NearSync Help

WhatsApp

AI Assistant Settings Six controls that decide how much the assistant does before a person steps in.

The assistant can answer customers on WhatsApp before anyone sees the message. Six settings decide how far it goes.

AI handles first

The master switch. On, the assistant reads incoming messages and responds where it can. Off, everything goes straight to a person.

Start off. Watch the inbox for a week, note what people actually ask, then turn it on once your knowledge base covers those questions. An assistant switched on before it has anything to answer from produces handoffs with an unhelpful message in front of them.

Classifier confidence

How sure it must be before acting on what it thinks the message is about.

Raise it and more goes to people. Lower it and more gets handled automatically, including things it has misread.

Err high at first. A handoff costs a person a minute; a confidently wrong answer to a customer costs more than that.

Knowledge base match

How closely a question must match your material before the assistant answers from it.

Loose matching produces answers that are almost about the question, which reads worse than not answering. Tight matching produces more handoffs and fewer embarrassments.

Reply debounce

How long it waits before responding.

People send three messages in a row while thinking. Debounce stops the assistant answering the first before the third arrives, and then answering that too. A few seconds makes the difference between a conversation and an argument with a machine.

Escalation SLA

How long a handed-off conversation may wait before it is chased.

Set it to something your team will actually meet. An SLA nobody meets stops being a deadline within a fortnight.

Create a ticket on escalation

Whether a handoff becomes a tracked support ticket.

Turn this on if handoffs are things somebody must resolve. Leave it off if they are usually answered in one reply, or you will build a queue of tickets that close themselves.

The order to set them in

Knowledge base first, then a high confidence threshold, then on. Loosen from there once you can see what it gets right.

4Updated 28 July 2026

Did this answer your question?

No, ask a person