Conversations and leads

Conversations

Inbox → Conversations lists everything the agents have been asked, newest first, with country, language, message count, sentiment and topics.

Open one and you get the whole transcript, plus what produced it:

• the sources cited for each answer and the retrieval score behind it
• any rating the visitor gave, and the reason they picked
• any files they attached
• any action the assistant ran, with its result
• the lead, if one came out of it

The retrieval score is the number worth learning to read. High score and a bad answer is a prompt or model problem. Low score and a confident answer is a content problem. Low score and an honest “I don’t know” is the system working.

A conversation belongs to a visitor’s browser and survives across visits, so a returning visitor continues where they stopped rather than starting over.

Note

Conversations are read-only records. You cannot reply into one — the widget is not a live-chat inbox. When a conversation needs a human, the assistant offers to open a ticket instead; see Email and ticketing.

Finding one

The conversation list filters by country, language, device, channel and date, by how many messages were exchanged, by whether a lead came out of it, by how the visitor rated it, and by the page the chat started on.

To search what was actually said, ask the assistant on Overview → Assistant: it reads the transcripts, which is the fast way to answer “has anyone ever asked about X”. The same search is available to your own tools through the MCP server as search_conversations.

Leads

When a visitor puts an email address or a phone number into the chat, the conversation is read once more and a lead is extracted:

Field	Where it comes from
name	Whatever they called themselves in the conversation
email	The address they gave
phone	The number they gave
interest	What they were actually asking about
country	The request’s geography, not something they typed
notes	The parts of the conversation a salesperson would want

Extraction runs again as the conversation continues, and fields are merged rather than overwritten — a later message adding a phone number does not erase the name from an earlier one.

Inbox → Leads lists them. Turn on lead alerts in Workspace → Notifications to have them emailed as they arrive, to whichever addresses you name.

Leads are records, not a pipeline

Engadin captures leads; it does not work them. There is no status, no owner and no follow-up state here on purpose — that belongs in your CRM. Send them across with a CRM integration and work them there.

Lead capture behaviour

Lead capture is a per-agent switch in Agent settings. When it is on, the assistant will ask for contact details at moments where that is a natural thing to do rather than at every opportunity. It can be configured to ask for a name alongside the address.

Turning it off does not stop extraction from working on details a visitor volunteers — it stops the assistant from asking.

Ratings and flags

Visitors can rate an answer up or down, and pick a reason: wrong, incomplete, outdated. Signed-in reviewers on the hosted chat page can additionally flag a message with a note, which is the fastest way to collect internal feedback during a rollout without anyone filing tickets.

Both show up on the conversation and in Overview → Analytics.