Every Checkout Pays Visa a 59% Margin. These Agents Are Learning to Skip the Toll.

ISSUE 04 · 8 MIN

Most shopping agents stop at the cart because the last step means holding a card on a site that did not invite them. Three builders crossed that line, and the ground they are standing on is the richest in commerce: a 2.4% swipe fee that earns Visa and Mastercard a 59 to 60% operating margin, and a 15 to 22% commission that Booking.com collects on $51 billion of bookings a quarter.

Most shopping agents stop at the cart. They find the thing, compare the price, and hand you a link, because the last step means holding a card and clicking buy on a site that did not invite them. The builders below crossed that line, and their agents finish the purchase.

Two shapes are live. One is infrastructure that lets any agent check out anywhere, and the other is a vertical agent that owns the whole trip from search to receipt. The money question is the same in both: who holds the card, and who decided this purchase was allowed.

The problem these builds attack is not that agents cannot find things to buy. It is that every purchase already pays two tolls before the agent arrives, and both toll collectors intend to keep collecting. The first toll is the card. The average swipe fee on a US card purchase in 2025 was 2.36%, rising to 2.9 to 3.5% when the card is not present, which is every purchase an agent makes. That fee is why Visa reported $11.6 billion of net revenue in its most recent quarter at a 59.1% operating margin, and why Mastercard reported $9.3 billion at 60.2%. Those are not software margins. They are the margins of a toll road with no competing route, and they are the single most comfortable position in the agentic economy.

The second toll is the platform. Booking.com charges hotels a commission that starts near 15% and lands at 18 to 22% once a property pays for visibility, and Booking Holdings collected $7.4 billion of revenue on $51.0 billion of gross bookings in the second quarter of 2026, which works out to about 14.5 cents of every dollar a traveler spends. Expedia charges 15 to 30% and took $4.3 billion on $33.9 billion of bookings in the same quarter. A corporate travel management company adds $10 to $35 per booking on top. Between the card and the platform, a $1,000 hotel stay booked through the usual channels can lose $200 or more before the hotel sees a cent, and the traveler pays for it all in the room rate.

That is the terrain, and it explains the pattern in the three builds below, where the buyer's agent is free, and the seller funds it. Otto earns a commission from the supplier the way a human travel agent does; Mindtrip earns affiliate revenue from the same platforms that charge the hotel; and CartAI charges the developer, not the shopper. An affiliate of Booking.com keeps 25 to 40% of Booking's commission, roughly 4 to 5.5% of the booking, so an agent that sends a traveler to Booking is being paid out of the 15 to 22% toll rather than around it. The guerrilla's opening is the difference between those two numbers. An agent that books direct with the hotel and settles without a card doesn’t need to be paid by the toll, because it removes the toll.

The incumbents have moved first and loudest here. Google announced its Agent Payments Protocol on 16 September 2025 with more than 60 partners. Stripe and OpenAI launched Instant Checkout and the Agentic Commerce Protocol on 29 September 2025. Visa Intelligent Commerce aimed for mainstream delivery by early 2026, and Mastercard launched Agent Pay for Machines on 10 June 2026. Every one of those announcements describes how an agent will pay. None of them publish what the agent will be charged, and until one does, assume the answer is the same 2.4 to 3.5% that funds a 59% margin today.

LISTED · AGENTIC CHECKOUT API

An agentic checkout API. Other people's shopping agents call it to search a catalog, build a cart, and complete the order on a merchant's site, with no integration on the merchant's side. Payments run through a PCI compliant layer on Visa and Mastercard rails, and instead of evading bot detection, CartAI identifies its traffic through Web Bot Authentication with Cloudflare, HUMAN, and Akamai. Product Hunt #2 Product of the Day on 21 July 2026 with 408 upvotes. Founder Manil Uppal, with Kunal Mestri and Aman Gupta.

Reach them · Manil Uppal on LinkedIn · cartai.ai · docs.cartai.ai

LISTED · BUSINESS TRAVEL AGENT

An executive assistant agent for business travel. It books, modifies, rebooks, and cancels flights, hotels, and, since July 2026, car rentals, inside its own interface, without you re-entering credentials. Commission based, free for individuals and small teams for the first 12 months, and no card to sign up. Nine months in closed beta before the December 2025 public launch. Founder Michael Gulmann, formerly of Expedia, incubated at Madrona Venture Labs and backed by Direct Travel and former Orbitz CEO Barney Harford.

Reach them · [email protected] · ottotheagent.com

LISTED · TRAVEL PLANNER TURNED BOOKER

A trip planner that became a booking agent. Flights went live in May 2026 and hotels, as Mindtrip Stays, in July 2026, with inventory from Booking.com, Expedia, Priceline, Agoda, and Nuitee. Payment runs through PayPal and partner processors, and the site states that Mindtrip does not store full card data. Team from Apple, Google, LinkedIn, and ShopStyle.

