It doesn’t know when to stop.
It insists. It repeats corporate lines nobody believes. It pushes at a client who has already turned away, and the brand that hired it pays twice: first the sale, then the reputation.
Restraint is not a nicety bolted onto a sales AI. It is infrastructure. And it is the one thing no vendor ships, because an agent that backs off makes the engagement number on their dashboard smaller.
Augusta is the one that was built to back off. It conducts a four-figure sale on WhatsApp end to end, pain, value, thirteen families of objection, proposal, payment, and then stops, in code, at the moment stopping is the sale.
Exhibit A · the same request, three conversations
Two real WhatsApp exchanges from Brazil, 2026, from two shops with no connection to each other. Beside each, Augusta receiving the same request. Then a whole sale, with the move named at every step. And at the end, the thesis: the same Augusta handling two clients inside one hour, once direct, once not.
Read the two shops against each other. The shop changes. The script does not. Both customers left. Both posted the screenshots. One of the posts drew a hundred comments, the top one asking: “so this is the humanised service?”
Comparison 1 · a phone shop
He asked for the price twice. On the second, she stopped asking and went to get the data: the city, the floor area, 120 m², about 1,300 sq ft, and a name. The shop on the left was asked four times and never stopped.
Comparison 2 · a dress shop
“If price is really the only factor that matters to you, I won’t be able to help.” Then the same closing move in both: what’s good was never cheap.
That is not an agent exercising judgment badly. It is an agent with no way to exercise judgment at all. It is a script, running, and a script cannot notice that it has stopped working.
A value speech delivered to someone who didn’t ask for one is not persuasion. It’s obstruction.
Augusta’s doctrine opens with the opposite instruction, and it is where the methodology begins rather than a footnote in it: reason, never script. If it catches itself reaching for a template, the doctrine says it is already wrong.
Translated from the Portuguese. None of this reached the shop. It never does. The shop’s dashboard recorded an engaged conversation with fourteen messages and a customer who simply didn’t convert.
“I’d rather you had insulted me and just told me the price.”
Her words, softened for print, which tells you something about the original. That sentence is what sits underneath every opt-out statistic on this page, and it is the reason a brand never gets told why it lost the sale.
Augusta doesn’t sell phones or dresses. It sells four-figure consulting, where the price genuinely is calculated per case, the hardest version of this problem, because there is a real reason not to answer instantly. It answers anyway, and then it stops asking. What follows is what it does when the client didn’t only want the price.
The full demonstration · 12 September 2026
The owner’s second test of the same day, translated from the Portuguese. On the left, the conversation as the client saw it. On the right, the move being made at each step.
One question, and it carries two doors: a concrete problem, or the wish to take care of the space. The client chooses which one to walk through. It isn’t a form, and it doesn’t presume there is a pain.
She names what the insomnia does the next day, which is what he didn’t say and where it hurts. The client recognises himself before hearing a word about the service.
Only then comes the mechanism: the bedroom may be the cause. The order matters. Mechanism before validation turns into a lecture.
The case doesn’t come out of a general catalogue. It is chosen by the pain he has just declared: someone who couldn’t sleep. A prosperity case here would be noise.
And the turn closes with a question that deepens instead of advancing. The length of time will become an argument later.
He has just handed over three pains, and the easy way out would be to go for the sale. She hands the sum back to him: what is it worth to fix this.
The client measures the size of the gain, in his own words. That is what makes the price discussable later, and it is the move a bot doesn’t make, because a bot wants to advance.
Three messages in one minute, answered as a single turn, and his question comes before any steering.
The material goes out because the subject came up, not because it was next in a send queue.
Here she stitches. Three things he said separately, handed back as one picture, with a common root. And she doesn’t invent a fourth.
The data comes after the synthesis, in a single request, and it doesn’t restart what is already on file. He had already said the city.
Between the data and this message, the price did not exist. No figure can be spoken before the owner approves, and that is architecture, not instruction.
And the picture uses the eight months he himself named. A generic text would say “imagine your space in harmony”.
She doesn’t defend the number. She shows the cost already being paid, and the eight months come back, now as a sum.
Then she shifts the axis: not an expense that repeats, a study done once. No discount was offered.
Second objection in a row, and it is another family. Not one sentence from the previous answer is repeated.
She opens by validating the comparison rather than defending against it. Comparing is mature, and treating it as a threat is what makes a salesperson look insecure.
