The new buyer doesn't ask on WhatsApp.

On 30 July 2026 Google announced that Gemini Spark could take over the user's desktop Chrome and complete errands on its own, and on 3 August it began rolling that capability out in the United States^[1][2]. The agent uses “your signed-in accounts and saved passwords” to resolve tasks such as scheduling an apartment viewing or starting a flight booking^[1]. The company promises to hand back to the user “sensitive actions, like payments” before executing them^[1]. The conversation that followed dealt almost entirely with that last phrase, that is, with who authorizes the spending when the buyer is a program.

That conversation skipped the step before it, which is the one that decides who ends up inside and who ends up outside the commerce that is coming. An agent does not walk down streets or enter into conversations. It chooses from within a catalog, and to appear in that catalog a business has to deliver a structured data file that somebody validated in advance. The larger phenomenon of which this launch is today the best available example is that the condition for selling is shifting from having premises to having a feed. Latin America is the region worst placed in the world to meet that condition, because nearly half its employment is informal and its counter tends to be a private conversation.

How an agent decides that a business exists

The standard that governs that door is called the Agentic Commerce Protocol, published by OpenAI and Stripe under an Apache 2.0 license, with a feed specification that is explicit about what it demands^[3][5]. The business registers at chatgpt.com/merchants and pushes a file in TSV, CSV, XML or JSON “over encrypted HTTPS to the allowlisted endpoint”^[3]. Before going live it submits a sample “so that our indexing team can validate the parsing”, and from then on it can update every fifteen minutes^[3]. None of that is a metaphor for formality. Formality is written into the schema: among the mandatory fields are seller_url, with HTTPS preferred, and seller_privacy_policy alongside seller_tos when payment is enabled^[3].

The door does not end at the file. The business must also implement the Agentic Checkout endpoints, so that ChatGPT creates and updates the purchase session while the seller validates, determines delivery options, calculates and charges tax and finally accepts or rejects the order^[4]. It also needs a payment processor compatible with the delegated payments specification^[4]. Meeting those three layers presupposes an incorporated company, with its own domain, published legal documents, tax accounting and somebody capable of operating and maintaining an application programming interface.

The search engine used to find you; the agent has to be gone out and found

The difference from the previous era is one of direction, and it explains why this resembles no earlier digital transition. A search engine discovers by crawling. It travels the web on its own, indexes what it finds and includes you even if you never did a thing about it, so that appearing is the default position and disappearing demands the deliberate gesture of blocking it. An agent's catalog works the other way around. Nobody is going to crawl your business in order to take you in: you are the one who registers, who pushes the file and who waits for an indexing team to validate that your format can be understood^[3][4].

That change turns visibility into a piece of paperwork. Paperwork has a property well known in the region, which is that it filters by administrative capacity before commercial merit. The seller who does the job well and does not complete the procedure ceases to exist for the new buyer, not because of a decision taken against them but because the system contains only those who enrolled. The default option moved from being inside to being outside, and that change alone reassigns demand without anyone announcing it.

Getting paid did get solved, and that is why the other gap shows up so clearly

It is worth saying in so many words that the financial half of this problem is ready in the region. The contrast is the argument. Mastercard introduced Agent Pay in Latin America and the Caribbean in December 2025, and on 25 March 2026 announced live, end-to-end agentic transactions completed across the region^[7][8]. Seventeen institutions executed those operations: Banco de la Nación Argentina, Banco Galicia, Banco Ripley and Banco Security of BICE; Bancolombia, of Grupo Cibest, and Davivienda; BAC and Oriental Bank; Banamex; Cencosud-Scotiabank; as well as Banco Falabella, Banco Itaú, Santander, Pomelo, Dock, Evertec and ueno bank^[8]. The instrument is called an agentic token, and it “protects the payment credentials stored in AI agents through dynamic, cryptographic data, giving issuers full visibility and control”^[8]. The 2026 Fórum E-Commerce Brasil recorded that almost one hundred percent of Latin American issuers are already enabled to operate them^[9].

The standard that documents consent advanced in parallel. The AP2 protocol that Google published in September 2025 with more than sixty partners moved in May 2026 to community governance within the FIDO Alliance^[6]. None of this is regional ornament. Latin America's payment infrastructure has run years ahead of almost the whole rich world's, and Pix alone moved R$ 35.3 trillion across 63.5 billion transactions during 2025^[11]. All of that solves the last step of a purchase whose first step is that somebody finds you. Collecting payment well is worth nothing to whoever never entered the catalog.

