge gaps in the knowledge base (see Knowledge base). Questions the AI couldn't answer or that were blocked are grouped and drafted as articles, with facts only from your team's answers and your policies.
Weekly insights email
Every Monday, owners and admins get one email summarising the week before:
- the AI conversation QA numbers;
- spike alerts;
- the top topics and how they changed;
- how many suggested articles are waiting.
Try it and releases
Chat with the AI on your draft configuration in Try it, publish to customers when you are happy, protect must-get-right questions with golden questions, and tell the AI how to handle specific situations with skills.
Inbox
How agents take over customers, reply in their own language to customers in any language, review AI drafts, use quick replies, and how ringing, SLAs and quiet customers are handled.