Live Chat AI Assistant

The Live Chat plugin can answer visitors for you. The AI assistant replies only from the Knowledge Base you write — it never improvises prices, policies, or commitments — and when the answer isn't there, it escalates: the visitor is told a human will follow up and your team is notified — while the assistant keeps helping with anything it does know until a human joins.

It is off by default. Nothing changes for a workspace until you switch it on.

The Knowledge Base

CMS sidebar → Knowledge (under Content). A workspace-scoped collection of documents — up to 300 of them, 2 million characters total — added three ways from the single New document button:

• Write text — free-form documents (pricing, policies, FAQ), up to 30,000 characters each.
• Web page — paste a URL and Brandfine fetches that one page and turns it into an editable document. Re-fetch anytime to pick up changes. (One URL = one document — no site crawling; pages that need JavaScript to render may come back empty.)
• Upload file — PDF, DOCX, TXT or Markdown (10MB max). The text is extracted immediately and shown for review — what you see after upload is exactly what the AI reads. Scanned PDFs aren't supported yet.

This is the assistant's only source of facts. If a visitor asks something your documents don't cover, the assistant does not guess — it hands off to your team. Practical consequence: the quality of the assistant is exactly the quality of your Knowledge Base.

Every document's text is editable — including URL and file extracts — and documents can be disabled without deleting them. The Knowledge Base is a platform section, not a live-chat feature: future AI surfaces read from the same place.

Order and pinning

The list is the order the assistant reads in: pinned documents first, then the rest top to bottom (Move up / Move down). Documents are sent in full until the Assistant reading budget runs out; the Assistant column shows where each one landed:

Pin what must never be missed (prices, rules, contact details) and put background material lower. A larger reading budget fits more documents in context, at more credits per conversation.

Index cards

Every document has a one-line card saying what it answers. The assistant reads the card, not the text, for documents that are on demand. When a document's text changes, its card is marked card to rebuild: open the document with Edit and either Rebuild card (a fraction of a credit) or Edit card to write it yourself (free). Hand-written cards are never regenerated. For documents that are in context, a stale card changes nothing until they drop out of context.

Turning it on

CMS → Plugins → Live Chat → Settings, per workspace. The AI assistant card has one main choice — when should it reply?

The optional Personality & instructions field shapes tone — see below.

Choosing the assistant's model

The AI settings let you pick the assistant's model per workspace: Fast (the default — quick and economical, under 1 credit per reply), Smart (sharper answers for nuanced questions, ~1–3 credits per reply), or Expert (our most capable model, best at subtle phrasing and long, detailed knowledge, ~3–8 credits per reply, and noticeably slower). All three follow exactly the same rules and answer only from your Knowledge Base — a smarter tier simply spends your credits faster in exchange for better answers.

The first reply of a conversation costs more than the ones after it: it loads your Knowledge Base into the model's cache, which the following replies reuse at a fraction of the price. The bigger the reading budget, the bigger that first-reply cost.

Personality and instructions

Personality & instructions (AI assistant tab, up to 4,500 characters) shapes how the assistant talks: its name, tone, what to recommend and when, what never to say. It's for behaviour, not facts — facts belong in the Knowledge Base, so the assistant never improvises them.

Instructions also decide reply length. A reply is capped at 1,200 characters and, past that, ends at its last full sentence; an instruction that asks for several parts in every answer (reflect, answer, exercise, next step…) makes every answer long. Ask for short answers explicitly if your visitors chat on their phones.

Recommending pages

AI suggestions are pages the assistant attaches to its reply. How your site shows them (cards beside the chat, a list, links) is up to your chat setup. In the AI assistant tab you choose how often and how many:

n is Suggestions per reply (1–6, default 3). Outside "Only for pages the reply mentions", suggestions don't have to be named in the reply, and they follow the topic as the conversation moves.

Pages the assistant can recommend (same tab) is the list they come from: add published posts by search, or any link by title and URL, up to 40. Labels group pages (beginner, paid…) for your own instructions to refer to. Once the list has pages, suggestions come only from it; any other page the assistant names becomes a link inside the reply. The assistant never suggests an address that isn't in your Knowledge Base or this list.

Conversation topics at a glance

Each conversation gets an AI-generated one-line topic ("Pricing for the annual plan") shortly after the visitor's first real message, so you can triage the inbox — and the WordPress plugin's conversation list — without opening threads. Early topics refresh once a few messages in, when the real subject is clearer. Topics are written in the language the visitor writes in (falling back to the page or workspace language only when the message is too short to tell), describe what the visitor wants in their own words, and work even when the assistant itself is switched off. Requires a plan with the AI assistant; each topic costs a credit or two.

While the assistant is composing a reply, the visitor sees an animated "typing" indicator in the widget — the same cue they'd get from a human agent — so the two-to-four-second wait never feels like silence.

What language it speaks

The assistant mirrors the visitor: write to it in Turkish, it answers in Turkish — whatever language your Knowledge Base is written in. The widget also passes the page's language along (WordPress sends the site locale automatically; other embeds use data-locale, the page's <html lang>, or the visitor's browser language), which settles borderline cases. When no page language is available and the visitor's message is too short to tell — a bare "Info?" — the assistant falls back to your workspace's default locale rather than defaulting to English.

How it stays in bounds

• Answers come only from your Knowledge Base and workspace profile (business, audience, voice, pinned facts from the AI Context tab).
• It never invents prices, discounts, availability, or policy.
• Off-topic requests are politely refused; manipulation attempts ("ignore your instructions…") are ignored or escalated.
• Replies are plain text; links are allowed only if they appear in your Knowledge Base entries.
• After 15 assistant replies in one conversation, it escalates to your team regardless — a hard cost bound.

Booking calls from the chat

If the Appointments plugin is active on the same workspace, you can also enable "Let the assistant offer call booking" in the AI settings. When a visitor asks to schedule a call, the assistant's reply carries a time picker right in the chat bubble — real open slots from your appointments availability. The visitor picks a time (and leaves an email if the conversation doesn't have one yet), and the request lands in your Appointments inbox as a normal pending booking for you to confirm or decline. Pausing the Appointments plugin silently removes the offer — no settings juggling needed.

Handing off to humans

When the assistant escalates, the visitor is told the team has been notified and your away-notification email fires — but the conversation doesn't go dead: the assistant keeps answering questions your Knowledge Base covers, asks once for an email so the team can follow up, and stays available — skipping messages that need no reply — until a hard turn cap. Once a human takes over, the assistant stays out permanently.

In the inbox, an AI-owned conversation shows a banner with a Take over button instead of the composer. Taking over — or simply sending a reply — silences the assistant for that thread. Assistant messages are marked with an "AI" stamp in the inbox and labeled AI assistant in the visitor widget, so people always know when they're talking to a machine.

AI credits

Assistant usage is metered in AI credits (1 credit = 1 cent of model cost). Your plan includes a monthly allowance that resets at the start of each calendar month; you can also buy top-up credit packs from Settings → Plan & Billing — those never expire and are spent automatically once the monthly allowance runs out. When both are exhausted the assistant simply stays quiet (your normal away-notification flow keeps working, and human replies are never blocked); it resumes on the next monthly reset or top-up. The billing page shows a live usage bar plus any extra-credit balance.

For SDK integrators

Assistant messages arrive with sender: 'BOT' (alongside VISITOR, AGENT, SYSTEM). Render BOT — and any unknown future sender value — on the business/agent side of your UI, ideally with an "AI" label. Don't switch exhaustively over the sender union; the server may add sender kinds before your SDK version learns their names. See the headless session docs for the message shape.