Recognise a signed-in visitor, and draft replies for your team

Two things that were missing on either side of a conversation: the agent did not know it had met this person before, and the colleague who took the conversation over answered from memory. Both change here.

Recognise someone your own site already knows#

If a visitor is logged in on your website when they open the chat, your site can tell us who they are and the agent will greet them by name and, if you want it to, remember something from last time.

Your site does that by signing a short-lived token for that visitor and handing it to the chat widget. Signing is the point: anything typed into a form is a claim, and a token you signed is a fact, so only the second can create a profile.

You need three things on the Recognised visitors page:

• the issuer your tokens carry,
• a signing secret, which we store and never show again,
• and the switch turned on.

Your developers need the article on end-user profiles in the API reference for the exact token shape. Nothing else about your chat changes: a visitor with no token is anonymous, exactly as before.

Three switches, not one#

Recognise signed-in visitors greets a returning person by name. On its own it remembers nothing.

Remember between conversations is the separate decision to let the agent carry something forward. Recognition without memory is a perfectly reasonable choice, and it is the default.

Personalise search result order lets what a person has asked about before lift matching results up the page. Order only. It can never add a result, remove one, or reach a document that person was not already allowed to see.

What it may remember, exactly#

At most twelve short items, each of one of four kinds: a preference they stated, an open item left unfinished, a topic they asked about, or a one-line summary of a past conversation. Nothing else, and no transcript.

Open any profile on the page and you see the whole list, item for item, in exactly the form the agent is handed it. That screen exists so you can answer "what does your assistant remember about me" without asking anybody to check a database.

Forgetting#

Every profile deletes itself after the number of days you set, counted from the last time it was used. The default is 90 days.

You can also erase one immediately: open it and press Forget this person. The profile, everything it remembers and the links to their conversations go, and it cannot be undone. Erasing a conversation also erases what that conversation taught the agent, so you do not have to remember to do both.

What we never receive#

No email address and no phone number keys a profile. Your visitor's real identifier is used once, to work out an opaque key, and then dropped: it is never stored and never written to a log. Two different projects that recognise the same person hold two unrelated profiles.

Draft replies for the person answering#

When a colleague takes over a chat, Suggest a reply on the live chat screen drafts an answer from your knowledge base and shows the passages behind it, the articles worth reading, and a summary of the conversation so far for somebody who just picked it up.

Nothing is sent. The draft goes into the composer, your colleague edits it if they want, and they press send exactly as before.

Two things on the screen are worth trusting:

• A draft with no sources is called out as unsupported. Nothing in your knowledge base backs it, and that is precisely the one to check before it goes out under a person's name.
• A reply that came from a draft is marked on the record, whether it was sent as-is or rewritten first. That is what keeps your quality numbers honest about what a machine wrote and what a person did. The chip in the composer shows the mark before you send, and you can clear it if you rewrote the sentence from scratch.

Your team can use the same drafts from another ticketing tool through the API, without a chat session.