Business Agent · Knowledge

It answers from your knowledge — and nothing else

A confident wrong answer about your prices costs you the customer. So the agent is locked to the material you give it: your text, your files, your links, your images. Ask it something outside that and it says so, and hands you the question.

Build your agent — free Try to catch ours out

Free to build. You only pay when you put it live.

The Business Agents dashboard: a character picker showing an agent called Xavier with an alert count and a lead count, tabs for Overview, Setup, Showcase, Conversations, Workspace, Leads, Meetings, Alerts, Test and Settings, and a row of counters for conversations, calls, leads captured, upcoming meetings, open alerts and messages.

One agent, one dashboard. The knowledge you write lives under Setup; what it could not answer lands under Alerts; who it spoke to lands under Leads.

Most AI chat has one job. This one has one rule.

A general assistant is rewarded for always having an answer. An agent standing in for your business has the opposite job: it must never say anything you did not authorise. That is the whole design here, and it is enforced in the prompt every single reply is built from — not left to the model's good manners.

  • Your knowledge is fenced into the prompt as everything it is allowed to say, on every channel it works
  • Asked something outside it, it says it does not know instead of producing a plausible number
  • The same lock applies whether it is on your website, in its inbox, in your DMs or on a live call
  • Nothing you wrote is ever paraphrased into a promise you did not make
Illustration: a white speech bubble held inside a violet ring with a closed padlock, while grey scribbles drift outside the ring unable to get in.

Everything you set, one card at a time

Setting the agent up is seven cards on one screen, and you can leave most of them alone. Open any of them here to see exactly what it decides before you ever sign up.

Status Whether this agent is live and answering.

One switch, and it is the only thing standing between a configured agent and a live one. Turning it off is not a reset — it stops the public page and the embedded widget answering, and everything you set up is still there when you turn it back on.

Enable business agent

On, and anyone with the link can chat to it. Off, and the link and the embed stop answering — nothing you configured is lost. Useful when you are mid-rewrite of the knowledge base and would rather it said nothing at all.

Identity & tone Who the agent is, how it speaks, and the language it leads in.

This card is about manner, never about facts. Changing the tone changes how a sentence sounds; it can never change what the agent is allowed to say — both settings stay strictly locked to the knowledge you gave it.

What is this agent?

A short brief in your own words — who it works for and what it is there to do. It is the frame the agent reads before your knowledge, so "a booking assistant for a dental practice" and "a technical pre-sales rep" behave very differently on the same material.

Tone

Professional writes formally, the way a company would. Personal is warmer and uses first names. Pick the one that matches how your customers already hear from you.

Primary language

The language it leads in — in chat replies and on live voice calls. Leave it on Automatic and it simply mirrors whatever language the visitor writes in, which is usually what you want on a site with international traffic.

Opening message

The first thing a visitor sees when the chat opens. Leave it blank and the agent greets people in its own words instead of reading a script.

Replies Whether each answer is written by AI or is always the same canned message.

Not everyone wants an AI answering. This card lets you keep the whole machine — the public page, the widget, the DM replies, the lead capture — and turn the intelligence off, sending one fixed sentence instead.

AI-generated

The default. A fresh, knowledge-locked reply is written for each message.

Generic message

One message you write, sent to everyone, word for word — in chats, in comment replies and in keyword replies, with no AI involved at all. The usual reason to reach for it is a holiday, a launch queue, or a business that just wants "thanks, we'll be in touch" and a captured email.

Voice & calls How the agent sounds, and whether visitors can call it from the chat.

A live voice call is the same agent, out loud, answering from the same knowledge. It is off until you switch it on, and it comes with the two limits you would want before putting a phone number on the internet.

Voice

The voice used on live calls. Leave it on Automatic to match your character's gender, or pick one from the library and hear it before you save.

Let visitors call the agent

Adds a Call button to the chat on the public page and in the embed, so a visitor can start a voice conversation without dialling anything or installing anything.

Call length limit

How long a single call can last before it ends itself. The caller sees the time remaining while they talk, so nobody is cut off by surprise.

Calls per visitor

How many calls one visitor can start in 24 hours — the setting that stops an open call button becoming an open budget. Set it to 0 for unlimited; each caller still has a daily minutes budget.

Email The address people can write to, and the subject line of onboarding emails.

The agent can have its own email address and work that inbox from exactly the same knowledge base — so an answer you fix once is fixed in the chat, the DMs and the email at the same time.

Email address

An address on the platform's mail domain that belongs to this agent. Anything sent to it arrives in the agent's inbox with a reply drafted from the whole thread, not just the last message.

Onboarding subject

The subject line used when the agent sends a welcome email. Write {agent} for the agent's name and {client} for the recipient's, or leave it blank for the built-in default — you can still override it per person later.

