Configure

Conversation Insights

Turn every finished conversation into a living knowledge map: what visitors actually ask your AI agent for, grouped by request type — and, for each ask, whether the agent could say yes or had to say no. Drill from the big picture down to the individual conversations, then export a summary for your stakeholders.

Turn it on

1. Open Analytics in the dashboard sidebar and switch to the Insights tab.
2. Pick the widget you want (if you have more than one).
3. Flip the Conversation Insights toggle on.
4. Click Analyze past conversations to classify your recent history (up to the last 90 days). The map fills in live as batches finish — new conversations are classified automatically as they end.

Read the map

• Bubble size is how often visitors ask for that thing.
• Bubble color runs green → red with the share of asks the agent had to decline. A big red bubble is your loudest unmet demand.
• Click a category to zoom into its topics; click a topic to see the actual requests — each with a one-line summary, the outcome, and a link to the full conversation.
• Before you select anything, the side panel shows your Biggest gaps — the topics visitors most often hear “no” about.

What the outcomes mean

• Fulfilled — the agent answered the ask or did the thing (booked it, quoted it, explained it).
• Declined — the agent said no, couldn't do it, or the thing isn't offered. Each declined ask carries a short reason (“out of stock”, “no same-day slots”) so the gaps are actionable.
• Partial — partly helped, or handed off to a human mid-request.
• Unresolved — the visitor left before an answer landed.

The headline stats show your fulfilled vs declined split over the selected window (7 / 30 / 90 days), plus coverage — how many of your engaged conversations have been classified so far.

Export the summary

• CSV — one row per topic (requests, fulfilled, declined, partial, unresolved, top decline reason) plus per-category subtotals. Drops straight into a spreadsheet.
• Report — a self-contained, print-ready HTML page with the headline numbers and the full category/topic breakdown. Forward it to anyone — it opens in any browser, no login needed.

How classification works

• When a conversation ends, a lightweight model reads the transcript and extracts up to three distinct requests, each labeled with a category, a topic, an outcome, and a one-sentence summary. It runs off the chat's critical path, so it never slows a reply.
• The classifier is shown your existing labels first, so the taxonomy converges — you get one “Pricing & Plans” bubble, not five near-duplicates.
• Greeting-only chats, spam, and internal test conversations are skipped.
• It works across channels — web chat, SMS, social, and phone-call transcripts all feed the same map.

Ask your conversations a question

The map above answers “what do visitors ask for?”. Sometimes you have a sharper question — “has anyone asked for a Sunday appointment recently?”, “is anyone confused about pricing?” — and you want the actual conversations, not a category.

Ask either kind. The box works out which you meant:

• “Did anyone ask about X?” — it finds the specific conversations and lists them for you to open.
• “What are the top 3 things people ask for?” — a ranking question, so it answers from the Insights data above instead: counted across every classified conversation in the window, not a sample. You get the ranking inline, plus which of them you are turning away and why.

Open Inbox and type the question into the box at the top of the list. Quincer searches in three ways at once and merges the results:

• The words visitors actually typed. Your question is expanded into the vocabulary a real customer would use first — ask about “appointments” and it also looks for book, booking, slot, come in.
• The topics already classified by Conversation Insights above.
• Meaning. This is what finds “can I pop in Saturday?” when you asked about the weekend and the word “weekend” never appears.

Then it reads the matching transcripts and tells you what it found, with each matching conversation listed underneath. Click any one to open it.

Read the small print under the answer

The line beneath the results is not decoration — it is what stops a confident answer being a misleading one:

• “3 matches” is always the number of conversations you can actually click. It is counted from the list, never taken from the AI’s summary.
• “searched 30 days” — the period is taken from your question. Say “recently” or give no period and it looks back 30 days; “in the last three months” widens it. 90 days is the maximum.
• “examined 40 of 402” — when more conversations matched than could be read, it reads the most relevant ones and says so. The answer will say “none of the ones I reviewed” rather than “nobody”, because a flat no over a partial search would be a claim about conversations nothing read.

Getting better answers

• Ask a question, not keywords. “Sunday” is a worse search than “has anyone asked about Sunday opening hours?” — the full question is what lets it judge which conversations genuinely answer you.
• One thing at a time. Two questions in one sentence get one blended set of search terms and a vaguer answer.
• It covers every channel. Web chat, voice calls, SMS, WhatsApp, Messenger, Instagram, email and Slack all write to the same conversation history, so one question searches all of them.
• Non-English conversations are searchable in English. Translated threads are matched on their English translation, so you do not need to know what language the visitor used.