Machine Payments Are Supposed to Be Sub-Cent. Visa's Own Data Says Seventy-Six.
Visa published the first volume figures for machine payments. Divide them: the average ticket is seventy-six cents, not a fraction of one. Then read what the same report concedes about liability.
Visa published the first real numbers on machine-to-machine payments. Divide them and the story changes. Then read what the same report says about liability, and the story changes again.
In July, Visa and Artemis Analytics published a report on agentic payments. It is the most substantive thing anyone has produced on the subject, it contains the first public volume figures for the machine-payment protocols, and it concedes — in a network's own words, in print — the thing I have been arguing since April of last year.
Two parts of it are worth more than the rest of the agentic coverage of the last eighteen months combined. One is an arithmetic problem nobody has done. The other is a sentence about liability that Visa chose to publish about itself.
Do the division
The report gives cumulative figures as of 21 April 2026.
x402, launched May 2025: over $135.7 million in volume across 178.3 million transactions.
Machine Payments Protocol, launched mid-March 2026: over $38,000 in volume across approximately 184,600 transactions.
Now divide.
x402's average transaction is about seventy-six cents. MPP's is about twenty-one cents.
That is the whole architecture argument, with the air let out of it.
The entire case for machine-native rails — made in this report, made by Mastercard in June, made by every protocol team for two years — rests on sub-cent payments that card economics cannot price. Fractions of a cent. API calls. Per-token compute. The report itself says gas on the newer chains has fallen to fractions of a cent and that Tempo offers sub-cent fees with five-hundred-millisecond settlement.
But the observed average ticket is not a fraction of a cent. It is seventy-six cents, and twenty-one cents, and those are different animals entirely. Seventy-six cents is not below the floor of card economics; it is awkwardly near it. It is the zone where the argument stops being about physics and starts being about interchange structure, minimum fees and who is willing to price a low-value tier — which is a commercial negotiation, not a technical impossibility.
I want to be scrupulous about what this does and does not prove, because it would be easy to overclaim.
These are cumulative raw onchain totals. Visa says so explicitly, and says that adjusted figures excluding wash and test transactions appear elsewhere in the report. I have not seen those adjusted figures. Stripping wash and test activity removes both dollars and transaction counts, so the true average could move in either direction, and I cannot tell you which. A long tail of genuinely sub-cent calls could sit underneath a small number of large transfers that drag the mean upward; a median would answer that and is not published.
So the honest statement is narrower than the headline and still worth making: the only public data on machine payments shows an average ticket two orders of magnitude above the one the architecture is justified by, and nobody has published a median. Anyone building a business case on sub-cent volume should ask for that median before they build it.
The other number
MPP has moved thirty-eight thousand dollars.
Not thirty-eight million. Thirty-eight thousand, across roughly 184,600 transactions, in the five weeks between mid-March and 21 April.
I am not mocking it. A five-week-old protocol moving thirty-eight thousand dollars is a perfectly respectable five-week-old protocol, and MPP has the distinction, per the report, of being the only agentic payment protocol developed through the IETF standards track — which, given everything I have written about governance, is the single most encouraging fact in the document.
But it is useful to say the number out loud, because the discourse around these protocols is not calibrated to it. A category that has moved less money than a used car is being described, in boardrooms in this region and elsewhere, as a thing institutions need a strategy for this year.
x402's $135.7 million over roughly eleven months is a real number — call it twelve million a month, which is a business, if a small one. MPP's is a pilot. They are being discussed in the same breath, and they should not be.
What waited twenty-nine years
One detail in the report is worth stopping on because it reframes the whole category.
HTTP status code 402 — Payment Required — was reserved in 1997. It has sat there, unimplemented, for twenty-nine years.
The report's explanation is correct and underrated: the protocol was never the obstacle. Card fees made sub-dollar payments unviable, so there was nothing to implement. What changed was not a standard. It was a cost floor, and a source of demand — capable agents, which the report dates to mid-2025.
That is a useful corrective for anyone in this industry who has sat through a vendor presentation about how a new protocol will unlock a market. Protocols do not unlock markets. Economics unlock markets, and then somebody finally implements the protocol that was sitting there the whole time. Twenty-nine years is the longest-running demonstration of that principle I know of.
Then Visa says the quiet part
Now the second half of the report, which is the half I would hand to a board.