Knowledge Everything the agent is allowed to say — text, files, links and media.

The card the whole product turns on. Four different kinds of material go in here, and between them they are the complete answer to "what does it know?" — there is no fifth, invisible source.

Knowledge base

A single text box, up to 16,000 characters, holding everything the agent may tell a visitor: what you sell, your prices, your features, your policies, your FAQs, your links. It answers only from what is written here and never guesses around it. A live counter tells you how much room is left.

Research files

Documents you upload — a PDF price list, a policy, a spec sheet. Files are indexed rather than pasted, so the relevant passages are pulled in for the question being asked. This is how a knowledge base gets bigger than one text box without the agent getting vaguer.

Reference knowledge

Links, document summaries, images and videos. Links are read automatically when you save: the agent opens the page, writes down what it says and keeps that as knowledge. Each link carries a badge showing whether it has been read, with Re-check after the page changes and Deep read for a longer write-up.

Images & video

Up to ten images and ten videos the agent can actually show in the chat when it genuinely helps — and, for images, look at and describe rather than just link. See it can read your pictures below.

Paid add-ons Bundles your clients pay to unlock — one price, one purchase.

Knowledge you sell rather than give away. Create an add-on, put any number of files or reference items inside it, and put a price on it.

What the visitor sees

The items stay visible in their workspace with a padlock and a price, but cannot be opened.

What the agent knows

It does not get the contents, so it cannot give them away by accident. It does know the add-on exists and what it costs, so it can offer it — with an Unlock button right in the reply.

After payment

One payment opens everything in the add-on, and the agent can answer from the whole bundle in that client's very next message.

Lead collection and meeting booking are not on this list on purpose — they live on their own Leads and Meetings tabs, next to the results they produce.

Illustration: a stack of documents, a PDF file, a chain link, a photograph and a video play button, each connected by a violet line flowing into one single rounded card.

Four ways in, one thing it knows

You do not have to rewrite your business into a text box. Most of what the agent needs already exists somewhere — as a page, a PDF, a photo of a menu — and each of those is a way in. Whatever you use, it ends up as one body of knowledge behind every channel.

  • Write it — the knowledge base text box, up to 16,000 characters
  • Upload it — research files, indexed so the right passage is found rather than the whole document stuffed in
  • Link it — paste a URL and the page is read, summarised and kept
  • Show it — images and video the agent can put in front of a visitor mid-conversation

Start with your website

Paste a link and it goes and reads the page

An attached link used to mean the agent knew a URL existed. Now every link you add is opened, read and written down — so "it's all on our website" is a genuinely complete answer.

It reads more than the one page

The agent takes the page you gave it plus up to four more it discovers from there, so pointing it at your homepage picks up the pages your homepage links to.

You can edit what it wrote down

The write-up it produces is yours to correct. Anything you had already typed about the link is treated as authoritative and merged in — your notes are never overwritten by the crawl.

A badge that tells the truth

Each link shows whether it has actually been read. Change the URL and the badge flips to "re-check" rather than quietly claiming a page it has never seen.

Deep read when the summary is thin

One button asks for a fuller write-up of a page — the escape hatch for a dense pricing page the automatic pass summarised too politely.

Sites that need a browser

A page that only renders with JavaScript is opened in a real browser instead of being given up on, so a modern marketing site is not a blank read.

Nothing silently rots

When a site is down or unreadable you get "check failed", with the reason, rather than an agent confidently answering from a page that was never fetched.

It can look at your pictures, not just link to them

Upload a photograph of a menu, a size chart, a floor plan or a finished job, and the agent can genuinely read it when a question needs it — and show it to the visitor when a picture answers better than a paragraph.

  • A freshly uploaded image is described automatically, and you can edit that description before saving
  • It only looks at an image when a question actually calls for it, rather than on every message
  • It can show an image or a video in the chat — at most one per reply, introduced in words
  • It never pastes a raw file URL at a visitor; what appears is what you configured
Illustration: a photograph being crossed by a violet scanning beam, with lines carrying small shapes out of the picture into a card beside it.

The same knowledge, out loud

Switch calls on and the chat grows a Call button. What the caller hears is the same agent under the same lock — it does not become chattier or looser because it is speaking.

A live voice call in progress with an agent named Leo: a portrait, the word Connected, a running timer, the remaining minutes for the day, a microphone button and a red End call button.

The caller sees who they are talking to, how long they have been talking and how much time is left today — the limits you set in Voice & calls, shown honestly rather than enforced by a sudden hang-up.

Give a client the knowledge, not just the answers

Every agent can hand out a private workspace: the files, the links and the media you chose, with the agent sitting beside them to explain any of it. It is a link, not an account — they open it and it is there.