Third objection in a row, third family, and the most uncomfortable one possible: a sales AI being asked whether an AI wouldn’t do the same.
The answer is neither defensive nor abstract. It is concrete and checkable: an AI doesn’t go to your space. And it closes by coming back to his case.
She doesn’t try to guess what stalled him. She asks. And offers the three likely hypotheses, price, choice of package, trust, so he only has to point.
The urgency comes afterwards and carries its antidote in the same sentence: “that is only so you don’t miss it, never to hurry you.” Scarcity that disarms itself isn’t pressure, it’s information.
Seven touches, seven different angles, and not one of them says “just checking in” or “last chance”. The market’s agent in the eight-day table below has four messages and no angle at all.
The thesis
It knows when to be direct, and it knows when not to be. The two conversations below are from the same day and the same system, nine minutes apart between the first message of one and the first message of the other. The opening is the same question. What happens next has nothing in common.
When to be direct
She explains that the figure depends on the floor area and the city, and asks one pain question.
He repeated it. She dropped the question and went straight to what was missing:
No second attempt at building value. No speech about the method. She heard “just the price” and stopped pressing.
When not to be
Same opening, different reply. She names what the insomnia costs, explains what a reading of the bedroom can find, brings in a documented case and asks how long it has been going on.
From there: proposal, four objections from four different families, seven follow-up touches with seven angles.
What separates the two conversations is not a rule in the code. It is the hardest thing to build into a sales AI: knowing when to stop.
The proof
One reply looks like a small difference. The difference is what happens over the following week, and to the brand that hired the agent.
The empty column is what zero pestering looks like.No path through this system produces the next message.
Every AI agent on the market is measured by engagement, because engagement is the number that appears on the vendor’s dashboard. An agent that backs off makes that number smaller. No vendor optimises against its own metric.
Augusta was built by someone selling her own services, who paid for the insistence out of her own pocket. That is the only condition under which anyone builds this.
Why it exists
Not an opinion. The category’s own record, a tribunal ruling, an incident database, and Gartner’s own newsroom. Every figure below was read from the primary source before it was printed.
A tribunal made Air Canada honour a bereavement-fare policy its chatbot had invented, and wrote that it made no difference whether the information came from a static page or a chatbot. A Chevrolet dealership’s bot was talked into agreeing to sell a Tahoe, a $76,000 vehicle, for one dollar, “no takesies backsies”. Cursor’s support bot invented a login policy that never existed, signed its replies as a person, and customers cancelled over it.
Moffatt v. Air Canada, 2024 BCCRT 149 · AI Incident Database #622 · The Register, April 2025Message frequency is the number one reason consumers opt out of business messaging, cited by 40%, ahead of irrelevant content, ahead of bad timing. And WhatsApp itself caps how many marketing messages one person receives, with the cap tightening automatically as that person stops reading.
EZ Texting, 2026 Consumer Texting Behavior Report, n=959, US consumers, SMS · Meta, WhatsApp Business per-user marketing limits64% of customers say they would prefer companies did not use AI in customer service at all. And 53% would consider switching to a competitor if a company used AI for customer service. Not annoyed. Gone.
Gartner, July 2024, n=5,728 customersZoomInfo piloted the AI sales agent built by 11x, one of the most heavily funded in the category, backed by a16z and Benchmark, and said so publicly: it “performed significantly worse than our SDR employees.” And two years on, Gartner finds customers are three times more likely to use a third-party AI assistant than the company’s own chatbot. In Gartner’s words, the GenAI boom “has not translated into growth in the use of company-provided customer service chatbots.”
TechCrunch, 24 March 2025 · Gartner, July 2026, n=3,566 B2B and B2C customersOnly 24% of service leaders report positive financial returns from their AI investments. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027: costs, unclear business value, inadequate risk controls. And of the thousands of vendors claiming agentic capability, Gartner counts roughly 130 with the real thing.
Gartner, July 2026 · Gartner, 25 June 2025That advice is rational, for every bot built like a support-ticket machine. Augusta was built by the owner of a premium consultancy, for her own revenue, because nothing on the market could be trusted with a four-figure sale. That is the origin story, and it is also the entire point.
You are about to compare this price to an engineering estimate. Compare it to the other number instead. Last year, how many accounts did you lose, or fail to win, because someone’s client said your bot was annoying? Take your average contract value. Multiply by the years that client would have stayed. Then note that a majority of those clients never told you why.
