The AI assistant answers customers only from your knowledge base: for each question it searches your articles and builds the reply from what it finds, in the customer's language. When nothing in the knowledge base supports an answer, it says so instead of guessing, and in a normal inbox it passes the conversation to your team.
Which conversations it answers
The assistant replies only in conversations that are assigned to it. A new platform already routes its channels to the assistant; a channel you add later needs a routing rule (see Routing a new channel to the assistant).
It works in the web chat, Telegram, WhatsApp and email. It answers open conversations only: a snoozed or resolved conversation gets no reply. When the customer writes again, the conversation reopens and the assistant answers if it is still the assignee. In Telegram and WhatsApp a resolved conversation reopens only within the inbox's Reopen window (hours); a later message starts a new conversation (see Handoff to agents and back).
Answers come only from your knowledge base
Before it states anything about your company, products, prices, policies or how something works, the assistant searches the knowledge base: your knowledge base articles and the published Help Center articles that are enabled for AI. Search looks for pieces of text in two ways: by meaning, and by the words of the question, other forms of the same word included, in a full-text index on the knowledge server of your platform. When a reranking model is set up on the Providers page and answers, it reads up to 30 pieces found by meaning and up to 20 found by words (up to 80 found by meaning when the knowledge server does not answer) and moves the pieces that actually answer the question to the top. Up to ten pieces reach the model. Without a reranker, search ranks by meaning only. More on how articles are prepared for search: Knowledge base basics.
On top of your settings, the assistant is given fixed rules:
- it does not use its own general knowledge, and treats anything missing from the knowledge base as unknown;
- it politely declines requests outside your support topics (writing code, maths, trivia, other companies), even when it could answer them;
- it never makes up prices, steps, policies or promises, and does not promise to check or follow up later unless it does so right away with a tool;
- when a request is unclear, it asks one short clarifying question;
- replies are short and conversational, with simple formatting (bold, links, lists) only where it helps.
Refusing instead of inventing is enforced in code
Instructions to a model are requests, and models do not always follow them. So the most important rule is checked by the system, not by the model: a reply longer than 160 characters that is not backed by any knowledge base result is not sent to the customer. The only exception is the closing reply after the customer has confirmed they are done.
Instead, the customer reads "I could not find an answer to this in our knowledge base, so I am passing you to a colleague. They will reply here." and the conversation goes to your team. In an AI chat without agents the customer gets a different message (see the last section).
Short replies such as greetings or a clarifying question need no knowledge base and pass. A successful call to one of your custom tools also counts as support: data from your own system is as reliable as an article.
It searches even when the model does not
The model decides by itself when to search, and sometimes answers straight away. When its first draft comes without any search or other tool call, the system searches the knowledge base with the customer's message. If that search finds something, the draft is discarded and the model answers again with the results. If it finds nothing, the draft stays, and a draft longer than 160 characters is then not sent, because nothing in the knowledge base supports it.
A short follow-up such as "and how much is it?" says little on its own, so the search also includes up to two earlier messages of the same customer (about 500 characters). Messages with an image are left to the model; the 160-character rule above still applies to them.
Several messages in a row get one answer
People often send a question in two or three messages. If the customer writes again while the assistant is still preparing a reply, that reply is dropped, and the next attempt answers all the messages at once.
When such a turn contains two or more questions, each message is searched separately before the model writes (up to five), so the second question is not lost behind the first. Greetings and short acknowledgements like "ok" or "thanks" in the turn are recognised by meaning, in any language, and are not searched.
The first message says it is an AI
When a new conversation starts in the web chat, Telegram or WhatsApp and a rule hands that inbox to an assistant, a first message goes out automatically on behalf of that assistant. An inbox where your agents answer gets no first message, so nobody is told an AI answers when a person does. In the web chat and Telegram it starts with your greeting from the inbox setting First message; a WhatsApp inbox has no such field. Then comes a notice added by the system: "You are chatting with an AI assistant. Ask for a human at any time and I will pass you on." In an AI chat without agents the notice is shorter: "You are chatting with an AI assistant."
The notice cannot be switched off. Disclosing the AI at the first interaction is a legal requirement (EU AI Act), and because the notice is part of the conversation, it stays on record. If an agent hands a conversation to the assistant later, the notice goes out right before the assistant's first reply. Email conversations get no automatic first message, since an automatic reply to every email is how mail loops start.
