BO Digital Workers · Agentic Sales Rep — turns your website and WhatsApp visitors into buyers.
A visitor who does not know what they need gets a price list and a chat window. The price list cannot ask anything; the chat window cannot work anything out. So the people most likely to buy the wrong thing are the ones left to choose for themselves.
This is an agent that sizes them. It reads the page it is standing on, does the work it can do without being told anything, and puts up a plan with the reason written out — before it asks a single question. Then it waits to be corrected.
The demo sells office coffee, and next door the same agent sells office cleaning for a firm with no basket at all. Not hosting, not software — two companies that do not exist. That is the point: an agent that only sells the trade it was written for has proved nothing.
One agent that knows how to sell and nothing about what. Everything about a particular catalogue lives in one file — an adapter. Swap the adapter and the same agent sells a different trade, with nothing changed above that line.
It opens, reads the page, and runs. No greeting, no menu, no question first. The first thing a buyer sees is a plan and a reason, not a form.
Named steps it moves through itself, with the progress shown. Pause it, step it, jump back. The run belongs to the agent; a person interrupts it.
It says what it guessed and will not go further until someone confirms or corrects it — never mid-run, never on the buyer's behalf. An assumption quietly confirmed is a lie about what the buyer said.
The size comes from the catalogue owner's own sizing rules, the price from their price list. If either cannot be read it says so and hands the buyer to a person. It never guesses at money. It quotes in the shop's default currency — or, for a buyer abroad, the one the shop's own site shows them — and a buyer can switch only to currencies the shop itself lists; it never converts a price.
A returning buyer is not sized from nothing. When it has been wrong the same way twice it proposes a change, shows the evidence from that buyer's own history, and waits. Nothing takes effect without a yes, and one click undoes it.
Every colour is a token with a fallback. Define your own and it adopts them, dark theme included. Define nothing and it still looks right.
Two halves and one frozen rule between them: adapters transform; no business logic.
The panel, the run and its pacing, the quiet status, the findings sheets, the decision card, the approval gate, the hand-over, the receipts.
One page per buyer: best summary on top, dated history below, rebuildable from it. Lessons, proposals, what is in force, and the undo.
What is for sale, the ways in, the checks behind each, the judgement, and which sizing rules and price list to read.
Four things, and the code is arranged so it cannot. It may not choose what to sell — the provider's sizing rules do. It may not work out a price — the provider's price list does. It may not draw anything — the engine does. And it may not decide whether an act is allowed — the manifest does. Break any one and the agent stops being safe to point at money.
Each agent carries a manifest: the acts it may take, and a plain-English list of what it must never do — never invent a price, never promise a date, never claim popularity it has no data for, never ask for a card or a password, never keep selling after the buyer has been handed an exit. The engine enforces the acts list rather than trusting it: an act the manifest does not name is not offered, and the run ends as uncertain and goes to a person.
The way live chat does. Press Start free trial, connect your catalogue — sign in, a key, a link or a file — and paste one line on your site. We host it; you host nothing. Sign-up, the trial and paying all happen in the same control panel as our other agents; the agent itself never handles a trial or a payment. Until that panel's sign-up is live, a short invite-only sign-up stands in for it.
Before anything goes live it tests the connection and shows you what it needs: what is connected, what is missing, and what it will and will not do until that is fixed. With nothing connected it still reads the page, qualifies the buyer and hands them over warm — it just never names a product or a price. Once the catalogue passes, it sells. If the catalogue later breaks, it steps back down on its own and tells you why.
A sale counts for the agent only if your own system recorded it, the agent recommended it to that buyer shortly before, and nothing else already claimed it. Results are shown against your own history, not an industry average. Every figure it states says where it came from, and it never offers a discount — not even to close.
Before the line goes on your site, it is run against sample buyers and you see every answer. It speaks the buyer's language, Arabic and English included, and stops the moment one of your people takes a conversation over.
Its memory and its learning run on our servers, one brain per shop, so it learns from every buyer that shop gets — and what it learns stays with that shop. The same brain runs in our other agents, each in its own lane and on its own accounts. It works on its own, under a shared readiness standard, or inside a control panel, and nothing about how it learns changes.
When a buyer hesitates, is about to leave, or comes back, it opens with a plan and its reason — never a bare hello — once a visit, and one click quiets it. Between two sizes it puts them side by side. If the next plan works out cheaper, it says so, with both numbers from your own price list. You choose how firmly it sells, from informative to promotional; no setting makes it discount or invent urgency.
The shop also says where a decision goes — its basket, its own checkout, its quote form, a call booked in its calendar, or a person — in its own link format. The agent fills in what it decided and never builds a link of its own. With no way to a person, it does not switch on: a sales chat with a dead end is worse than none.