A client workspace: a left sidebar listing files, links, document summaries, images and videos; a welcome document in the middle explaining the three panels; and a live chat with the agent on the right.

Files on the left, the document you are reading in the middle, the agent on the right — answering about the exact thing on screen. Voice calls and meeting booking are in the same header.

Tell it what to change, in words

You do not have to be at the Setup tab to edit a knowledge base. Turn agent mode on in the chat and say what is wrong — "add our new opening hours", "we don't do same-day delivery any more" — and it makes the edit while you watch, then shows you what it changed.

  • Add to the knowledge base without retyping the parts that were already right
  • It will not put an agent live with an empty knowledge base — it drafts one and comes back to you
  • Everything it writes is still an ordinary edit you can open and change by hand afterwards
Agent mode switched on in a chat panel with an agent called Leo, and a dialog over the dashboard reading Okay, Leo is making some updates. Please wait.
Illustration: a speech bubble with a question mark, and a line arcing from it to a small flag planted on a board.

What it could not answer is the most useful thing it produces

An agent that says "I don't know" would be a dead end if the story stopped there. Instead the question is flagged for you with the customer's own wording attached, so the gap in your knowledge base is handed to you rather than quietly costing you business.

  • Every unanswered question becomes an alert on your dashboard, in the visitor's exact words
  • Visitor feedback is flagged the same way, so complaints do not vanish into a transcript
  • Read them as a to-do list: each one is a paragraph your knowledge base is missing
  • The agent picker shows the open count, so you see it without going looking

One house knowledge base, many agents

If you run agents for several clients, the things that are true of all of them — how you work, what you charge for, your terms — do not belong in every knowledge base separately. Group the agents under one enterprise and the shared material reaches all of them at once, on every channel.

  • Shared text, links, images and files behind every member agent
  • Each client keeps their own knowledge, conversations and leads — nothing bleeds across
  • Fix a house answer once instead of in eleven places

For agencies

Illustration: one large card connected by violet lines down to four smaller cards, each carrying its own coloured dot.

Common questions

How is this different from ChatGPT answering for my business?

A general assistant is rewarded for always having an answer. This agent is fenced: your knowledge goes into every reply as the complete list of what it is allowed to say, on every channel it works. Asked something outside it, it says it does not know rather than producing a plausible price or policy you never set.

What can I actually put in the knowledge base?

Four things. A text box of up to 16,000 characters that you write. Research files you upload, which are indexed so the relevant passage is found rather than the whole document pasted in. Links, which are opened and read for you. And images and video the agent can show a visitor mid-conversation.

Do I have to write it all myself?

No. The build reads the website you give it and drafts the knowledge base for you, so the first version is already there to edit instead of a blank box. After that you can also just tell the agent in chat what to add or correct, and it makes the edit while you watch.

What happens when I paste a link?

It is crawled on save — the page you gave it plus up to four more it discovers from there — and written up as knowledge you can edit. Each link carries a badge showing whether it has really been read, with Re-check after the page changes and Deep read when the summary was too thin. Anything you had already written about that link is treated as authoritative and merged in, never overwritten.

Can it read an image, or does it just link to one?

It can read it. A freshly uploaded image is described automatically and you can edit that description, and the agent looks at the picture itself when a question calls for it. It can also show an image or a video in the chat when that answers better than a paragraph — at most one per reply, introduced in words rather than pasted as a raw file URL.

What happens when it cannot answer something?

It tells the visitor honestly and raises an alert for you with their own wording attached. Read as a list, those alerts are exactly the paragraphs your knowledge base is missing.

Can I turn the AI off and still use everything else?

Yes. Switch replies to a generic message and one sentence you write is sent to everyone, word for word, with no AI involved — in chats, comment replies and keyword replies. The public page, the widget and the lead capture all keep working.

Can I sell some of the knowledge instead of giving it away?

Yes, as a paid add-on. The items stay visible to your client behind a padlock and a price, and the agent does not receive their contents, so it cannot give them away — but it does know the add-on exists and can offer it with an Unlock button in the reply. One payment opens the whole bundle, and the agent can answer from it in the next message.

I run agents for several clients. Do I maintain the same answer in each one?

No. Group them under one enterprise and shared house knowledge — text, links, images and files — reaches every member agent on every channel, while each client keeps their own knowledge, conversations and leads separate.

Give it what you know

The build reads your website and drafts the knowledge base for you, so the first version is already there to edit rather than a blank box. Free, and you can talk to the finished agent before deciding anything.

Build your agent — free See pricing

Aria, in the corner of this page, is one of these agents. Ask her something we never told her.