The first message is not counted as a reply: the assistant still answers the customer's first question. When the customer's first message is only a greeting or an acknowledgement, the first message has already answered it, and the assistant stays silent instead of greeting a second time. In the web chat and WhatsApp its language is the language the assistant will answer in, unless your greeting is written per language and has no version for it. Otherwise, and always in Telegram, it is the browser or Telegram language, then the language of your platform.
Reply language
The assistant answers in the language of the conversation, judged from the customer's recent messages rather than the last one alone, so a single foreign word does not switch the language. If your knowledge base is written in another language, the assistant translates what it finds.
The Languages field of the assistant changes this:
- empty: the customer's language, whatever it is;
- one language: always that language, even if the customer writes in another;
- several languages: still the customer's language, even when it is not on the list; the list helps only when the language is unclear, for example in a very short message.
For a short message the choice is made among the languages you expect: the browser or Telegram language, the languages on the list and your platform language. If none fits, the browser or Telegram language is used, then the first language on the list, then the platform language. The same order applies if language detection is unavailable for a moment.
A reply that comes out in another language is sent back to the model once to be rewritten. Messages from the system, such as the AI notice or the handoff message, are sent in the customer's language when the panel has a translation for it, otherwise in your platform language.
Screenshots and files
When Model supports image input is on (it is on by default on our platforms), images that customers send are passed to the model: up to three per message and four per conversation, up to 8 MB each, the newest first. The assistant may describe what is on a screenshot, but it diagnoses a problem or gives advice only when the knowledge base covers it. Otherwise it says it cannot help with that and, when Offer handoff to a human is on, offers a person.
Files it cannot read, and images over these limits, are not ignored: the assistant is told about them and can ask the customer to describe them in text. Voice messages are turned into text and answered like typed ones; when a voice message cannot be turned into text, the assistant asks the customer to type the question (see Voice messages).
Links in answers
When a customer asks for a link, the assistant gives the address of the page its answer came from: an imported web page, a Confluence page or a published Help Center article. Articles written by hand in the panel and uploaded files have no page address. The assistant is told to give only addresses that appear in the knowledge base or in the conversation, and never to make one up.
What the assistant knows about the customer
The assistant sees the contact's details that are known to the panel: name, email, phone, country, your own user ID and custom attributes, plus the conversation subject and attributes. They are passed to the model as reference data from the customer, never as instructions it must follow.
It reads up to the last 30 messages of the conversation. For a customer whose identity is verified (see Custom tools), it can also look at their other conversations from the last 30 days, up to three of them.
Closing a conversation after a real answer
In a normal inbox, when a reply fully answers a question, the assistant adds a short question asking whether that helped, unless the reply already ends with a question. In the web chat the question comes as a separate message. In email, Telegram and WhatsApp it is added to the end of the answer. The web chat, Telegram and WhatsApp also show buttons such as Yes, thanks and I need more help in the customer's language; when the assistant asks another question with two or three obvious answers, these channels show them as buttons too. In the web chat the buttons stay under the assistant's message until the customer writes.
The assistant marks the conversation resolved only after the customer confirms they are done, and only if an earlier reply has already answered them: it never closes a conversation on the customer's first message, and never in the same reply that asks for confirmation. When a conversation is resolved, the customer receives a satisfaction survey if CSAT Surveys is on for the inbox.
When there is no answer
What happens depends on the inbox mode:
| Situation | Normal inbox | AI chat without agents |
|---|---|---|
| No answer in the knowledge base | The customer is told and the conversation goes to your team | "I could not find an answer to this in our knowledge base. Please ask another question." The conversation stays with the assistant |
| The model fails or takes too long | "I am passing you to a colleague. They will reply here." and the conversation goes to your team | "I could not answer just now. Please try asking again." |
| The customer asks for a person | The assistant passes the conversation on, if Offer handoff to a human is on | It explains that only the AI assistant answers in this chat |
| Reply limit (Max turns) | Applies | None |
| Confirmation question, closing, survey | Yes | No |
In an AI chat, "ok", "thanks" or a greeting are not searched, and if the model has nothing to add, the assistant simply stays silent. A closing question such as "Anything else I can help with?" at the end of a reply is removed. Details on handing over: Handoff to agents and back.