The Unit of Account War: Four Agentic Payment Launches in Eight Days, Zero Interoperable Definitions

On October 20, four companies shipped four different ways to let a machine spend your money. Three of them shipped on the same day, inside the same trade-show news cycle. None of them agreed on what they were selling. Evervault sells sessions — $0.25 apiece, sitting behind a $995-a-month platform fee. Stripe, working through Sierra, sells tokens at $0.15 per successful transaction, stacked on the 2.9% + 30¢ it already collects. Visa, alongside BILL and Billtrust, sells invoices. Mastercard sells sub-cent: fractions of a penny, terms undisclosed, detail locked behind a login wall. Run the arithmetic and the absurdity surfaces fast. A $0.15 token fee on a $100 cart is 15 basis points — roughly five percent of the 290bp you already hand a card network. Invisible. The same $0.15 on a $5 micro-purchase is 300 basis points, and total cost of acceptance pushes toward 6%. The model doesn't bend under that. It breaks. Same product, same price, a twenty-fold swing in economic meaning, decided entirely by the size of the ticket. That ratio is the story. This is not a payments race. It is a metering race — and whoever wins the metering argument owns the customer's cost ledger for the next decade. The window was October 20 through 27, and the clustering was not coincidence. Money20/20 ran inside it. Four announcements, one exhibition floor, one category attempting to be born in public. Evervault launched at the show. Sierra and Stripe announced jointly. Visa brought BILL and Billtrust. Mastercard went into early access with almost nothing disclosed at all. I have watched this movie before. In 2017 I spent six weeks inside the 0x whitepaper and its earliest contract interactions, and what I found was not a token story — it was an atomic swap standard story. The token was a fundraising instrument wrapped around an open-source settlement primitive. Everyone traded the wrapper; the wrapper was the least interesting object in the stack. The lesson I carried into every cycle since: in a young protocol market, the durable asset is never the thing being priced. It is the thing being standardized. Agentic payments sit at exactly that stage. The rails underneath are thirty years old and entirely unremarkable. Stripe still clears at 2.9% + 30¢ through card networks. Visa, BILL and Billtrust still route through existing supplier portals. Mastercard still runs its own network. Not one of the four products changes the clearing layer. What changed is the credential. What changed is the authorization abstraction bolted on top. So the relevant historical analogy is not 2021 PFPs or DeFi Summer. It is closer to the ERC-20 fight, or the EIP-712 signature standard — a moment when multiple teams present incompatible primitives for the same verb, and the market must decide which one becomes default before any of them can scale. Except there is a twist that did not exist in 2017: the base rails are already commoditized and already owned by someone else. That changes the strategic geometry entirely. Interface innovation layered on an owned rail has a well-documented fate — it gets absorbed, standardized, or repriced as a free feature. Every team that built a clever checkout abstraction over card networks in the 2010s learned this the hard way. And still, three launches landed on one day. That is not product competition. That is narrative competition, a scramble for the category-definer slot inside a single news cycle. Which means the differentiation has not been tested. Nothing has been through a procurement cycle. Nothing has been price-compared side by side. The sorting happens in Q4 budget season, and the sorting mechanism will be brutal, because the units do not convert. Look at where each company places its meter. That is not a billing decision. It is a confession about where they believe they are exposed. Evervault meters the session — a one-time secure container that exists before the transaction resolves. Revenue scales with agent activity, not with transaction value, and it is entirely decoupled from whether the payment succeeds. That is a defensive posture from a company that is not a licensed payment institution and has no intention of becoming one. It sells isolation. The $995 monthly platform fee is the tell: usage pricing here is a sales instrument calibrated to reduce trial friction and speak the developer's dialect, not an economic model. The subscription is the business. Stripe meters the successful transaction — $0.15 per token on top of interchange. It gets paid on outcome, which means it owns incentive alignment and can afford to look patient. It is also the only participant here that is already an acquirer, already holds the merchant relationship, already lives inside the developer's build pipeline. Charging fifteen cents is not a revenue strategy. It is the creation of a billable line item, a meter installed today so the rate can be adjusted later. Anyone who has watched platform pricing evolve recognizes the pattern. Visa, with BILL and Billtrust, meters the invoice. Highest value per unit, lowest frequency, and — critically — an object that already exists in every ERP chart of accounts on the planet. No education required. No new accounting category. The invoice is the only unit in this entire race that a controller can book without asking a question. Mastercard meters the sub-cent. Fractions of a penny. The machine-native unit, and the one with the weakest evidentiary basis of demand. Four meters, four different answers to the same question: what is a payment? Whoever defines the unit also defines