When your agent holds the card: the war to become the Visa of machines

THE THESIS. The agentic payments war is not about technology: it is about who certifies trust. An AI agent is a program that performs tasks for you autonomously, and for it to buy things it needs something new: digital proof that you gave it permission, how much it can spend, and on what. Whoever turns their «mandate» format — that signed permission — into the de facto standard will control the plumbing of an enormous market. Gartner projects that by 2028, 90% of B2B purchases will be intermediated by agents, over $15 trillion (a prediction unveiled at its October 2025 Symposium). Our reading, against both euphoria and panic: the prize is real, but the first battle already ended in a retreat — OpenAI switched off its buy button in ChatGPT in March 2026 — and that tells us this transition will be slower and deeper than the marketing suggests.

ONE YEAR, SEVEN PROTOCOLS. The timeline matters because it reveals the incumbents' anxiety. Mastercard unveiled Agent Pay with its «agentic tokens» in April 2025; a token is an encrypted stand-in for a card number, here extended to also identify the agent. Google launched AP2 in September 2025 with more than 60 partners, from Mastercard to American Express, PayPal and Coinbase. That same month, OpenAI and Stripe released their Agentic Commerce Protocol (ACP), and Coinbase and Cloudflare created the x402 foundation for stablecoin payments. Visa answered in October 2025 with its Trusted Agent Protocol (TAP), and PayPal introduced its agentic commerce services on October 28, 2025 (now «Agent Ready»). In January 2026 Google added a second standard, the Universal Commerce Protocol, for merchant handoff. Seven protocols in roughly twelve months. We wrote it when analyzing the Google-Microsoft «alliance»: the AI war is shifting from the smartest model to the plumbing — distribution, integration, standards. Payments are the extreme case: nobody can afford to let someone else write the standard, because whoever issues agents' identities will collect the toll on the machine economy. That is why everyone swears interoperability: Visa's TAP moves no money — it is an identity layer that sits on top of any rail — and AP2 supports both cards and stablecoins. It is VHS versus Betamax, but with compliance lawyers.

HOW A MACHINE ACTUALLY PAYS. The mechanics are worth understanding, because they are what separates this from handing your card to a bot. In AP2, every purchase is broken into three signed mandates: intent (what you want and within what limits), cart (what the agent assembled) and payment (what gets charged). Each mandate is a W3C verifiable credential — a cryptographically signed digital document any party can check without calling the issuer. Since AP2 version 0.2 (April 2026) there are even «human not present» payments: you sign the conditions upfront and the agent executes alone, for instance when concert tickets go on sale. Visa's TAP adds the other half: it cryptographically signs the agent's messages (the RFC 9421 standard) so the merchant knows this bot is legitimate and acting under real authorization. Above all, «know your agent» (KYA) is emerging: just as banks must identify their customer (KYC), here the person or company responsible behind every money-moving agent must be identified. The underlying idea is sound and we buy it: turning delegation into an auditable trail. If something goes wrong, there are signatures, not excuses.

THE FIRST RETREAT. The year's most informative data point is not a launch — it is a pullback. OpenAI launched Instant Checkout — buying without leaving ChatGPT — in September 2025 and retired it on March 6, 2026, as Shopify confirmed to Forrester. Fewer than 30 Shopify merchants ever went live (Forrester's figure), and industry outlets reported that Walmart converted three times worse inside ChatGPT than on its own site. Forrester's reading is sensible: payment is the most regulated and most emotional part of shopping, and merchants do not want to surrender their checkout, their data or their loyalty programs. The pendulum swung to «discover in AI, buy at the store». Google, by contrast, doubled down: its Universal Cart — a unified basket pulling products from Search, Gemini and YouTube — reached the US on May 19, 2026 with partners like Walmart, Target and Nike. And OpenAI is preparing the next round on another flank: its Dots, introduced on September 29, 2026, are always-on agents with their own cloud computer and access to over 4,000 apps. They do not shop today; the question is how long that lasts. Our reading: controlling discovery does not automatically hand you the payment. Checkout is retail's last trench, and retail will defend it.

STABLECOINS: THESIS AND MIRAGE. The crypto thesis has one serious argument: machines will make millions of micropayments — buying a data point, an API call, a second of compute — and cards cannot serve that, because their minimum fee hovers around 30 US cents. According to a Coinbase Institute report (October 2026; an interested party, worth remembering), 76% of agent transactions fall below that floor. The x402 protocol repurposes a forgotten web code, HTTP 402 «Payment Required», and settles in USDC, a stablecoin: a cryptocurrency designed to always be worth one dollar. The numbers impress — roughly 160 million payments and close to 69,000 active agents through April 2026, per ecosystem data reported by CoinDesk — until you look closely: analytics firm Artemis estimated that about half of that volume was wash trading, fake transactions between one's own accounts to inflate the figures. A lesson we already learned with benchmarks applies here: adoption metrics are also a property of the harness that measures them. Our position is multi-rail: agents are rail-agnostic and will use whichever lane suits each case. Stablecoins have won the cents-level machine-to-machine niche; cards and wallets will retain delegated human purchases, which is where the dollar volume lives. Neither «crypto is inevitable» nor «crypto is irrelevant».

WHO ANSWERS WHEN YOUR AGENT BUYS WRONG. Here lies the debate with years of runway, because the law is trailing. Today no country has specific legislation on liability for an agent that buys on its own. The default allocation is uncomfortable: the user who authorized the agent eats the bad purchase, and the merchant usually eats the chargebacks — the forced refund of a charge — because classic fraud evidence (device fingerprints, browsing history) does not exist when a bot does the buying. The UK's competition regulator (CMA) set a reasonable marker in March 2026, according to legal analyses in the sector: consumer law applies equally with a human or an agent, and the business is responsible even if a third party built the agent. And the front is heating up: the US FTC opened an investigation into OpenAI and Anthropic on September 30, 2026 over possible consumer risks, including episodes — as reported by several outlets — of agents slipping out of their test environments. That fits what we have been saying: the real risk from agents is short-term and very concrete, not distant superintelligence. There is also a silent casualty: advertising. If a machine does the comparing and deciding, brand storytelling loses power to structured product data — price, specs, availability readable by the agent. And a new antitrust problem is born: whoever ranks the results inside the agent holds a market power regulators do not yet know how to measure.

OUR LONG-TERM READING. Short term, honesty: there will be industrialized fraud, diffuse liability and a worrying concentration of power in whoever controls identity and mandates. But the long horizon looks bright to us — not on faith, but on arithmetic: every friction removed from a transaction — searching, comparing, negotiating, claiming — is time and cost returned to people. A world where machines handle the logistics of buying is one more step toward the abundance we stand for: fewer hours on errands, more hours for what each person actually cares about. The condition is not technical, it is governance: auditable mandates, real «know your agent», and rules written on evidence, not panic. Machines with wallets can be great news — as long as the signature, and the final word, remain yours.