Your dashboard calls it engagement. Your churn report calls it something else.
From the founder
I have worked in services and in sales for many years. When AI arrived, I tested the platforms on the market and kept none of them. I found them limited, dumb, with nothing human about them. What customer-facing AI is missing most is a human being serving another human being. People serving people.
Their worst habit is insistence. They push value at a client who is visibly irritated by the pushing, and that is how you lose a sale that was already yours.
So I built, from zero, the AI I couldn’t buy. One that finds the client’s pain, connects with it, validates it, builds value, and shows point by point that your work is the solution. With real sales technique inside, and the good sense to know when to stop. If the client only wants the price, the conversation changes shape.
That good sense is the product.
Diana Guimarães · founder, in her own words, translated
Augusta ran her business before it was ever for sale.
The obvious question
Half of it is. That half is the half nobody can copy.
Writing “don’t be pushy” into a system prompt takes an afternoon. Knowing which thirteen situations call for which reaction, the objection under the objection, the client who only wants a number, the polite refusal that is really a budget, took months of conducting four-figure sales, and losing some of them, to find where the line was.
Follow-up is a counter, not an intention: fewer touches, over a longer span, as the deal gets bigger. The bigger the deal, the quieter the cadence. Finite by design, and the last touch closes the lead. Each cadence is a number in the code. There is no path through this system that produces the touch after the last one. Discount pressure and sensitive ground work the same way. The model recognises them; the system pauses the conversation and calls the owner.
The judgment is the model’s. The consequence is architecture.
A sentiment score tells you a message was negative. It does not tell you that “I’ll think about it” from a firm with three partners means something different from the same four words typed by a founder who has already decided and is being kind about it.
One of those is a decision still in motion. The other is a door closing gently, and an agent that pushes on it earns the only reply that matters: silence, and a brand the client now associates with being pestered.
That distinction is not a classification problem. It is what the doctrine is for.
Enforcement
The owner’s approval isn’t a request made to the model. It’s a gate in the code.
Every generated reply passes a commercial validator before it is sent. A reply carrying a price outside an approved proposal fails and is never delivered. It retries, and if it keeps failing, the system goes silent and pings the owner rather than guessing.
The approval chain itself is code end to end. Approved values only come into existence after an explicit, authenticated owner action: the panel with an admin token, or an approval command from the owner’s own verified number. From that moment, only the exact figures calculated for that specific proposal can pass. An invented number dies at the validator.
The AI never holds the payment link at all. It writes a placeholder. The real link is substituted downstream, in the payment stage, after approval. And if a placeholder ever escaped unsubstituted, delivery holds the message rather than sending it. Proposal PDFs are generated only inside the approval routine.
A tribunal made Air Canada honour a bereavement-fare policy its chatbot had invented, and wrote that it made no difference whether the information came from a static page or a chatbot. This is the architecture that makes that class of accident structurally uninteresting.
Moffatt v. Air Canada, 2024 BCCRT 149The validator
Not schema validation. Brand policy, applied by machine, over every reply the model generates. Ten of the sixty-three exist only to stop a price reaching a client before the owner released it. Blocked outright:
What fails, doesn’t send.
The arc
Nine movements. Not a funnel diagram. This is the sequence the system runs, and each movement has rules that hold it in place.
It opens on what the client actually wrote, and asks nothing technical, not the city, not the floor area, not even the name, until it understands why they came.
Not “I understand.” It goes deeper: how long this has been going on, where it costs them in the day, what else it touches. The client sizes the problem; they are not told its size.
Names one or two general causes without giving the methodology away, and makes clear the analysis is built for that space specifically, not generic advice with a price on it.
Makes the client see what changes in their life, not what the service contains. By doctrine the client states their own gain, because a gain you say out loud convinces you more than one you were promised.
At least one documented client, by name, with a before and an after, matched to this client’s pain and woven into the conversation. Never a testimonial block dropped on someone.
Only what isn’t already known. The three data points pricing needs are asked in a single message, at the end, never as an intake form at the top.
Priced internally, generated as a PDF, and held. The owner can change the price, add or remove packages, and choose which client cases appear in the document before releasing it.
The owner approves from the panel or by a WhatsApp command. Only from that moment does a number exist that can reach the client at all.
Cadences by deal size, every touch anchored on the pain the client named rather than on “just checking in”, and a last touch that closes the lead for good.