the customer's cost aggregation and reconciliation taxonomy. That is the actual prize. Not the fee. The ledger. When a controller builds next year's budget, they need a line item. Session? Token? Invoice? Sub-cent event? Whichever one is adopted into the accounting system becomes sticky in a way no API contract can replicate, because switching means re-mapping the chart of accounts, re-instrumenting the audit trail, re-training the AP function. That is a moat built out of workflow inertia, and it is deeper than anything in the technology stack. The arithmetic here matters more than any roadmap. Two ticket sizes, one fee, two incompatible conclusions. At $100: $0.15 is 15 basis points. Against a 290bp card cost, the token fee is a rounding error — five percent of what the merchant already pays, easily absorbed, easily bundled, easily hidden inside a blended rate. At $5: $0.15 is 300 basis points. Add 290bp of card cost and the merchant is approaching 6% all-in for a five-dollar sale. That is not a viable commerce margin for almost any physical or digital good. The model self-negates at exactly the price point that makes agentic commerce interesting in the first place. This is the most important quantitative inference in the whole cluster of announcements, and almost nobody has run it. Any per-event fee has a break-even ticket size below which it destroys the transaction it is metering. For a $0.15 fee, that size is somewhere north of $30 if you want the fee to stay under fifty basis points. Which means the industry cannot converge on a pure per-event unit. It will converge on a two-part tariff: a per-transaction floor plus an ad valorem component. Every payment network in history has arrived at this structure for the same reason — fixed cost per event, variable margin per dollar of value. The current confusion is not a disagreement about pricing philosophy. It is a disagreement about which half of the two-part tariff each vendor wants to be judged on. Mastercard's insistence on the sub-cent framing is the loudest signal in the room, because Mastercard is the only participant pushed into a corner by small-ticket economics. When your unit is a fraction of a penny, a per-event fee of any size is a catastrophe. Sub-cent is not a feature Mastercard wants. It is a corner Mastercard has been backed into, and the early-access framing is the polite way of saying the netting mechanics have not been solved. Because that is what sub-cent actually requires. There is no card rail that clears individual transactions below a cent. Sub-cent pricing forces netting and batch processing — Mastercard would have to stand up a netting ledger inside the network, effectively operating a small clearing house within a card organization. That is the highest technical difficulty of the four products, and the least discussed. It is also where settlement finality and risk retention collide with rules written for gross, consumer-scale transactions. Here is the gap that no one in the cluster has addressed, because no one can: no existing legal framework assigns authority to an agent that decides to pay. Four very different institutional roles are in play. Evervault is a security and technology vendor, not a licensed payment institution. Stripe is an acquirer. Visa and Mastercard are network operators. BILL and Billtrust are accounts-payable automation vendors. The licensing regimes cover who clears. Nobody holds a license for who decides. The regulator's question is not whether the payment was processed correctly. It is whether the agent had authority — and whether that authority was granted ex ante by a human buyer or generated autonomously by a model. Visa's own framing gives the answer away: eligible payments are those already approved by the buyer. In one sentence, the agent is demoted from decision-maker to executor. That is the only defensible compliance posture available today. It is also a quiet rewriting of the word agentic into something much older and much safer: pre-authorized automation. Read the product names, then read the terms. The distance between them is the gap between a narrative and a rulebook. The most elegant piece of engineering in the cluster is also the most commercially cynical. Sierra routes the card number directly to Stripe; the agent only ever sees payment status and the last four digits. On its face this is textbook data minimization. It is also, and more importantly, a procurement weapon. Keeping the PAN outside the model boundary is what keeps an enterprise out of PCI DSS audit scope for its agent deployment. Whoever lets a company run agents that pay without forcing the company through a card-data compliance program holds the key to enterprise purchasing. That, not the fee schedule, is what decides deals. I have sat through enough security reviews at crypto-native firms to know that compliance scope is the single most effective veto in enterprise sales — a nineteen-cent fee will never lose a deal that a PCI audit requirement can kill outright. The corollary is unglamorous but decisive: the winning primitive may not be the cheapest, the fastest, or the most decentralized. It may simply be the one that keeps a Fortune 500 CISO from having to open a ticket. Of the four scenarios in play, only one has an accounting category that already exists: B2B accounts payable bound to an invoice. The volume is large, the frequency is low, the counterparty set is known and KYC'd, the dispute surface is contractual rather than consumer-protection, and the ERP already has a field for it. Visa, BILL and Billtrust are not selling a future. They are selling automation of a workflow that predates the