The AgenticSelling Manifesto

For twenty years, using a website meant translating yourself into a machine’s terms.

You arrived with a need you could have said in one sentence (a coat that will last, a knee that hurts after a long walk, a payroll that has to cover two countries, a lawyer who has argued this kind of case before), and the site handed you a category tree, a keyword box and a wall of filters, then asked you to do the compiling.

We called that convenience. It was, in fact, a tax paid by every guest on every visit, for one reason: the machine could not understand a sentence.

Language models lifted that tax, and they lifted it in three ways at once: the machine now reads our language, it has begun to buy on our behalf, and it can compose a page while a person is reading it. Together, those three are rearranging buying and selling from the inside.

The work of translation has changed sides: every interface built on the old arrangement (the dropdown, the filter, the ranked grid) was a structure the buyer maintained for the seller’s convenience, and whatever replaces it has to earn its place by understanding. A seller who still asks a person to speak in categories is charging a tax that no longer has to be paid, and we need to understand that almost everything written about this concerns the buyer. In fact, the seller has no discipline yet, so this is our attempt at one.

The intermediary has ancient roots: the etymology of the word comes from the Persian sapsar and referred to the figure who connected buyer and seller. Arabic carried it along trade routes, reshaping it into simsar; Italian received it as sensale and applied it to every kind of broker, from agricultural commodity trading to those who, in a now-distant era, arranged marriages. Every age reconstructs this figure, updating it to the technology of the moment, and always asks it the same question: whose side are you on? Software doesn’t change the question.

What has changed is the audience, because a seller now has two listeners: one is a person, moved by story, atmosphere and attention; the other is a machine, moved by structure it can verify. They must be answered from the same truth, since a seller with two versions of itself will be caught by a reader that checks. For now, answering both takes two crafts, and we expect them to merge as buying agents learn to question a brand’s own agent, ask it for evidence and negotiate with it, and as that agent learns to answer with real context on the brand.

The second listener is unforgiving in a specific way, as buyers meet a seller inside an answer before they ever reach its site, and every answer is a verdict: an agent applies one mandate across more options than a person can hold in mind, and it rewards only what it can parse. A real difference that can’t be read by a machine retains its full value in a human buyer’s eyes...but has none whatsoever for the agent. This is why making an offer legible is not a technical detail: it is what lets a seller’s real quality reach the buyer looking for it.

There’s a loss that leaves no trace - when an AI agent finds a buyer a good coat, she’s happy, never knowing there was a better one the agent simply didn’t show her; the better coat stays on its hanger, and no receipt records what she missed.

The seller has it worse: nobody tells them that a real buyer searched for exactly what they make, that an agent visited their site, missed one small detail, and left: no abandoned cart, no failed checkout, nothing to measure or fix, just a sale that never had the chance to happen.

It’s clear that a market whose losses appear invisible cannot correct itself through the disappointment of those who buy - it needs evidence: claims and confirmations a machine can verify in the exact moment they are made, records a third party can check afterwards, and a brand that takes responsibility for what its agent communicates. Sellers tend to read this verification as compliance, when in fact it is the opposite: where nobody feels what they lost, verification is how quality still earns its price.

Some of the buying will never be delegated. For a printer cartridge, for example, or a software renewal, the want is settled before the search begins, and the search is pure cost. For a coat, a wine, a room for an anniversary or an agency for the year ahead, the want is assembled while looking, and the looking is where the value is made. Delegation takes the labor and leaves the judgment with the person, because a taste inferred perfectly still returns an object nobody chose, and for much of what we buy we want to have been the one who chose (a gift carries some of the time spent choosing it). The seller’s problem now is that both buyers arrive at the same door.

For the buyer still choosing, attention is what a brand sells. Elements such as the window, the fitting room, the remembered name, the founder who runs the demo, the agency that calls back on a Friday: all of it stages wanting, and wanting is most of the pleasure.

AI assistants have understood this by accident, as with no warmth and a plain interface they listen, answer the question that was asked, and stay through the fifth question without tiring; then the person who has been attended to for an hour arrives at the brand’s own house, where a page of prepared sentences turns them back into a stranger. This way, the brand becomes the only party in the conversation that stopped listening.

The best salesperson has always been the one who understands customers and rearranges the sale around them: this is the highest form of selling, though it never scaled.

Now you can get pretty close: answering the sentence a person actually wrote is a different act from showing the same fixed page every time, because prepared interfaces meet only the questions somebody anticipated and, honestly, what a person asks next is almost never one of those. A generated product card that can handle the unexpected question, and the hesitation that follows it, stops being a page and starts behaving like that salesperson who always seemed to know already what you needed.

A page that composes itself has to be governed, because it can also manipulate, drift and fabricate. The materials are the brand’s own and every claim carries a source; order, emphasis and depth can vary, while the offer itself (its facts, its terms, its safety information) stays fixed. What one person was shown has to be recoverable later by someone who was not there, or generated commerce cannot be audited, and since no widely adopted convention of that kind exists yet, it is the standards work we would put first.

The same discipline governs action: an agent on live systems needs a mandate before it needs intelligence, with bounded actions, a record and a way back (for the most part it observes, proposes, acts with approval, and only then acts alone).

We call this discipline agentic selling, and the one that practices it a brand agent: one agent per brand, carrying that brand’s mandate and no one else’s. It studies the market, learns continuously from the brand, its channels and its competitors, then sells in conversation, answerable for every word: it is one part of what we build, alongside the models that compose interfaces, the shopping agents that buy on a person’s behalf and a protocol for speaking to agents. Within five years we expect most brands to run an agent of their own, as ordinary as a website is today and for the same reason: without one, a brand is described by whoever has one.

Underneath all of it is agency in both senses of the word, the capacity of software to act and the capacity of people to remain the authors of their choices. While the first is being engineered with extraordinary care, the second is nobody’s job yet - which is why our research led us to build what we call the “Brand Identity Protocol”.

Buying is among the most human things we do: what we choose, how we look for it, whom we trust to advise us, what choosing feels like.

For twenty years we accepted that the surfaces we bought from could not understand a sentence, because that was the price of the technology, and the technology has changed: the machine now reads and composes. It can attend to us with a patience no shop could ever afford, and it can act for us with a consistency we do not have (including the consistency of showing a million people the same three answers). We will know what got built by counting how many sellers stayed distinguishable, and how many decisions people still made themselves.