How the AI answers

The AI answers only from your sourcesCopy Anchor Link

When the AI writes a reply, this is all it may use:

• Your knowledge base (customers get the published version; see Knowledge base).
• Policy numbers (政策数字): return windows, shipping fees, warranty periods and similar figures come only from here.
• Skills: handling instructions you write for the AI (see Test chat and releases).
• Synced product data: prices and variants. Stock is only described as in stock, low or sold out, and data not updated for more than 7 days is flagged as possibly out of date.
• A signed-in customer's own orders: the latest 5, without shipping address or email.

If the answer is not in those sources, the AI does not make one up; it hands the customer to a person.

Order tracking does not go through the model. When a signed-in customer asks where their order is and is not upset, the reply is built from the order and shipment data with a fixed template, and the autonomy levels below do not apply to it. An anonymous visitor who asks about an order is asked to verify who they are first and is told nothing about any order. To have customers signed in, see Identity tokens.

Every reply is checked before it is sentCopy Anchor Link

Before any reply the AI wrote reaches a customer, a separate judging model checks it:

A blocked reply is never sent. The customer gets a short message in their own language saying the team will reply shortly, the conversation goes to a person, and the team can see the blocked draft in the inbox.

The check is strict about answers that drop a condition. If your source says "unopened items can be returned within 30 days" and the reply only says "you can return it within 30 days", it will usually be blocked as contradicting the source. That is deliberate: a hand-over to a person is better than a wrong promise.

If the judging service is unavailable, the AI sends nothing on its own and every message goes to a person.

When the AI hands over to a personCopy Anchor Link

• The customer clearly asks for a person.
• The answer is not in your sources, so the AI declines to answer.
• The reply fails the check above.
• The question type is set to R0 or R1 (the reply goes to a person as a draft; see below).
• The judging service is unavailable, or writing the reply failed.
• The AI has already sent 30 replies in this conversation.
• This month's AI spend cap has been reached (see below).
• On the Free plan, this month's AI conversations are used up (paid plans never stop; extra conversations are billed, see Plan and billing).
• The customer's segment is set to human first, it is within working hours and someone is online (see Customers).

Who gets the conversation, and what happens when nobody is online, is covered in Teams and routing.

Autonomy levels R0–R4Copy Anchor Link

In Settings (设置) → AI autonomy (AI 自主程度) you set, per question type, how far the AI may go on its own. Only owners and admins can change it. The console shows the levels with these labels:

The question types you can set separately:

The "other" type (其他) is not in the list and always uses the default, R1.

Highest level per planCopy Anchor Link

Levels above your plan cannot be selected. After a downgrade your setting is kept but the plan's highest level is used (the page says so), and upgrading again restores it.

In a test workspace you can switch levels freely, without the evidence below.

Moving from drafts to auto-answers: promotion evidenceCopy Anchor Link

Raising a question type from R0 or R1 to R2 or above requires its promotion evidence to be all green. The tag next to the level shows "evidence N/8" (凭证 N/8) or "evidence met" (凭证达标); open it to see each figure.

The evidence covers the last 30 days of model-written replies for that question type (order-tracking templates are not included):

You can still raise the level before the evidence is met, but you must write a reason (at least 4 characters). The reason is recorded with the level.

Automatic demotionCopy Anchor Link

Demotion needs nobody's approval. It happens in two cases:

1. Someone marks an answer as wrong: when anyone on the team clicks Wrong answer (答错了) in the inbox on a reply the AI sent on its own (no person reviewed it), that question type drops one level immediately.
2. Repeated low ratings: 3 ratings of 2 stars or fewer on the AI's replies for one question type within 24 hours drop it one level. Ratings from before the last level change are not counted.

Demotion only applies above R1 (R2 to R1, R3 to R2, R4 to R3), one level at a time; two triggers at once still demote only once. The AI is never promoted automatically; raising the level again is up to you.

After a demotion, the settings page shows the reason and time in red under that question type.

Monthly AI spend capCopy Anchor Link

In Settings (设置) → Working hours and assignment (工作时间与派单), set the monthly AI spend cap in US dollars (AI 每月花费上限):

• Paid plans have no cap by default; Free and test workspaces default to $50. It counts what the AI's model calls actually cost when answering, including calls made in test chat.
• It is counted per calendar month (UTC). The page shows this month's spend against the cap and the number of replies, and warns when you reach 80%.
• At the cap, every new message goes straight to a person and the AI is not called. Customers are told someone will reply; they never see an error. The page shows how many messages were handed over because of the cap.
• This cap is separate from the AI conversations included in your plan; see Plan and billing.

Only owners and admins can change it.

Customers know they are talking to an AICopy Anchor Link

The AI says it is an AI (as required by Article 50 of the EU AI Act). This cannot be turned off:

• Website chat: the greeting at the top of the window says this is the store's AI assistant and that a person can take over any time. AI replies are labelled "AI assistant", replies from your team "Customer service". Both are shown in the customer's language, in 12 languages. An AI draft a person sends unchanged is still labelled as the AI; an edited one is labelled as customer service.
• WhatsApp and Instagram: the AI's first reply in each conversation starts with a line saying so; for a store called Lina Toys it reads "(This is Lina Toys's AI assistant. Ask for a person any time.)".
• Email: every reply the AI sends on its own ends with that line.

Messages your team sends (including drafts sent after review) do not carry that line on WhatsApp, Instagram or email.