internet, and that is precisely why it will land first. Compare the maturity gradient across the four use cases: B2B payables and invoice binding, clear and large-ticket; embedded checkout inside a browser session, essentially existing e-commerce with a new input method; consumer conversational shopping, high frequency and heavy dispute exposure; machine-to-machine micropayments, no validated demand at all. The last one is a call option on a future narrative, not a current product. Micro-payment models have failed repeatedly over two decades, and they failed on demand, never on engineering. The one operational number published across the entire week was a completion rate of 71.3% for Cedar's agent, Kora, with no methodology disclosed. Invert it. Roughly 28.7% of initiated payments did not complete. That is the largest leakage point in agentic commerce right now, and it arrives as a single unaudited figure with no denominator, no definition of initiation, no breakdown between user abandonment, out-of-stock, and risk-model rejection. Strip out voluntary abandonment and the true failure rate might be far lower. Add risk-model false positives back in and it could be higher. A single completion rate without methodology should not survive a procurement review, and yet it is the only operational benchmark the entire category has produced. The silence around it is more telling than the number. If you had a genuinely strong story on authorization success and dispute handling, you would publish the methodology. The people who decline to publish are the people who have looked at the tail. The consensus reading of this week is that agentic payments finally have product-market fit, and that the fragmentation of four incompatible units is a temporary teething problem that standards will fix. I think that gets it backwards on both counts. First, fragmentation is not a bug at this stage — it is the price of admission. Every category goes through a period where the definition of the atomic unit is contested, and the contest is not waste. It is how the market discovers which abstraction survives contact with a controller's spreadsheet. What looks like chaos is actually a distributed search. The danger is not that four units exist. The danger is that the search is being run by parties who do not control the rules that will eventually govern it. Second, and more uncomfortable: the payment is not the valuable part of this. The payment is the commodity, and it is being priced to zero by three of four participants within twenty-four months. What is not commodity is the authorization audit layer — the immutable record of who authorized what, within which scope, revocable at what point, provable after the fact. That product does not exist yet. There is no vendor selling it, and none of the four launches even gestures at it. It is a genuinely independent position, because it sits outside the card network rulebook rather than underneath it. Every hack is a lesson in trustless verification. The agentic payment stack has, so far, learned nothing from that lesson. It has built four beautiful ways to verify a payment and not one way to verify a mandate. Third, and this is the blind spot I keep circling: Evervault's session is a security boundary and a billing unit at the same time. Those two responsibilities are welded to one abstraction. A session is a one-time container — used and discarded, accumulating no cross-customer authorization history. That means no data network effect, which is a structural weakness dressed as privacy discipline. Worse, it means the risk model and the revenue model fail together. Prompt-inject an agent into swapping the funding instrument or the payee mid-session, and the security container is compromised at the exact moment the meter is still running. You have a single point of failure that also happens to be the invoice. And the deepest structural risk is the one nobody wants to price: the chargeback reason code set has no entry for an agent acting outside the scope of its mandate. Disputes of that shape fall into manual review, at a handling cost that can exceed the disputed amount. That alone kills sub-cent micropayment economics long before the technology is ready. Every hack is a lesson in trustless verification; every dispute is a lesson in verification that never happened at the point of authorization. Watch three things over the next twenty-four months, and ignore everything else. Whether a fifth non-convertible unit appears. Four units is a search. Five is a fragmentation crisis, and a fragmentation crisis is the moment a conversion layer becomes a business — a reconciliation product that maps sessions, tokens, invoices and sub-cent events onto something a CFO can actually book. That is the highest-value unoccupied position in the entire stack, and it is currently empty. Whether a card network updates its rulebook to require token-requestor certification and define liability for agent-initiated payments. If that happens with a transition window shorter than six months, every private primitive in this cluster gets rebuilt to someone else's spec, and the first-mover advantage converts cleanly into migration cost. And whether anyone publishes an agent transaction decline rate with a methodology attached. The day that number becomes auditable is the day this category becomes investable at scale. Until then, everything on offer is a narrative wearing a fee schedule. Every hack is a lesson in trustless verification. The question for 2027 is simpler than it sounds: when a machine pays on your behalf, who holds the receipt that proves you said yes — and can you read it without a login?