It stays with the buyer on the cart and checkout pages to answer last questions and to catch a plan that does not match what they said they need — before they pay. It never touches the payment form, and it never sees a card.
Written as it stands, not as it would read best. Where there is no evidence yet, it says so.
One thing is settled: it is never free — not even on our own storefront, and not for the resellers running shops on it. It is its own agent, sold on its own: a monthly minimum plus a fee for each sale it is credited with, capped. A sale counts only if your system recorded it within seven days of the agent's recommendation. Tests, rehearsals and hand-overs to a person are never charged. The amounts are not set yet, because none have been tested.
Any business with a catalogue the buyer cannot confidently choose from: tiers that differ by capacity, anything counted per unit, anything with options that change the price. The harder a catalogue is to choose from, the more the agent is worth — and the less a price list does.
There is no market sizing here yet. It would be invented, and an invented number in this section is worse than an empty one.
Live chat and support bots answer questions from documents. They do not size anybody, and when pushed they invent prices — which is the specific failure this agent is built to make structurally impossible. Configurators size well and cannot converse: they assume you already know you need twelve of something. Quote builders face inward, at a salesperson rather than a buyer. The gap is an agent that does the sizing work itself, in front of the buyer, and cannot lie about the price.
Buzinessware first — its storefront is where the agent was born, and it connects the same way as any other shop. Then its resellers, who get a shopfront of their own and have the same problem one layer down, and pay for the agent like anyone else. Beyond that, any shop signs up and connects its own catalogue; nothing of theirs moves.
One adapter in use on a real demo storefront, and two invented trades on this site — the second written by somebody who was shown the rules and not the engine, and it worked first time. Memory and the learning loop are built and tested, and since phase 2 they run on our servers, one shop at a time — a shop can sign up, connect, and paste the line. Today's invite-only sign-up is a stand-in until the control panel's free trial is live. It credits itself with a sale only by a strict rule, against the shop's own history — so the first real numbers will be honest ones. No paying customer yet.
Phases 1 to 3 are done, with 2b and 2c: a shop can switch it on, and it speaks up, compares, books, speaks Arabic, and proves its sales honestly. A shop can now start a free trial and pay through the control panel our agents share. It is packaged, tested in both modes, backed up and sized, and reported ready to go live — going live is the next step, and it is a decision, not a build.
Engine, memory, learning, one adapter, and a demo on an invented catalogue.
Sign-up, catalogue connection with a test first, the what-it-needs screen, the one line a shop pastes — and its own brain on our servers, so it learns across every buyer that shop gets, not one at a time.
Proof of each sale against your own history, the buyer's language, human takeover, sources on every figure, a rehearsal before going live, and a catalogue other AI assistants can read.
Opens at the right moment, compares two sizes side by side, books a call or goes straight to your checkout, and a setting for how firmly it sells. On your cart and checkout it stays for last questions, and speaks up once if the plan in the cart does not fit what the buyer told it — never touching the payment form.
Try it in the demo, press Start free trial, then upgrade whenever you like, any time in the trial — sign-up, plans and billing run in the same control panel as our other agents. When a shop is paused your buyers never see a paywall; the panel simply isn't there.
Packaged the way every agent in the control panel is admitted, tested in both modes, backed up and sized — then reported ready, with the evidence, before anything goes live.
Rehearsed against Buzinessware's live catalogue with every kind of buyer, and with the catalogue pushed until it breaks, including people trying to make it break its rules — before any real buyer meets it. What it learns comes back as suggestions to approve.
The Launchpad: give it your website (or a catalogue link, key or file), let it test itself on your data, review every answer and suggestion, and go live on your yes — then it moves into Settings. You say where you sell and in what currency; buyers pick their currency and data centre, see the regular price with the prepaid one beside it, and are asked how many users or sites a per-user plan is for. Your website is read the way a visitor in your own country sees it, and the Launchpad shows whether that held. The Dashboard: how it is doing, every day after.
One experience in every agent: sign up, choose a plan, the Launchpad (Connect, Analyse, Test, Go live), the Dashboard, one Connect screen for every connection, and Settings come from one shared kit, the same in all our agents. This agent adds only the selling itself and what it reads from your site. By default it never starts a new conversation at night (21:00–08:00 in your time zone), on your public holidays, or more than 200 times a day — you can only make that stricter; answering a buyer is never held back.
Give it your website. If you run on Shopify or WooCommerce it reads your real catalogue itself; either way, within minutes it has your catalogue and is selling on a preview of your own site, with the results on its Dashboard — no setup from you. Upgrade any time — the link is there throughout, not only at the end — and it connects properly and goes live.