Under trust, Visa writes that agent trust is the hardest problem. Then it lists failure modes with a candour that is unusual in a document with a logo on it.
On mis-purchase: a single misconfigured agent can execute hundreds of bad transactions before anyone notices. On adversarial attacks: prompt injection has no complete solution, and the attack surface grows with autonomy. On agent-to-agent chains, where one bad output contaminates every payment upstream of it: no established resolution mechanism exists for unwinding them.
On disputes, three sentences that I would have been accused of exaggerating if I had written them myself. Chargeback windows and evidence requirements were designed for human-speed commerce with clear buyer intent. The concept of a single disputed order does not map to hundreds of sub-cent API calls. Disputes lack established resolution mechanisms.
And on liability, the plainest of all: liability remains legally ambiguous. The delegating human, the hosting platform, the model provider and the merchant all have plausible claims and defences. No jurisdiction has established clear precedent.
I have made versions of this argument eight or nine times since April 2025 — that the industry shipped credentials and left the rulebook closed, that better evidence with unchanged rules produces better-documented disagreements rather than fewer of them, that nobody has said who pays. It is a different thing when a network writes it down.
So: credit, unreservedly. This is an honest document. It would have been easy to publish a thought-leadership PDF that said the problem is being solved. It says the opposite.
The gap is the product
And then the report reaches its conclusion, and the conclusion is the part I want people in this region to read twice.
Among the conditions for scale, it frames incumbents as potential issuers of agent credentials, processors, or providers of the trust and dispute infrastructure that crypto-native protocols still lack.
Read that next to the liability section and the architecture of the argument becomes clear.
Visa is not describing an unsolved problem. It is describing an unserved market.
The missing dispute mechanism, the ambiguous liability, the absence of a way to unwind a cascading agent failure — those are not, in this framing, defects the industry will collectively repair. They are the moat. They are precisely what the networks have and the protocols do not, and the report's closing move is to point at the gap and note, politely, who is positioned to fill it.
That is not cynical of Visa. It is accurate, it is good strategy, and I would do the same in their seat. The dispute apparatus took sixty years and billions of dollars to build, and it is the single thing in the card system that cannot be replicated by writing better software. If agentic commerce needs one, the people who already own one have an enormous advantage, and they are right to say so.
But the consequence is worth stating in plainer language than a thought-leadership report will use: the liability question is unlikely to be answered by a neutral body, because the parties best placed to answer it are the parties who profit from it being unanswered elsewhere. The gap will close as a product, sold with a price, not as a standard, published for free. That is a different future from the one the open-protocol enthusiasts are describing, and the report quietly tells you which one is coming.
From a Panamá desk
The practical reading for an institution in this region is shorter than the report and less exciting than the discourse.
On volume: you have time. The entire machine-payment category, by the only public numbers that exist, has moved an amount of money that would not register in a mid-sized acquirer's monthly report. If your 2027 plan has a line item for machine-to-machine payment acceptance, that line can wait a year without costing you anything. The institutions that lose here will not lose by being late. They will lose by having no opinion.
On the average ticket: ask for the median. If a vendor is selling you a stablecoin micropayment rail on the argument that card economics cannot serve sub-cent transactions, ask what their observed median ticket is. If it is seventy cents, you are not being sold a new category. You are being sold a cheaper price on an existing one, which is a fine thing to buy and a different negotiation entirely.
On liability: this is the part that actually arrives first, and it arrives at you. Visa has now said in writing that no jurisdiction has established clear precedent for who bears an agent's mistake. Panamá is one of those jurisdictions. So is every other country in this region. When the first disputed agent transaction lands in a Panamanian issuer's queue — and it will, before anybody here has a product, because our cardholders will meet agents from the other side of the border long before we offer them one — somebody in a service centre will be improvising an answer that no rulebook contains.
That is not a technology project. It is a page of guidance, a position agreed with your legal team, and a line in a script. It costs almost nothing and it is late the day the first call arrives.
And on the moat: notice where it leaves us. If dispute infrastructure is the product the incumbents intend to sell, then institutions in markets like ours are not going to build it. We are going to buy it, at a price set elsewhere, under rules written elsewhere, for a problem created elsewhere. That is the familiar shape, and the only lever we have ever had against it is being early enough to the conversation to ask what we are being charged for.
My judgment
Three things, in order of confidence.
