An Exa search costs seven tenths of a cent.
That is not a rounding error or a promotional rate. It is the published price. Send a request to Exa’s search endpoint with no API key and no payment attached, and the server answers with HTTP 402 and a header that says, in machine-readable form, that this will cost $0.007 in USDC and here is where to send it. Content retrieval runs $0.001 per page. CONFIRMED
Now put an AI agent in front of that endpoint. It is halfway through a research task. It needs one search. It has no account with Exa, no API key, no billing relationship, and no human awake at 3am to create one.
On every payment rail described in the rest of this piece, that transaction is not expensive. It is impossible.
A card cannot process it, because the fixed fee alone is larger than the purchase. A bank transfer cannot process it, because the relationship has to exist before the payment can. An instant payment scheme cannot process it, because the agent and the API are almost certainly in different countries and no scheme spans both.
The money exists. The willingness exists on both sides. The rail does not.
I have spent a fair amount of time this year reading about what AI agents will do to finance, and most of it skips this. The interesting question is not whether agents will transact. It is that the entire global payment system was built on an assumption that is about to stop holding: that on at least one end of every transaction, there is a person who set something up in advance.
This piece is about why that assumption is load-bearing, and what breaks when you remove it.
What the rails actually are
Three families of payment rail carry almost everything. Each has a structural property that matters later, and it is not the one most people would name.
Correspondent banking and SWIFT. McKinsey’s Global Payments Map records roughly $179 trillion in cross-border payments in 2024, of which lower-value flows account for about a tenth. SWIFT connects more than 11,500 institutions across 200 countries and carries around 53 million messages a day, with a single-day record of 68 million set in 2025. CONFIRMED
The detail most coverage skips: SWIFT is messaging, not settlement. No money moves through SWIFT. Money moves across nostro and vostro accounts that banks hold with each other, and the message tells them what to adjust. Where two banks have no direct relationship, the payment relays through intermediaries who do, each one adding a hop, a fee and a delay.
The relationship has to exist before the payment can. Not as a policy choice. As an architectural fact.
Domestic instant schemes. These are genuinely good, and it is worth saying so before the criticism. FedNow launched in July 2023 and settles 24/7/365 in central bank money, irrevocably, in under fifteen seconds, with the transaction limit raised to $10 million in November 2025 and more than 1,500 participants. The Clearing House’s RTP network, bank-owned and live since 2017, processed $1.3 trillion in 2025 after its own limit went from $1 million to $10 million. CONFIRMED
Outside the US the numbers are larger. India’s UPI runs around 660 million transactions a day and accounts for roughly 49 percent of global real-time payment volume, more than three times Brazil’s Pix share of 14 percent. Pix handled 63.4 billion transactions worth about $4.6 trillion in 2024, exceeding combined card volumes in Brazil by 80 percent. CONFIRMED
These systems work. They are also national islands.
Cards. Interchange runs 1.5 to 3 percent plus a fixed component per transaction. That structure is fine for a $40 purchase and inverts completely below about a dollar, which is the point where the fixed component eats the entire transaction.
Six pain points, with numbers
Cost. The World Bank’s Remittance Prices Worldwide puts the global average total cost of sending $200 at 6.49 percent as of the first quarter of 2025, easing to 6.36 percent by the third quarter. Large corporates typically pay 1 to 3 percent on cross-border B2B. Smaller businesses pay above 5 percent. Some corridors exceed 10 percent. Correspondent wire fees commonly run $25 to $75 per transaction, split across a correspondent handling charge, a lifting fee, and an FX spread that is frequently not disclosed as a fee at all.
Time and cut-offs. SWIFT reports that 75 percent of its payments reach the destination bank within ten minutes, which is a real improvement and also not the whole journey. Fedwire operates weekdays only. ACH runs in four daily batch windows with settlement in one to three days. The last mile still lands on somebody’s business hours.
Trapped liquidity. To make cross-border payments feel fast, banks pre-fund nostro accounts in advance, which means capital sits idle in foreign currencies purely to be available. Meanwhile the number of correspondent banking relationships has fallen by more than half over the past decade, which lengthens the chains that remain.
The minimum viable transaction. This is the one that matters most for what follows. Card interchange makes anything under roughly a dollar uneconomic to process. Not difficult. Uneconomic, in the sense that the processor loses money. So the rail does not serve that segment, and everyone has quietly built around the gap using subscriptions, monthly invoices, prepaid credits and API keys. Every one of those workarounds assumes a human signed up in advance.