A small decision model for its quick yes-or-no calls — what kind of buyer, which need, whether to hand over — used only where it beats today's rules on your own past conversations, switched on by you, and never for a price, a size, a number or a date. Today it runs on a small model trained on your data alone; the hosted model is added when its account is connected.
The agent moves to its own address, sales.betteroutcomes.ai, on a server in the UAE shared with our other agents — its own database, behind one shared secure front door, deployed only when a release is approved, with a nightly encrypted backup and an automatic step back if a new release is not healthy. Its first customer is Buzinessware, testing on a private copy of its site. It starts the way every customer does: its email and its site on the start page, then everything else on the Launchpad.
Before anything else: the whole agent tested end to end on made-up shops and buyers — every way in, every way it can break, people trying to trick it, the full journey from sign-up to going live, and thousands of conversations at once — fixed until everything passes, then run again on every release. In chat it now shows the prepaid price, says when a bigger plan costs less, and gives the total for per-seat plans. A messy price list is no longer sold from: a price of zero or below, a plan that holds nothing, a product listed twice or with two different prices is refused, and the owner is told which product and what to fix. A price a buyer quotes is no longer mistaken for how many they need. A shop whose plans count in Arabic (orders, users — طلبات) now connects and sells, and two such measures are never mixed up in its memory. What a shop has learned and switched on — what its buyers typically need — now shapes WhatsApp answers too, said out loud, never silently. Next round tests the brain itself: does it read a shop right, does it pick right, does it really get better with use, and where the small decision model makes it better — with the numbers. Its memory stays one record per shop, and each shop can open it as linked pages — every plan, buyer, lesson and change, clicking from one to the next. It ends with the whole loop working end to end — from a buyer's message, through understanding, deciding, remembering and learning, back to a better answer — each step proved by its own test — on a hundred made-up shops and every kind of buyer, fixed until nothing fails.
Sign in with a code sent to your email — no password. Invite your team as owner, admin, approver or viewer: everyone can see and ask; approvers approve; admins change settings; only the owner goes live or touches billing. Above every shop sits the Publisher — the one person who runs the agent itself, on its own Publisher page. The same sign-in and roles as our other agents. Codes and invites come from asr@betteroutcomes.ai; setup is one Connect screen, opened with a setup code from the deploy page — the mail password and your email in one go, the same screen as our other agents and never stored anywhere else.
No menu of pages. After go-live your first screen shows what the agent is doing now, with a Pause button; one box to ask it anything; what needs you first; what happened since your last visit; your conversations and hand-overs; and how it is doing. A list opens wide with the evidence on every row. Everything else sits behind the gear, including a record of every action and why. Pause takes the agent off your site within a minute, and the record says who pressed it. We see every shop on one page of our own, the Publisher page — its plan, its use this month, whether it is working — and can pause it too.
One bar across the top of everything: Read, Watch, Try, Set up, Use, Analyse & improve — this page, the demo, the trial, the Launchpad, live work and the Dashboard. You always see where you are, where you can go, and what opens next and why.
Through the whole trial an Upgrade link sits at the top of the Launchpad, on the Dashboard and on the preview of your own site — not only when the trial ends. Upgrading moves the same shop on at once, with everything it has learned. Your buyers never see it.
The shop says where a decision goes — basket, quote or a person — and a third trade is written by somebody who did not write the engine.
For every credited sale, what it recommended against what was actually bought — the same, bigger, smaller — per page and per adapter. Ready; the numbers start with the first real sales after go-live.
Outside AI assistants can ask your agent for a size and a reason, when you allow it — read and propose only, through the same checks as the panel. Messaging: you connect WhatsApp, email, Slack or Teams once in the control panel — your passwords and keys stay there, never with the agent. Messages arrive signed; your team is told apart from buyers; when a person takes a conversation over, the agent goes quiet; every reply goes through the same checks, never in a trial, and a refusal is shown to you with its reason. None is required; the agent sells on your site without them. On WhatsApp the same agent now answers your buyers: it sizes from your own limits, quotes your own prices, sends your own checkout link, says plainly how it read what they asked, and hands them to a person when it cannot size them or they ask for one — then it goes quiet. Next: your team giving it commands in chat. You choose where hand-overs go — your sales inbox or a team channel — in Settings → Channels.
Reads a shop's site and drafts its catalogue for a person to confirm.
The amounts: the monthly minimum, the fee per credited sale, and the cap. And whether a decision model reads what the buyer typed, which is allowed only for typed yes/no and choice judgements, once it beats the current method on past data, and never for a ranking, a number or a date.