Reach them · @mindtripai · mindtrip.ai/contact

Three ways to hold the card. CartAI holds it for the developer's agent inside a PCI layer, so the shopping agent never sees the number and the merchant receives an ordinary card transaction. Otto stores your credentials and loyalty numbers and spends them as you, so you never reenter them. Mindtrip refuses to hold the card at all and pushes the payment through PayPal. Who gets paid is the other half of the picture: CartAI is priced per developer and does not publish the number, Otto earns a commission from the airline or hotel, and Mindtrip earns an affiliate fee from the booking platforms, so in all three cases the merchant pays for the agent, and the shopper pays nothing. If you plan to build here, the vocabulary of that arrangement is worth learning, because it decides who your customer is.

Start with interchange, the part of the swipe fee that goes to the bank that issued the card and makes up most of the 2.36% average; the network itself keeps about 0.14%, and the processor keeps the rest. A merchant who sells a $1,000 item to an agent over a card gives up about $24 to $35 before anything else, and that cost is the merchant's cost of revenue whether the buyer was a person or a machine. An agent that settles the same purchase in a stablecoin over a public chain moves the cost of revenue on that line from $24 to a fraction of a cent, and that gap, not the convenience, is why a merchant would let an agent pay a different way.

Next is the difference between a take rate and a commission, which sound alike and are different. Booking's 14.5 cents on the dollar is a take rate: a share of gross booking value that the platform keeps because the transaction ran through it. Otto's commission is a payment the supplier makes for a customer it would not otherwise have reached, and it is a marketing cost on the hotel's books, which means Otto's customer is the hotel and the traveler is the product. Mindtrip's affiliate revenue is a share of somebody else's take rate, 4 to 5.5% of the booking, so Mindtrip's margin is capped by Booking's willingness to share, which is the weakest position of the three.

The last term is customer acquisition cost, and here it is the whole story. Every build in this issue has zero acquisition cost to the shopper, because the seller pays it. That is a fine way to grow and a dangerous way to build, because a business whose revenue is a commission from the incumbent can be repriced by the incumbent at any time, and Google, Stripe, Visa, and Mastercard have all announced their own agent rails without naming a price. The builder who is paid by the toll is a tenant. The builder who removes the toll owns the ground.

Do not build a general shopping agent. Build a buyer for one category where reorders are predictable and the catalog is messy, such as restaurant supplies, dental consumables, or replacement parts for a single brand of equipment. Put CartAI underneath so you never touch a merchant integration, set a monthly budget per customer, and charge the office manager a flat fee for the hours you save. The moat is the reorder history, not the checkout, because after three months you know what a customer buys before they do.

Then move the payment. Once a supplier sees the same agent reorder every month, offer to settle directly in a stablecoin and split the 2.4 to 3.5% the card would have taken. If you price the service at the platform's commission, you win on cost. If you price it at the toll you removed, you win on margin, and as we always say, “margin is the better business to be in.”

google-agentic-commerce/AP2 · Apache 2.0, about 3.2k stars, with a working Python SDK under code/sdk/python/ap2 and runnable reference implementations in Python, Go, and Android under code/samples. This is Google's Agent Payments Protocol, the one with the 60 partners, and it is the more useful of the two big specs to read because it ships code rather than only a document.

Why it is worth an hour: the pattern to study is the signed mandate. AP2 separates an Intent Mandate, what the human authorized the agent to buy, from a Cart Mandate, what the agent actually assembled, and a Payment Mandate, the instruction that moves the money, and each one is a signed credential the merchant can verify. That is the structure every checkout agent needs whether or not it ever touches AP2, because it is the only clean answer to the question in The Tell. If you want the leaner spec that powers ChatGPT's Instant Checkout today, read agentic-commerce-protocol, about 1.5k stars, Apache 2.0, and start with the latest dated folder under spec.

An agent that can check out anywhere can be pointed anywhere. The web page an agent reads to compare prices can also tell it where to buy, and with a stored card and no recipient list, the difference between a legitimate reorder and a $900 purchase from a store that appeared last week is one line of hidden text.

Merchants see the other side. An agent arrives with a valid card and a Web Bot Authentication header, and the merchant still doesn't know whether the human behind it authorized the purchase. If you want a machine to answer that before the order ships, FLINT Scout does it for a cent, and that is the only pitch you will get from us.

NEXT ISSUE

Agents that chase money for a business: a collections agent that charges 1% of what it recovers, a procurement agent with 1.8 million negotiations behind it, and a 3 person AP department that never sleeps, on ground where a collection agency takes 35% and 43% of US B2B invoices were paid late last year.