First, with high confidence: machine-native micropayments are a real category and a much smaller one than the architecture debate implies, and the sub-cent framing is currently a projection rather than an observation. The data in this report is the strongest available evidence on the question and it points the other way. I would revise that view immediately on seeing a published median, and I would like someone to publish one.
Second, with moderate confidence: the liability gap will be closed commercially rather than collectively. Visa's own framing — incumbents as providers of trust and dispute infrastructure the protocols lack — is the clearest statement of intent anyone has made, and it is consistent with everything the networks have done for sixty years. I would be glad to be wrong and would welcome a neutral body. I have stopped expecting one.
Third, with low confidence and stated anyway because it is the useful one: I think the lasting contribution of the agentic protocol era will turn out to be the mandate structures rather than the rails. AP2's separation of intent, cart and payment into three signed objects is a genuinely good idea about what evidence a transaction should carry, and it is portable — it works just as well on a card message as on a chain. Rails come and go. A better definition of what was agreed, and by whom, survives.
Three questions for the people who own this
1. When a cardholder calls to say their agent bought something they did not want, what does your institution's script say today? Not the policy you intend to write — the words a person in a service centre will actually use this month.
2. For any machine-payment rail being proposed to you: what is the observed median transaction value, excluding test and wash activity? If the vendor cannot produce it, you are being sold a projection.
3. If dispute infrastructure for agentic commerce becomes a product your institution buys rather than a standard it inherits, who in your organisation will negotiate that contract — and are they in the room now, or will they meet it as a renewal?
Reply to this email with a line or two. The sharpest answers will shape the next issue.
Sources
Visa and Artemis Analytics, July 2026 — agentic payments report, with transaction data cited as of 21 April 2026 and no individual authors named. Source for the macro and micro commerce distinction; the figure of over $135.7 million in volume across 178.3 million x402 transactions since May 2025; the figure of over $38,000 in volume across approximately 184,600 Machine Payments Protocol transactions since mid-March 2026; the statement that these are cumulative raw onchain totals, with adjusted figures excluding wash and test transactions appearing in sections I have not read; the reservation of HTTP status code 402 in 1997 and the argument that card fees rather than missing protocol were the obstacle; the dating of capable agents to mid-2025; that Tempo offers sub-cent fees with 500-millisecond settlement and that gas on Base and Solana has fallen to fractions of a cent; the description of Visa Intelligent Commerce as announced in April 2026, protocol-agnostic, supporting TAP, x402, MPP, ACP and UCP, with over 100 partners and 30 building in the sandbox; AP2's Intent, Cart and Payment Mandates and its roughly 60 partners; x402's facilitator model, its "Upto" capped authorization, and the move of its stewardship to the Linux Foundation in April 2026; the statement that MPP is the only agentic payment protocol developed through the IETF standards track; and the trust section, including that agent trust is the hardest problem, that a single misconfigured agent can execute hundreds of bad transactions before anyone notices, that prompt injection has no complete solution, that no established resolution mechanism exists for unwinding cascading agent-to-agent failures, that chargeback windows and evidence requirements were designed for human-speed commerce with clear buyer intent, that the concept of a single disputed order does not map to hundreds of sub-cent API calls, that disputes lack established resolution mechanisms, that liability remains legally ambiguous, and that no jurisdiction has established clear precedent. The conditions-for-scale section frames incumbents as potential issuers of agent credentials, processors, or providers of the trust and dispute infrastructure crypto-native protocols still lack.
On the arithmetic. The averages of approximately 76 cents and approximately 21 cents are my own division of Visa's published figures, not figures Visa states. They are means, not medians, and they are computed from raw onchain totals that Visa itself flags as including wash and test activity. The report publishes no median. I have read an excerpt of the report rather than the complete document; sections on MPP's design and a referenced comparison table were not available to me, and adjusted figures may appear there. If a median is published and contradicts this reading, the reading is wrong and I will say so.
A note on what is mine. That the observed average ticket undercuts the sub-cent justification for machine-native rails; that seventy-six cents is a commercial rather than a technical problem; that the liability gap described in the report is being framed as a market rather than a defect, and will therefore close as a priced product rather than a published standard; and that mandate structures rather than rails will be the lasting contribution of this period — those are arguments, not findings. The praise for the report's candour is sincere and I would rather read ten documents like it than one more announcement. Nothing here reflects the position, portfolio, clients or vendors of my employer.