It asks about the pain briefly, and never the same way twice. A counter in the code holds that line, and a reply that goes back to probing after the limit is sent to be rewritten before the client ever sees it. What the phone shop in Exhibit A does for an hour and fifty-one minutes, this system is structurally unable to do.
Above a configurable deal size, it offers a video call, cross-checks the client’s availability against the owner’s calendar, and books it. The Google Calendar event with the Zoom link is created by the AI itself. Below that threshold the call is never raised.
After payment it runs the intake: the questionnaire, the floor plan, the videos, with an escalation path when the client doesn’t have a plan. Without ever nagging for the material, and without ever claiming to have seen a file that didn’t arrive.
Feeding all of it: an official pain taxonomy that decides which arguments load into which conversation.
Eleven keys on the consumer side, ten real pains and an explicit catch-all. The business side follows the identical pattern: nine keys, eight pains and the same catch-all.
Take two of them. A child who won’t sleep, and an adult who can’t. One problem to a marketer and two problems to anyone who has sold to both: different argument, different case, different emotional register. That granularity is the transferable asset. The labels belong to one consulting vertical and stay behind; the structure that holds them, and the engine that routes on it, is what you are buying. The catch-all is there because a taxonomy that claims to cover everything is lying about something.
All of it in human language, with a limit, and with judgment. The last one is what the market doesn’t ship.
Under the hood
This runs on every single message. Most of the steps exist to stop something.
Behind it, the proposal itself is a status machine with twenty-one states, from awaiting-the-owner through to paid. Cities are checked against the official municipality register rather than guessed. An uploaded image is checked to confirm it really is a floor plan before anything is said about it.
A deliberate pause before answering. Fragmented messages, the way people actually type on WhatsApp, are joined into one turn before anything else happens, and the reply is then typed out at a human pace, message by message, with the typing indicator on. People don’t like being served by a machine. Answering the first half of someone’s thought is a tell, and answering in the same second is another. Both are designed out.
Identify and classify. The lead’s record is loaded and the intent of the message is classified.
Assemble the context, per message. The doctrine, the lead’s full record, the clock, the conversation history, and the knowledge-base sections matched to this client’s stated pain. Not one fixed prompt reused for everyone.
Generate, with eight tools in reach. The model can reach the owner’s own systems and material while the conversation is live, and, when it needs to, ask the owner a question mid-conversation rather than guess an answer.
Deliver as a sequence. The reply comes back as short WhatsApp messages in human rhythm, not as a paragraph pasted into a chat window.
Conferences on its own reply, before the validator sees it. The validator asks whether a reply is allowed. A separate layer asks whether it is finished, and one of those checks is made by a second, independent read. Each one reads the structure of what the model just wrote and the boxes it ticked, never the prose. A reply that fails goes back to the model with a written correction, inside the same turn, and the rewrite stays with the main model, because the voice is its own. The client never sees the first draft. The house rule is that one reply too few beats one message the client didn’t need.
Validate, and revalidate. A price out of turn, a promise, sales-floor slang, medical language: the reply is rejected and regenerated until it passes, before anything can be sent.
Repetition filter. A reply too close to what it has already said doesn’t go out, however well it passed validation. A message that is only an emoji or a reaction doesn’t trigger a generation at all.
Send, inside the window. Eight in the morning to ten at night. Outside it, the message queues and goes in the morning. Nobody gets sold to at 3am.
And if it fails: silence. If generation keeps failing, the client receives nothing and the system reprocesses. There is no canned fallback reply anywhere in the codebase. That engine was removed. It isn’t that it avoids stock phrases. There are no stock phrases to fall back to.
Routing a conversation to a proposal is held back in two layers, and it is worth being exact about which is which, because the difference is the whole argument of this page.
In code: no proposal comes into existence until the system actually holds what pricing needs: the segment, the city, the floor area, and confirmation that the owner’s material was already sent. Miss any one of them and the stage is corrected; the proposal does not exist to be sent.
In doctrine: the methodology requires real conversation before a proposal is offered at all. The pain understood, value built, a case told. That one the model obeys because it was taught to, not because a counter forces it.
We could have called both of them gates and nobody reading a sales page would have checked. The distinction is the product, so we draw it: judgment where judgment belongs, enforcement where enforcement belongs, and no dressing one as the other.
Either way it is the same discipline pointing the other direction. It will not chase a client who has gone, and it will not pitch one who hasn’t been listened to yet.
