An AI agent that answers your own customers — on your site, in the app they already use, and on the telephone. Here is the whole thing, in six parts.
One assistant, with your knowledge and your rules, answering your own customers — in text, in voice, and face to face.
A member of staff that does not sleep. It answers questions about your products, checks an order, books a slot, takes a message, and hands over to a person the moment it should. It is not a chat window with a script behind it: it reads from your own material and it can act in your own systems while it talks.
The job decides how it opens, what it volunteers and what it must never say: a pre-sales assistant may not invent a price, an after-sales desk may not promise a refund it has not confirmed. Customer service, pre-sales, after-sales, bookings and reception, product help, course and community questions, lead capture and more — each is a different set of opening lines, guardrails and hand-over rules, not a different product.
Your brand, products, prices and policies, indexed into a knowledge space and retrieved by meaning at the moment it is asked. It answers from your material rather than from a paragraph pasted into a prompt — and when your material does not cover the question, it says so instead of inventing an answer.
Pick a voice from your asset library — cloned voices carry a consent record — then choose a character preset and move four dials: warmth, formality, initiative and humour. Answer length is set with them, because a companion and a support desk disagree about what a good answer is.
Three routes. Voice and text only, which is what a phone call or a WhatsApp thread is. A 3D figure drawn on the visitor's own device, which costs nothing per minute. Or photoreal video from a provider, billed by the minute. A face is worth paying for where the visitor is choosing something and wants to be walked through it; it is dead weight on a two-line question.
The same assistant answers in the language it is spoken to, with one knowledge base behind it. You write your material once. A shop whose customers are in Shenzhen and whose buyers are in Los Angeles does not run two desks.
Live conversation is metered by the minute: a model call, a voice, the room, and — on the photoreal-video route — the avatar stream, which is most of it. Credits are reserved before a session and settled against what it actually used. Text conversations are metered per message and cost a fraction of a spoken minute.
Four doors, and you can open as many as you like. The same assistant, the same knowledge, the same transcript behind all of them.
Paste one line into your page and the assistant is there. No platform to apply to, no account to verify, no licence: it is our software on our address, inside your page. This is also the only door where the face and the full-duplex voice work completely — every messaging app is text and voice notes.
A key can be a QR code on a poster, a table card, a shop window or a delivery note. The customer points a phone at it and is in a conversation — nothing to install. One key per shop, per poster, per campaign, so you can see which one brought the conversation, switch one off, or cap one without touching the rest.
Telegram, WhatsApp, Facebook Messenger, WeChat Official Accounts, WeCom and Discord. The customer writes to the account they already have in their pocket; the assistant answers there. This is how most of the world reaches a business — the dominant app is simply a different one in each country, which is why more than one is built.
Buy a number in the back office and it rings the assistant directly. Take a code on one of ours and your callers dial that number and key four digits. Or keep the number your customers already know and forward it — the number on your door does not change, and the assistant answers behind it. The phone is a capability you switch on per market, not the foundation: it needs a carrier, and in several countries a licence or a local company.
Your own site and a QR code: nothing at all. Telegram: nothing — a bot is one command. Messenger and Instagram: your own page, and no company needed. WhatsApp: a Meta business account and a dedicated phone number; without verification you can still answer everyone who writes to you first, which is what a service desk mostly is. WeChat Official Accounts: a registered company, in any country — an individual cannot get the service interface. Telephone: a carrier account, and the rules of the country you are calling into.
A key that is public — in your page source, or printed on a poster — can be walked through by anyone who reads it. Each one carries its own ceilings: minutes per session, sessions per day, how many at once, which sites may frame it, an expiry date, a total number of uses, opening hours in your own timezone, an address range, and a passcode. Every one is off by default, and each is a sentence you can explain to a customer who is refused.
A group chat is a channel like any other, except that the assistant stays quiet unless it is addressed — by name, by mention, or by a reply to something it said. An AI agent that answers every message in a busy group is not helpful, it is noise, and nobody invites it back.
Every door leads to the same desk: one contact, one history, one place a person takes over.
A customer who wrote on WhatsApp last week and rings today is one contact with one history, not two strangers. Your team sees the whole thread — what was asked, what the assistant answered, what it looked up and why it said what it said.
The assistant hands over when it is asked to, when it is out of its depth, and when the conversation has stopped going well. The person picking it up gets the transcript and what the assistant had already found, so the customer does not start again. Your own systems can be told the moment it happens, as a webhook, and open a ticket where your team already works.
