AI agent

How does the agent learn from my team's replies?

Replies your team writes, and drafts they rewrite, become answers the agent may reuse once someone approves them.

Updated 2026-10-15 · 4 min read

Open Workspace → AI Agent → Learning. Every time someone on your team answers a customer — by email, on WhatsApp, or by rewriting one of the agent's suggested replies before sending it — that answer can become a candidate. You approve the good ones, and the agent may reuse them for the next person who asks the same thing.

This is separate from your documents. A document is something you authored for the agent; a learned answer is something a colleague already said well.

Choosing how the agent learns

The card at the top of the Learning tab sets the mode for your whole team:

  • Review first (recommended) — answers are collected as suggestions, and the agent only reuses one after someone on your team approves it.
  • Automatic — answers that are general enough are reused straight away, with nobody checking them first. You can still remove them later.
  • Off — the agent does not learn from your team at all.

Automatic carries a warning for a reason: a reply that was right for one customer becomes a reply sent to strangers. Use Review first unless you have a specific reason not to.

Whatever the mode, an answer that mentions a price is never learned — prices always come from your price list, so a remembered amount cannot go stale in the agent's mouth.

What counts as a correction

  • Your own reply — you or a colleague answered the customer yourselves.
  • An edited draft — the agent suggested a reply and someone rewrote a real part of it (roughly 15% or more) before sending. Fixing a typo or a greeting is not a correction and is not collected.
  • A discarded draft — thrown away. The agent is told the draft was never sent, and it is never learned from.

Each candidate in the queue carries a badge — Own reply, Edited draft or From history — so you can tell a direct correction from an everyday answer.

This week, in numbers

The This week card shows what your team did with the agent's replies over the last seven days, against the week before: the share sent unchanged, how much text was changed on average, how many suggestions were discarded, and how often your team answered instead of the agent. Rising "sent unchanged" and falling "discarded" mean the agent is getting closer to how you answer.

Suggested improvements

Once a week the agent reads how your team corrected it and may suggest a change to its tone of voice, something it should always know, or something it must never say. They appear under Learning → Suggestions. Nothing changes until someone on your team presses Accept — an accepted suggestion is saved exactly as if you had typed it in the Behaviour or Knowledge tab. A rejected suggestion is not offered again.

Who can review

Anyone on your team can approve or reject answers and suggestions. When new answers are waiting, the team gets a notification that opens this tab.

The review queue

Learning → Queue lists candidates with the reason each one qualified and who wrote it. For each you can Approve, Edit & approve, or Reject.

Edit before approving. A reply written to one customer usually names them, or their car, or last Tuesday — strip the specifics and save what you would happily send a stranger. The colleague's original is always kept, so editing the candidate never rewrites history.

Build from history turns replies already in your message history into candidates. It runs in the background and the queue fills as replies qualify.

Each candidate shows why it qualified and who wrote it.

Conflicts

Learning → Conflicts appears when a knowledge document now covers a question that an approved answer also covers, and the two read differently.

The document already wins when the agent answers, so a conflict is not an outage — it is a cleanup prompt. Keep answer re-dates it as current; Retire answer stops it being reused and leaves the document to answer.

Keeping approved answers current

Learning → Approved is everything the agent may reuse, with how often each has been used and when it was last confirmed.

An approved answer can stop being used without being wrong — it ages out. Press Still true to re-date one without rewriting it, which is the cheap way to keep a good answer in circulation.

An answer I approved is never used

That is usually age or ranking rather than a fault. See Why is the agent using an old answer?, which explains the retrieval preview and the settings behind it.

Was this helpful?

Didn’t find your answer?

Sign in and open a ticket, or contact us — a question that needed a person is also how this article gets better.

Last verified 2026-10-15.