Inbox

AI reply drafts

Let AI draft replies you review and send.

Drafts, not replies

Above the composer, a sparkle button asks AI to draft a reply for the conversation you have open. It reads the thread so far and writes something you can send as is, edit first, or throw away. The important part is what it never does: it never sends on its own. Every draft still passes through your hands before a customer sees it.

Why it never auto-sends
A drafted reply can be wrong in ways a person catches at a glance and an automated send cannot: the wrong price, the wrong tone, a question the AI misread. Keeping a human in the loop is what makes it safe to use on every conversation, not just the easy ones.

Three lengths and credits

Pressing the sparkle produces one answer at three lengths, shortest first, so the pick is a quick judgement about how much this customer needs rather than a comparison of three different replies. Each press uses AI credits from your plan, the same pool the AI chatbot spends from, so drafting for every message on a busy day does add up.

  • Short is one paragraph: the answer and nothing else, for a thread that is already moving.
  • Medium is always two paragraphs: the answer, then whatever usually gets asked next.
  • Detailed is always three: the answer, the detail behind it, then what happens next. For a question that genuinely needs explaining.

All three commit to the same facts. The longer ones explain more, they never promise more, so picking a length can never quietly change what you have agreed to. Pick one, edit it if needed, and send it through the composer like any other message. Ask again if none of the three fit; a fresh set costs credits again.

Written like your customer

The drafts are written the way the person on the other end writes: how they greet you, what they call you, how long their sentences run, whether they use emoji, and the short forms they type. A customer writing full careful sentences gets careful sentences back; one typing brape ek does not get a formal letter.

Mirrored, not copied
Typos, missing spaces and shouting are never handed back, whatever arrives. And your own settings still win: your Reply language and Tone choices in AI, then AI reply, decide the language and the register, and mirroring only fills in what those leave open.

What a draft reads

Two things: the last few messages in the thread, and your knowledge base. The second is what lets a draft answer berapa harga with the real figure from your price list instead of offering to check and get back to them. It looks up whatever the conversation is actually about, not just the last line, so a short follow-up still finds the package you were discussing three messages ago.

  • The popover names the sources a draft quoted, under the three candidates. Nothing there means it only saw the thread.
  • Facts outside both are still never invented. The draft asks for what it needs, or says you will check.
  • By default it reads every ready source, including ones you add later. Tick specific sources in AI, then AI reply, to narrow it, or turn it off there entirely.
Ready sources only
A source still ingesting, or one that failed, has nothing to retrieve. Check the Knowledge tab shows it as ready if the drafts do not seem to know about it.

Where it is configured

AI reply has its own tab under AI, next to Chatbots and Knowledge base. That is where you configure what powers the sparkle button, kept separate from the AI chatbot that can answer customers directly. See AI chatbot for the difference between a bot that replies on its own and a draft that waits for you, and Knowledge base for what the draft actually reads before it writes.

Because the drafts come from the same knowledge your chatbot uses, keeping that knowledge current pays off twice: better bot answers, and better drafts for your team to send by hand.