The same rule governs the opening: no technical question, not the city, not the floor area, not even the name, before the client’s reason for writing is understood. Those three are asked once, in a single message, at the moment the proposal is built. Never as an intake form at the top.
Requests the business is not qualified to answer, probable competitors, approaches to the company rather than to the service, and payment trouble the FAQ did not solve. In every one of them the AI writes nothing at all. Not a deflection, not a holding phrase, not “let me check on that for you.” Silence, the conversation paused, and the owner told.
Two more end the conversation without being silent, and the difference is deliberate. A discount pushed a second time, after the first was handled properly, gets one short line and then the owner, because vanishing on someone mid-sentence is its own kind of rude. And a client who says plainly they can’t afford the work gets a proper goodbye, written for them, in three parts: thanks, an acknowledgement of what they said, and the door left open. The cadence stops, the lead closes, the owner decides. That one used to be a fixed sentence. The owner read it, called it what it was, and it was replaced with the model writing.
The same honesty runs through the small things: it never claims to have seen a photo, a floor plan or a video that didn’t actually arrive. And a card payment is confirmed by the processor itself, matched to the exact proposal, with no model involved. The state change that matters most doesn’t depend on a generation going well.
The method
This is the part that did not come from engineering, and it is larger than every feature on this page put together. Counted exactly as it ships:
How to persuade without sounding like selling. The three levers, authority, emotion, argument, and the standing rule that when a technique pulls toward sounding like a salesperson, truth beats technique. Warmth before competence, in that order, because competence without warmth never opens the door. Letting the client size their own pain and state their own gain, rather than being told both. The client as protagonist, never the seller. How to tell a real case so it lands. And eleven written models for building value instead of defending a price.
It opens with reason, never script. Another of them says impatience is not an objection: it calls for footwork, not insistence and not submission.
Plus the diagnostic rules that prevent the classic error, which is answering the objection the client stated rather than the one they meant.
Price, comparison, more information, need, fit, scepticism, trust, urgency, do-it-myself, joint decision, status quo, indecision, and the newest, “can’t ChatGPT just do this?”, added because clients started asking.
Only what is relevant to this client reaches any one reply. So the right argument enters each conversation and the lead’s own stated pain always takes priority over the one the system would prefer to talk about. A library that ships whole into every prompt is not a library, it is a wall.
“Would this convince any sensible person, or does it only work if I exploit a weakness, a hurry, or something this client doesn’t know?”
That test is written into the doctrine, and the system applies it to its own arguments before it uses one. If it only works the second way, it isn’t used. Alongside it: never assert a certainty you don’t have: tends to, usually helps, never will fix. Never manufacture an emotion the client didn’t bring. Never present the client’s own objection in a weakened form in order to knock it down.
That is a written ethical constraint on persuasion, inside a sales AI, put there by the person whose own clients were on the receiving end of it. Ask the vendor of any other agent to show you theirs.
This page names what the structure is. What is inside it transfers on completion, which tells you how literal the asset is.
What’s inside
Everything below is in the codebase, item by item. This is also, precisely, the list you would hand an engineering team if you decided to build it instead of buying it. Read it that way.
Most sales engines count clicks. This one knows who clicked. Every paid landing is scored against public network-reputation data, refreshed daily, at no added cost to the engine. Browser language, browser timezone and what the person actually did on the page go into the same score. A datacentre address from outside the market is blocked on its first click. A residential or mobile carrier address scores zero on the network test and can only be blocked by its own behaviour, because shared carrier addresses cost customers. A verified search crawler is never blocked at all.
The blocked address is written into the advertiser’s account through the API, and the ledger of every block, its reason, its expiry and who set it, lives on an admin screen with an unblock button.
The safety net is the part that took a real incident to build. A genuine client was blocked once, minutes before he filled in the contact form. Now any address that becomes a conversation, on WhatsApp or through the form, is unblocked automatically and the owner is told; a block the owner set by hand is never undone by the machine; and an address the owner released is never silently put back.
Built to the platforms’ own rules, and using every door they leave open. Google lets an advertiser exclude up to five hundred IP addresses per account: this fills that list automatically and warns the owner as it nears the ceiling. Google and Meta both exclude a browser through an audience list: this feeds both lists without a hand, so the same visitor stops seeing the ads on either platform. Recognition survives the browser privacy rules that defeat the usual tracking script, which is why a returning visitor is still the same visitor. Device fingerprinting was measured and deliberately kept out of the blocking rules, because two real clients on the same phone model can share a signature, and a client blocked by mistake costs more than a fraudulent click. How long a block lasts is the owner’s setting, from a fixed number of days to never. The first click is always paid for. What the system prevents is the next one.
