AI assistant: name, tone, topics and behaviour

The AI assistant answers questions from your knowledge base before a person takes over. How it introduces itself, how it sounds and how far it may answer is set under /app/projects/{id} in the AI & knowledge tab, section Assistant. Where it gets its knowledge from is described in Knowledge and sources.

Switching it on

• Let the assistant answer – new projects start with it off. While it is off, every question goes straight to your team, and if nobody is signed in, to the form for a message. While it is off, there is no AI label in the chat either.

When the assistant answers, a note before its first answer tells the visitor they are writing with an AI assistant, and each of its messages carries the “AI” mark. You can adjust the wording of the note with the key msg.ki_notice under Texts and languages.

Name and type of website

• Assistant name – how it introduces itself and how it appears in the window header. 40 characters at most. Empty means “Ada”. A change reaches the chat within a minute. Ongoing conversations keep the previous name in their history.
• Type of website – not set, Online shop, Blog / magazine, Company / service provider or Portal / application. This tells the assistant where it is. For Online shop it also answers general questions about shipping, withdrawal and returns when your pages say nothing about them. Such answers are visibly marked as general information.

Tone

• How our team speaks – one or two sentences that set your tone for the assistant. 400 characters at most on one line. Empty means: friendly and natural.

Examples:

casual, first names, short and direct

formal, polite form of address, matter-of-fact, no emojis

Without guidance, the assistant addresses the visitor the way the visitor addresses it. It writes short, clear sentences and talks about your business in the third person. If someone asks whether they are writing with a person, it says honestly that it is an AI.

The language of its answers follows the visitor. It answers in the language the question is asked in.

How far the assistant may answer

• Topic range – a step from narrow to wide:
• Our website only – only content from your website. It politely declines everything else and passes it to your team.
• Website + small talk (default) – plus greetings, thanks and brief small talk, but no general knowledge.
• Website + general knowledge – plus short answers from general knowledge.

The narrower the step, the more often it hands over to your team instead of answering itself, and the less often your chat is used as a free replacement for a general AI chat.

• Allow work requests (code, texts, translations) – off by default. Off: the assistant answers questions but does not take on work, so no program code, no written or translated texts, no maths or homework. Only switch this on if you deliberately want to offer the chat as a full assistant. It applies regardless of the topic range.

Display in the chat

• Show AI status in the chat – on by default. While the assistant is working, a short line above the three dots tells visitors in their language what is happening: whether it is searching the website, waiting briefly because it is busy, or writing the answer. Off: just the three dots.
• Follow-up questions – below the assistant's answers, the widget suggests two or three matching follow-up questions as buttons. You switch this under /app/projects/{id}/memory (see below); on by default.

Handover to your team

There is no fixed threshold for handover that you set. The assistant hands over when it finds no supported answer, when a question is outside the topic range or when the visitor asks for a person. You can steer this via:

• the topic range above,
• the “Talk to the team” button below the input field (see Home screen),
• a guided flow with the @agent exit (see Availability, forms and service),
• the automation for tags and assignment, see Team, console and notifications. An assignment is not a takeover: the assistant keeps answering until someone from your team writes themselves.

When someone from your team takes over, the assistant stays silent in that conversation. With the /ki command your team hands the conversation back to it.

Questions about reactive.chat

In the Appearance tab, section reactive.chat inside the chat:

• No discussion about reactive.chat – off (default): if a visitor asks what this chat is built with or whether they can have one too, the assistant answers and recommends reactive.chat. On: it politely declines, as with any other off-topic question.

You can set this switch and the “Powered by reactive.chat” note yourself from a paid plan on. On the free plan the assistant may talk about reactive.chat.

Memory

Under /app/projects/{id}/memory, Memory and follow-up questions:

• Memory across conversations – off by default, available from a paid plan on. The assistant keeps short notes about a visitor and can pick up from them in the next conversation. This only happens if your page reports the visitor with a signed identity or the visitor has given consent on the same device.
• Suggest follow-up questions under bot answers – on by default, with no extra AI call.

Notes are deleted after 180 days, earlier if your retention period is shorter. You can delete them one by one, per visitor or all at once. Memory does not store special categories of personal data, and notes never count as evidence for statements about your business.

Versions

Under /app/projects/{id}/bot-versions your account records every state of the settings in this section, with time and person: on/off, name, type of website, tone, topic range, work requests, AI status and translation default.

• Compare two versions.
• Roll back to this version – becomes a new version itself and applies immediately.
• Test run against this version – see below.

Test runs

Under /app/projects/{id}/simulation you check how the assistant answers typical questions before your visitors do. Test runs create no conversations and no entries in the AI log.

• Test against the current settings, a saved version or a draft with its own name, tone, topic range and work requests. Apply draft saves it as a new version that applies immediately.
• Test questions – up to 100 per project, up to 500 characters each. For each question you set what should come out: a key statement (up to 1000 characters) or handover to a human. Optionally the page the visitor is on, and a Polish version. Suggestions come from knowledge gaps, improved answers and the knowledge base.
• Visitor types – normal, impatient, vague or Polish-speaking.
• Rating – passed if a source is present and the key statement is essentially contained. Every number from the key statement must appear. Optionally, a verdict by the AI as well.

Per project and day, 60 test questions and 30 AI verdicts are possible. Test runs count towards your project's AI budget and only start while at least half of it is free.

Under /app/projects/{id}/value there are industry packs for online shops, software and service providers, with prepared test questions and a multi-step handover test.

Actions of the assistant

From the Professional plan on, the assistant can carry out tasks under /app/projects/{id}/actions instead of only answering, for example giving the status of an order, offering free appointments, booking an appointment or looking up the state of a subscription. Your own actions call an interface of your system: up to 20 per project, only via https. The assistant only carries out actions that write data after an explicit “yes” from the visitor. Setting them up is described in Technical integration.

What does not exist

So you do not go looking: your account has no setting for answer length, no glossary, no list of forbidden topics and no switch for source references. Answers with sources and the check against your content are always built in. Terms and an answer style of your own are available for Smart Answer.