Who cannot meet the requirement

Labor informality stood at 46.7% in the first half of 2025, according to the ILO^[10]. The differences between countries are enormous: it hovers around 25% in Chile and Uruguay, approaches 70% in Peru and Ecuador and exceeds 80% in Bolivia^[10]. Behind that average lies an economy that works without being written down. The seller takes orders over messaging, agrees the price in the conversation and collects by instant transfer, with no published catalog, no product listing and no stock declared anywhere. To enter the catalog they would have to incorporate a company, contract a domain, draft two legal documents and sustain a file that updates only if somebody programs it.

That seller is not failing through carelessness to meet the requirement to publish terms of service. The requirement describes a legal form they do not have and an administrative apparatus they never needed. It is also worth not reducing this to a shortcoming of the smallest players, because a formal company that sells through social media and messaging likewise fails to produce the file the agent needs, since its catalog lives inside somebody else's platform and its price is negotiated in a chat. A price agreed inside a private conversation has no machine-readable existence, and that sentence describes a good part of the region's commerce.

Those are the facts. What they mean for anyone selling in the region is the business of the section that follows.

The usual way of telling this story puts the emphasis on the risk that the machine buys badly, and the real risk runs in the opposite direction. An agent that does not find a seller commits no visible error and leaves no trace, because it delivers the best result it could build with the catalog it was able to read. The seller does not find out either, since nobody receives a notification for the purchases that never arrived, and that property makes the problem hard to perceive precisely while it is settling in.

The consequence is one of concentration before it is one of competition. The mechanism switches on the moment demand passes through an intermediary that sees only validated files: the small percentage able to publish them receives demand that used to be shared out by proximity and word of mouth, while the rest keep the customers they always had and lose the ones they would have won. A neighborhood shop never depended on a search engine and for that reason weathered the previous era well, but the neighbor who tells an assistant “get me this today” is delegating precisely the part of the process where that shop existed.

The advantage also accumulates. A business inside the catalog keeps adding order history, deliveries fulfilled and ratings, which are the signals by which an agent chooses it again. The one outside generates no signal at all, good or bad. The distance does not hold steady: it widens on its own, and whoever adopts the protocol early will compete with a growing advantage over whoever does not.

Agentic commerce does not sleep either. A feed updates every fifteen minutes and a purchase endpoint answers at three in the morning^[3][4]. Whoever sells over messaging sells while replying, and the message left unanswered at night is an order that went to whoever had an endpoint listening. Some buyers will always prefer to look in the face of the person selling to them, and that share is not going to disappear.

The counterpoint is serious and has to be conceded in full. Messaging platforms can close a good part of this gap whenever they choose, given that the catalogs their customers already upload are structured data and publishing them as a feed is a product decision before it is a technical problem. It is also worth remembering that real agentic volume remains tiny, and that the feature setting all this off was rolled out in a single country, so none of this describes a sentence.

What it does describe is an order of arrival, and orders of arrival decide markets. The demand that moves first toward the agent is the formal demand, the high-ticket demand and the demand of whoever already buys online, which is to say precisely the demand that sustains the margins with which a small business finances everything else. The previous edition described the same mechanism on other ground: when an input becomes abundant, value moves to the scarce complement^[12]. Here the abundant input is the capacity to buy and the scarce complement is being legible, which in Latin America is not a problem of technology but of representing an economy that was never written down.

THE PLAY

1. Publish a feed before a pretty page. The minimum existence in front of an agent fits into a file with identity, price, availability and delivery method, and that file is cheaper than a redesign and more urgent than a campaign. You can keep selling and collecting over messaging in the meantime, because the feed does not replace your channel but makes you findable to the buyer who does not use it.
2. Claim the structured records that already speak about you. Your listing on maps and directories exists even if you have never touched it, and today it is the cheapest layer of representation there is: a wrong opening time or an old address is what the agent is going to read and believe. Correcting it costs an afternoon and depends on no bank and no standard.
3. Build the missing layer, because it is a business and not only a diagnosis. The payment half is already solved and sold by the card networks, whereas making the informal economy legible — records, catalogs and prices of sellers who today do not exist in the form of data — is something nobody has done at regional scale. That is the work with structural demand, and it arrives with an uncomfortable obligation: to represent somebody without exposing them to a regulatory risk they did not choose.