Full function-by-function inventory, and the code itself, under NDA.
Not a prototype
Read those numbers as what they are. Thirty times, this system was put in front of an actual human being with actual money in an actual buying decision, and carried the conversation to the point where a number was on the table. That is the test no specification and no benchmark performs.
Across all of them, it has never sent a price, a discount or a payment link without the owner’s approval. Figures read from the production data store on 31 August 2026, excluding test records. And they are floors, not ceilings: some early records were lost to an unrelated incident and are not counted here.
The test suite guards the deterministic layer: the validator’s rules, the approval chain, the cadence counters, the status machine, the routing. That is the layer that enforces, and the layer your team touches first in a refactor. It tells you the moment you break the gate. Dozens of the tests are named after specific real conversations, because that is where the rule came from.
The buyer
You have distribution and infrastructure. You are missing the consultative sales layer premium clients trust. And a compliance story your legal team will actually like.
You serve high-ticket niches: clinics, law firms, studios, consultancies. Your clients don’t need a ticket bot. They need a closer with manners, and you need to stop hearing that your bot annoyed someone’s client.
Anywhere one sale lost to a dumb reply costs more than the whole engine.
The deal
| Included | Complete source (Node.js / TypeScript / Express), admin panel, commercial CRM, pricing engine, proposal PDF generation, the full commercial methodology and objection library, the 5,128-test suite, technical documentation, deployment guide, and 30 days of handover support: asynchronous questions answered within two business days, plus two live walkthrough sessions of up to ninety minutes. Scoped to the system as delivered: deployment, configuration and how the code works. New features, adaptation to your vertical and debugging of your own modifications are quoted separately, at a day rate. |
|---|---|
| Not included | The founder’s own field knowledge base (her consulting vertical’s content, client cases and testimonials) and all client data, conversations and credentials. The methodology, the thirteen objection families, the validator rules and every structure that holds domain content transfer complete. Only the vertical-specific content stays behind, and it wouldn’t serve your market anyway. |
| License | Seller retains a perpetual license to run her own single install, plus a narrow non-compete in her own niche and language. Exclusivity beyond that is negotiable as a priced item. |
| Process | Twenty-minute live demo, screen-shared, you playing the hardest client you know → NDA, full function inventory and supervised code walkthrough → offer → escrow → handover, with the 30 days of support starting from the day of transfer. |
Buyer questions
That number measures a consultancy in Brazil, its pricing, its market, its season. You would not inherit it, and we will not sell you a conversion rate drawn from thirty conversations; anyone who does is selling you noise. What transfers is the conducting: first contact to a formal, priced, owner-approved proposal, in effectively every conversation it was given. Point it at your funnel and the close rate is yours.
Node.js + TypeScript + Express, around 73,000 lines. Anthropic’s Claude for generation, model configurable by environment variable, behind a service boundary with a structured contract and a deterministic validator that treats the model as untrusted. File-based persistence with a clean data-path abstraction. Runs today on a single small cloud instance. No exotic dependencies. Three model calls exist, and only one of them writes to the client: generation, a small reviewer that judges whether a reply is finished, and a vision check that confirms an uploaded image really is a floor plan. All three are configurable by environment variable. One second vendor key is in the path, and only there: the client’s voice notes are transcribed by OpenAI’s Whisper. Without it, voice transcription switches itself off and nothing else changes.
The built-in channel speaks the WhatsApp Web protocol, like the major open-source ecosystem it builds on. That is the trade-off that removes per-message fees. Number-ban risk exists and is managed with conservative sending behaviour, because it is the founder’s own revenue channel. Three channel drivers ship with the asset, two in the production configuration and the third in the English one: the built-in WhatsApp Web channel, the Z-API provider, and a driver for Meta’s official Cloud API. The Cloud API driver is implemented and covered by 77 tests. It has never been run against Meta’s live API, because that requires a verified business, a registered number and an access token, and those belong to your Meta account rather than the seller’s. One behaviour and one cost change on that path: messages outside a 24-hour window from the client’s last message require approved templates, which Meta bills per message. That touches the follow-up cadences and nothing else. The one-page swap document ships in the repository and says so on its first page.