Interoperability. There is no universal language for real-time payments. One top-ten US bank described spending eighteen months building translation between FedNow’s ISO 20022 implementation and its own legacy systems, then another six months for Brazil’s modified Pix schema. The BIS is addressing the gap directly through Project Nexus, linking the instant schemes of India, Malaysia, the Philippines, Singapore and Thailand, targeting go-live in 2027 and potentially reaching 1.7 billion people. ANNOUNCED
Reversibility. Cards have chargebacks. FedNow and Fedwire are irrevocable by design. That difference is manageable when a human can call their bank and explain. Nobody has resolved what a dispute means when neither party is a person and the transaction was one of forty thousand made that hour.
Stablecoins: the bridge that got halfway
The obvious objection at this point is that stablecoins already solved this, and the objection deserves a serious answer rather than a dismissal, because it is substantially correct.
Stablecoin transaction value reached $33 trillion in 2025, up 72 percent year on year, with USDC at $18.3 trillion and USDT at $13.3 trillion. Supply sits around $300 billion. Roughly ten million addresses transact daily and more than 150 million hold a nonzero balance. CONFIRMED
Read those numbers carefully, though, because the headline overstates the case. Visa’s adjusted measure, which strips out bot activity and internal transfers, puts the figure nearer $10.2 trillion on a trailing twelve month basis. The gap between $33 trillion and $10 trillion is not fraud. It is methodology, and any figure quoted without its definition is close to meaningless.
Even at the conservative measure, this is the most important thing that has happened in payments in a decade, and it happened almost entirely outside the banking system.
Here is what stablecoins actually proved, and it is bigger than the volume. They demonstrated that tokenized cash works at scale, that value can move between parties with no prior relationship, in seconds, at a cost that does not scale with the size of the transaction. That last property is the one traditional rails cannot reproduce. A $0.007 payment and a $7 million payment cost roughly the same to settle on-chain. On a card rail they differ by orders of magnitude in relative cost.
This is why the tokenization argument stops being theoretical. Not because a bond became a token somewhere, but because a form of tokenized money reached meaningful daily volume and kept working. Every institutional tokenization project now under way depends on there being a credible cash leg. Stablecoins are the existence proof for that leg.
So stablecoins are not a sideshow to this story. They are the bridge between the rails described above and everything that follows, and any account of agent payments that treats them as a crypto curiosity has misread what happened.
But the bridge only goes halfway, and it is worth being precise about where it stops.
Stablecoins solved settlement. They did not solve identity.
When a stablecoin moves, the network knows an address paid an address. It does not know who they are, whether either is authorised to transact, whether a supervisor is entitled to see the flow, or what happens if one of them should not have done it. For two individuals moving money, that is fine and arguably the point. For an institution, it is the whole problem, because an institution cannot transact with a counterparty it cannot identify, and it cannot use a rail a regulator cannot examine.
Add an autonomous agent to that picture and it gets worse rather than better. Now the party initiating the payment is not merely pseudonymous. It is not a legal person at all.
Settlement is largely solved. Identity, authority and supervision are not, and those are the three things that decide whether an institution can use any of this.
Enter the machines
It is worth separating what is actually happening from what is being announced, because the gap is wide.
The Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services Report finds 52 percent of surveyed industry respondents actively adopting agentic AI, split between 23 percent scaling or transforming and 29 percent piloting. CONFIRMED
The European Central Bank reports that more than 85 percent of banks under European banking supervision use AI in some form. Read the detail and the picture narrows considerably: the uses are overwhelmingly monitoring, fraud detection, coding assistance and compliance. Analysis, in other words. Not execution. CONFIRMED
The Bank for International Settlements has gone further in a controlled setting, testing a generative AI agent on intraday liquidity management in a simulated wholesale payment environment. Simulated is the operative word. CONFIRMED
So the honest read is this. Agents are nearly everywhere in analysis and almost nowhere in execution. Institutions have been comfortable letting software form an opinion and have not been comfortable letting it move money.
That is not caution for its own sake. It is a rational response to the fact that the machinery for supervising a machine that spends does not exist yet.
What an agent actually needs
Assemble the requirements from what the rails above cannot do, and you get a list of seven. Each one is a real gap rather than a preference.
Sub-cent economics. The transaction has to cost meaningfully less than the thing being bought. At $0.007 per search, the payment overhead has to be a fraction of a cent, which rules out every rail with a fixed per-transaction component.
Machine-speed finality. An agent working through a task cannot block for three days, or overnight, or until the recipient bank opens. It needs to know the payment is done in the same order of magnitude of time as the API call itself.
Delegated authority with limits. Someone has to be able to say that this agent may spend up to this amount, on these categories, for this long, and have that constraint enforced by the system rather than by the agent’s own good behaviour.