Consent is recorded per contact and per channel, and a stop is honoured everywhere immediately — not at the next sync. Outbound to somebody who has not agreed, or who has withdrawn, does not leave the building. This is a legal state, so it is acted on here rather than exported for another system to decide about.
Upload a schedule — a spreadsheet, a document, or a PDF — and it fills a table you can edit like any other: who, when in their own timezone, why you are calling, what to ask, what came back. The assistant works the list under ceilings you set: a spend limit, quiet hours in the contact's clock, a cap on how many calls may be abandoned, and answering-machine detection so a voicemail is not treated as a conversation.
Each call writes its own outcome back onto the row: what was answered, what was agreed, what to do next. A field a person typed is never overwritten by the machine — you can always see which of the two filled a cell.
Every tool call is written into the transcript with its arguments and its result. When a customer is told something surprising, the answer to "where did that come from" is a line in the record, not a guess.
An assistant that can only talk is a brochure. Four ways in, and you choose per system — we host the AI agent, not your business software.
Many platforms now publish their own MCP server — Shopify, and a growing list of SaaS. Connect it and the assistant can use it. Nothing to write, and it keeps working when the vendor changes their API, because the vendor maintains it.
Where there is no server to connect to and the API is first-party and documented, we write the adapter. Odoo is the one that is built. Never one adapter per shopping platform — that is a treadmill, and it is the reason the first route exists.
Describe an endpoint in your own systems — look up an order, check stock, book a slot — and the model decides when to call it. The description and the parameter schema are the whole interface. A tool failing is not a conversation failing: a timeout comes back as something the assistant can speak about, because the person is still waiting.
Where a system has no API at all, the assistant can drive its web interface after you have signed in yourself, in a browser we host. It is a last resort, it runs in the background, and it never happens during a live call.
We do not host your ERP, your shop or your helpdesk, and we never will. The assistant's job is to reach whatever you already run — which is only possible if we are not competing with it.
Who pays, who can see what, and how your own software talks to it.
Subscriptions grant entitlements and monthly Credits; each variable-cost operation shows its Credit price up front, reserves before running and settles what it actually used. Failed operations release their reservation, and a job delivered twice is charged once. No plan sells unlimited conversation, because the cost of a plan is fixed and the cost of an hour on the phone is not.
An organization holds billing, Credits and members; workspaces separate the actual work — one per brand, client or market — so contacts, conversations and knowledge never mix. Teams are an optional grouping that grants access to workspaces without owning anything. Roles run from owner and billing at the organization down to workspace admin, editor and viewer.
Everything the interface does, the API does too: start a conversation from your own front end, read its transcript, manage content and publishing, read metrics. Keys belong to an organization, can be restricted to one workspace, carry explicit scopes, are stored only as hashes, and can expire or be revoked. Outbound webhooks push what happened into your systems. The Developers page has the endpoints, the scopes and the event list.
Tenant separation is enforced in the database itself, not only in the code above it: a read that belongs to one organization is executed as a role the database will not let see another's rows. Privileged support actions leave an immutable audit record.
The models, the voices, the storage, the carriers and the publishing connections are adapters, chosen per task, per user, per workspace or per organization, in that order. Which provider ran a job is recorded on the job, so a change of provider is visible rather than mysterious.
Two things the platform does that are not the service desk. Both are real and both are secondary — said plainly so you can plan around them.
A full pipeline sits behind the same account: topic, draft, per-platform packages, images, audio and video, four review gates, and publishing or scheduling to connected accounts. It works, and it is deliberately not the priority — the AI agent is. Use it if it suits you; do not choose the platform for it.
Assistants can be hired by the minute, and a seller keeps 70% of the margin on what they sell while the buyer gets 5% back as Credits. The cost of serving the minute comes off the top first, which is why that share is worth more than the same headline percentage of gross elsewhere. Today every listing is ours; the seller side is built and opens to other organizations with white label.
Upload from your computer, scan a QR code and send straight from your phone, or sync a folder from storage you already use. Each asset carries a role — fact source, brand asset, voice reference — and that role decides how it is read. Documents indexed into a knowledge space are what the assistant answers from.
Conversations by channel, by language and by outcome; what was handed to a person and why; what a campaign cost and what came back. Nulls stay nulls — we never show a fake zero, because "no data yet" and "zero" are different facts and only one of them is a problem.