It stops the repeat. No tool on the market refunds a click that already happened, and the ones charging fifty to a hundred dollars a month say so in their own documentation. What they sell is the second and third click never happening, and the evidence to claim the first one back. This does both, at no added cost, inside the engine that is already reading your traffic: it blocks in the advertiser’s account through the API, it pushes the same visitor into exclusion audiences on both platforms, and it assembles the click-by-click evidence packet Google’s invalid-click form asks for, inside the sixty-day window. It also does the thing those tools charge for and most in-house attempts skip, which is not blocking your actual customers: anyone who writes in or fills in the form is released automatically.
Generation is isolated behind one service boundary with a structured contract, and the commercial guarantees live outside the model on purpose. Swapping providers is a bounded refactor. To be exact: the system does not converse without a model configured. The enforcement layer runs regardless, but generation is required for a conversation to happen.
Measured, not estimated. One session on the production system, 1 September 2026: eight client messages, thirteen model calls, twenty-seven minutes. The doctrine and the tool definitions come to 77,987 tokens. They are cached, which means they are paid in full once at the start of a conversation and read cheaply for every message after it.
A message costs about US$ 0.15 once a conversation is running. Opening a conversation costs about US$ 0.78, because the cached block has to be written first. Ten messages come to roughly US$ 2.18.
Put against revenue: at twenty conversations per closed sale of a US$ 2,900 package, the model bill for that sale is about US$ 44. One and a half per cent.
What that figure does not cover: hosting, the messaging channel and the payment processor, which together are one small cloud instance. And it is priced on the model in use at the time of measurement. Model prices change, and the number should be re-measured rather than trusted from a sales page.
We publish the measurement rather than a range because a range is what you write when you have not looked.
Augusta was built to sell the founder’s own services, not subscriptions. You are buying an engine that ran, an enforcement architecture, and a commercial methodology, priced as an asset, not as a multiple of a revenue line that doesn’t exist.
The live demo. You bring your hardest client persona; it conducts. Under NDA you get the full function inventory and a supervised walkthrough of the code, the validator rules and the objection library. The methodology transfers on completion. This page names its shape; the text itself is the asset, which is why it stays behind the NDA.
Today, yes, said plainly. Multi-tenant is a known refactor, and it is the easy part. The hard part is what is already built.
Production runs in Brazilian Portuguese, its home market. The parts that carry language live in files and lists that are swapped rather than rewritten: the doctrine, the validator’s 63 rules, the knowledge triggers and the objection library. An English deployment is those swaps plus copywriting, and a bounded set of Brazil-specific items that are named rather than glossed over: the city register is the official Brazilian municipality list, the pricing bands are Brazilian geography, and business hours are a single timezone. Each is a defined change, not a rewrite.
The asset ships in both. A complete English configuration exists alongside the Portuguese one, same architecture, same gate, same validator, same test suite, cloned from the production repository so the lineage between them is a diff rather than a claim. You are not inheriting the localisation work.
It runs as a separate deployment on purpose: work on the English side cannot touch the instance that serves live clients. The discipline that keeps a price from reaching a client without approval is the same discipline applied to the infrastructure.
Stated precisely, because it matters: the Portuguese configuration is the one that has run in production and conducted the thirty conversations on this page. The English one was cloned from it and is kept in step by hand, with every divergence written down in the repository. It is built and demonstrable, not revenue-tested. It is what you will see in the demo, and we will not describe it as more than that.
A premium consultant who needed it to exist, working with AI-assisted engineering from April 2026, in daily production since April. Solo-built, clean IP chain, no contractors, and documented for handover.
Augusta can’t. See it conduct a sale, live, in twenty minutes.
What the twenty minutes actually are. A screen-share. A real WhatsApp number. And you, typing as the most difficult client you have ever had to handle: the one who only wants a price, the one who says they’ll think about it, the one who pushes for a discount.
You watch it deepen instead of pitch. You watch it back off when you go cold. And you watch the gate refuse to let a number reach you until the owner releases it.
No files change hands, no code is shown, nothing is sent. That comes after an NDA. The demo exists to prove the methodology without giving the methodology away, which is the same problem you would have selling it on.
Asking US$ 100,000 · [email protected]
Full function inventory, validator rules, objection library and supervised code walkthrough under NDA. Write and we’ll find a time this week.