Identity for a counterparty that is not a person. Every existing identity framework in finance assumes a human or a legal entity. An agent is neither. The International Monetary Fund put it plainly in its 2026 note on agentic payments: know-your-customer and multifactor authentication are designed around human users who explicitly approve transactions. CONFIRMED
Per-call metering rather than invoicing. The subscription and the monthly invoice exist because per-transaction billing was uneconomic. Remove that constraint and the natural unit becomes the individual call.
Dispute handling with nobody in the loop. If the API returns garbage, or the agent asked for the wrong thing, there is no cardholder to file a claim and no service desk to call.
Auditability sufficient for a supervisor. Not a log the operator keeps voluntarily. A record a regulator can demand, read and act on.
None of this is a criticism of the existing rails. They work extremely well for the thing they were built for. The list is simply a description of a customer they were never designed to serve.
Why this is happening now rather than in 2019
Three things changed in roughly the last two years, and they are worth naming because “AI got better” is not an explanation.
The first is technical. Stablecoin authorisation patterns now let a payer sign an approval without holding the network’s native token to pay a fee, which removes a step that made machine payments impractical. Second, layer two transaction costs fell below a hundredth of a cent, which is what makes a $0.007 payment viable rather than absurd. Third, models became capable of reading a structured payment challenge and signing a response in milliseconds without a human confirming each one.
Then the commercial and regulatory opening. The GENIUS Act passed in July 2025 in the United States and MiCA came into force in the European Union, which between them gave large regulated institutions a defensible answer to the question of what a stablecoin legally is. That question had been blocking enterprise adoption far more than any technical limitation.
The capability and the permission arrived at roughly the same time. That is the actual reason this is a 2026 conversation.
Six answers to the same question
Something notable happened between April 2025 and March 2026. Six organisations shipped protocols addressing this gap, and most of them are not competing with each other in the way the coverage suggests.
Coinbase released x402 in May 2025, reviving the dormant HTTP 402 status code so a server can quote a price and a client can pay it inline. Mastercard announced Agent Pay in April 2025, extending its existing tokenisation infrastructure to agents. Google published AP2 in September 2025 with more than 60 launch partners, structured around signed mandates and deliberately settlement-agnostic. Visa introduced its Trusted Agent Protocol the same month, focused on verified agent identity and issuer-signed consent. OpenAI and Stripe published the Agentic Commerce Protocol for checkout. Stripe followed with its Machine Payments Protocol in March 2026. CONFIRMED
Six protocols, six organisations, roughly one quarter.
The instinct is to call that a standards war and wait for a winner. I think that reading is wrong, and getting it wrong matters, because these protocols mostly operate at different layers and several are designed to compose with each other rather than replace each other. One of them even ships an extension to use another for settlement.
Part two takes them apart properly: which layer each occupies, what x402’s history on Base actually demonstrates about the relationship between a protocol and the network underneath it, and the thing all six have in common that nobody seems to be saying out loud.
Because they do have something in common. Almost every one of them was built for an agent buying from a merchant on behalf of a consumer. Your assistant books the flight. That is a real and large market.
It is not the same transaction as one institution’s agent paying another institution’s agent, and the difference is not a matter of scale.
Somewhere in your organisation there is a number nobody wrote down deliberately. It is the smallest payment you can actually process without losing money on it, and it was set by the rails you happen to use rather than by anything you decided.
What is that number where you work, and what sets it?
Six standards and one blind spot. The layer taxonomy, x402’s governance history, the security findings, and what all six have in common.
The counterparty you cannot onboard. What institutional agent-to-agent payments need, who can actually serve them, and what would prove it wrong.
References
- Exa x402 documentation, pricing and payment flow (exa.ai)
- McKinsey Global Payments Map, cross-border volumes (mckinsey.com)
- World Bank, Remittance Prices Worldwide, Issues 53 and 54 (remittanceprices.worldbank.org)
- Federal Reserve, FedNow service details and limits (frbservices.org)
- ACI Worldwide real-time payments data, cited by India’s Ministry of Finance, December 2025 (pib.gov.in)
- Artemis Analytics stablecoin volumes, reported by Bloomberg, 8 January 2026; Visa adjusted measure (visaonchainanalytics.com)
- Cambridge Centre for Alternative Finance, 2026 Global AI in Financial Services Report (jbs.cam.ac.uk)
- European Central Bank on AI adoption in supervised banks, June 2026 (bankingsupervision.europa.eu)
- Bank for International Settlements, generative AI agent in wholesale payments (bis.org)
- International Monetary Fund, Note 2026/004, How Agentic AI Will Reshape Payments (imf.org)
- BIS Project Nexus, cross-border instant payment linkage (bis.org)