AI automatic replies
The AI combines the recent conversation, the buyer's order profile, the conversation product and shop best sellers, writes a reply and sends it — then steps aside as soon as a human takes over.
Turning it on
- Open the AI reply page;
- Switch it on per shop (each shop is controlled independently);
- Optional: write a shop prompt covering selling points, after-sales policy and the tone you want.
The more concrete the prompt, the better the reply. Free shipping over R$99; returns within 7 days; no dispatch on weekends beats be friendly by a wide margin.
What the AI can see
For every reply the AI reads four kinds of context:
- Recent conversation — what has already been said here;
- Buyer order profile — what this buyer bought from you and how much they paid;
- Conversation product — which listing the question is about;
- Shop best sellers — for related suggestions.
All of it is read from cache. The AI never triggers a full shop fetch on the reply path, which keeps replies fast and protects your platform API quota.
Four behavioral safeguards
| Safeguard | Behavior |
|---|---|
| Only answers the buyer | If the latest message in a conversation is not from the buyer, the AI stays quiet — no self-triggering loops |
| Steps aside for humans | After an agent replies (including the seller replying from the platform app), the AI stays silent for 30 minutes by default |
| Clear data boundaries | Price, stock and order status come only from system data; the AI does not invent them |
| Sensitive cases go to a human | Refunds, cancellations and complaints: acknowledge the buyer, hand off, and raise a to-do |
The silence window is adjustable on the page. To stop everything temporarily, use Pause automatic replies in the Assistant, or switch the shop off.
When the AI stays silent
- The latest message is not from the buyer;
- The conversation is inside the human-handover silence window;
- The AI decided there is nothing to answer (the buyer sent only a thumbs-up) — that is a normal outcome and does not raise a to-do;
- The model judged the case needs a human — it raises a to-do instead of guessing;
- Credit ran out, or a platform or model call failed — a to-do is raised to alert you, and the message itself is stored as usual.
When credit runs out the AI stops replying. Watch the balance notice on Credits and billing.
Billing
AI replies are billed by actual usage, broken down per agent or per shop on Credits and billing. Welcome messages, timeout replies and translation are not billed.