the first economy without humans in it

right now, somewhere, a piece of software is paying another piece of software. a few cents for data, a fraction of a cent for an api call, machine to machine, no human approving, no human even watching. on the experimental rails built for exactly this, tens of millions of such payments settle every month, averaging around 32 cents each.

32 cents is nothing. that's the point. nobody builds payment infrastructure for 32-cent transactions unless they expect trillions of them.

commerce is splitting into two economies. one where humans shop. one where machines procure. the second one barely exists today, and i think it ends up bigger than the first.

the honest scoreboard

first, the part the hype gets wrong, because the gap between the rails and the volume is the actual story.

the rails got built absurdly fast. in roughly eighteen months: openai shipped a buy-button inside chatgpt. google launched an agent payments protocol and donated it to a standards body. visa shipped a trusted-agent framework. mastercard launched machine-to-machine micropayment rails with thirty-plus partners. google, shopify, walmart, and target co-authored a protocol that lets any store publish a machine-readable storefront at a standard address, the way websites publish sitemaps.

the usage, so far, is a rounding error. openai's buy-button reached about 30 live merchants and got pulled within six months. walmart measured checkout inside chatgpt converting three times worse than its own site. visa's agent pilots have processed hundreds of transactions. not hundreds of millions. hundreds.

so the skeptic's read is: it flopped. i read it differently, because one number did flip. ai-referred traffic to us retailers grew roughly 7x year over year, and it now converts better than search traffic, a complete inversion from a year earlier. salesforce counted about 20% of global holiday orders as ai-influenced. that measures recommendations, not robot checkouts, but it measures something real: the machine already decides what humans see before they buy.

discovery has flipped. execution hasn't. that ordering is normal. people browsed the web for years before they trusted it with a credit card, and "nobody will type their card number into a website" was a respectable opinion in 1997. rails first, volume later, then all at once.

what dies when the buyer is a machine

whole categories of economic behavior were built for a buyer with eyes, habits, and a limited attention span. when the buyer is software, they stop making sense.

persuasion dies. verification wins. you cannot retarget an agent. you can't make it feel a jingle, anchor it with a fake discount, or catch it in a weak moment at 11pm. an agent reads structured data, provable claims, verified reviews, actual prices. a very large share of the hundreds of billions spent on advertising every year is aimed at a human attention span that is exiting the purchase. marketing to machines is a real discipline, and it looks like engineering: publish machine-legible truth, or be invisible.

the forgetting economy dies. an enormous amount of modern revenue is monetized forgetting: the gym you don't visit, the saas seats nobody logs into, the auto-renewal you meant to cancel. agents don't forget. they audit every renewal, every month, forever. and the reverse: micropayments failed for thirty years because the mental cost of deciding "is this worth four cents" exceeded the four cents. software has no mental transaction costs. pay-per-use, which humans hate and machines prefer, comes back from the dead. mastercard shipping rails for fraction-of-a-cent machine payments is the loudest possible signal of where the networks think this goes.

confusion margins die. entire industries earn their margin by being too confusing to compare: telecom plans, insurance tiers, banking fees, airline pricing. call it what it is: an obfuscation tax on human patience. agents compare everything, always, at zero cost. when the buyer holds every plan in memory simultaneously, that business model has nothing left to hide behind. loyalty as habit dies with it, because the agent re-runs the comparison every single time. the only loyalty that survives is earned preference, logged where machines can read it.

what commerce catches from finance

markets where machines trade against machines already exist. we call it finance, and its pathologies are instructive. in 2011, two repricing bots on amazon bid an ordinary biology textbook to $23 million. nobody intended it. the algorithms just kept reacting to each other.

as agents take over ordinary buying and selling, flash-crash dynamics leak out of finance into retail, logistics, energy. pricing wars at machine speed. inventory whipsaws at machine speed. commerce will need circuit breakers, and it will get them after the first famous accident, not before.

regulators have noticed the other half. when many sellers price with the same algorithm, "we never talked" stops being a defense: the doj settled its realpage case on exactly that theory, and states are writing algorithmic-pricing laws directly. enforcement built for humans colluding over dinner now faces tacit collusion between models. that fight is just starting.

the biggest buildout: trust between machines

here's where the real money is going, and almost nobody frames it this way: the first great infrastructure buildout of the agent economy is not intelligence. it's trust.

think about what a merchant needs to know before accepting an agent's order. is this agent who it claims to be? does it actually have authority to spend its principal's money? will a human honor the purchase, or dispute it? there's a name forming for this: know your agent. kyc for software.

the pieces are landing fast. the ietf is standardizing cryptographic identity for agents, so requests arrive signed. amex now sells agent purchase protection: if a registered, credentialed agent botches a purchase, amex covers the customer. an insurer underwriting the behavior of software is a milestone in economic history that went almost unreported. identity, escrow, dispute resolution, liability cover, reputation: for machines. each one of those is a company-sized opportunity, and most of them don't exist yet.

the fight that decides the shape

one court case matters more than the rest. amazon sued perplexity for letting its agent shop on amazon on a user's behalf, and won the first round: the court held the agent accessed the platform "without authorization" even though a human directed it. then, on august 4, the ninth circuit vacated that injunction: the user did the accessing, the agent was just the hands. one ruling, and the question below lands on the "authority transfers" side, for now.

strip the legalese and the question is still: does your authority transfer to your agent? if yes, agents roam the open market and buy wherever their principals could. if no, every platform can lock agents out and sell access back, and the agent economy is born feudal: gated, tolled, owned by whoever already owns the traffic. the uk regulator has already planted a flag on the adjacent question, ruling that the business deploying an agent is responsible for what it does. the us hasn't. watch this case the way people should have watched the app store rules in 2009.

what stays human

status goods are immune, and it's worth being precise about why: the point of the watch, the bag, the table at that restaurant is being seen by humans. an agent can procure the object. it cannot procure the being-seen.

so shopping splits cleanly. procurement, replenishment, comparison, renewal: machine, utility, invisible. shopping as identity, gift, entertainment, status: human, and more valuable per minute than ever, because human attention in commerce just became scarce. brands will follow that attention, and the ones that survive will speak two languages fluently: machine-legible truth for the agents, meaning for the humans.

who owns the buyers

now the question underneath all of it. every agent in this economy is a productive asset. it works, it transacts, it generates surplus. and someone owns every single one.

the last infrastructure shift gave us a warning. the cloud concentrated into three companies. the web before it distributed into a hundred million hands. both outcomes are live here, and the stakes are higher this time, because agents aren't where you rent capacity. they're who does your buying, your selling, your earning.

i'll say where i stand. the healthy outcome is the distributed one: the people who use this economy owning pieces of the actors that run it, surplus flowing wide instead of pooling. if agent ownership concentrates the way cloud did, the agent economy becomes the most efficient rent-extraction machine ever built, a toll booth on every transaction on earth. this is the highest-stakes open variable in the whole picture, it's being decided right now by defaults nobody is voting on, and it's worth being loud about early.

the decade

forecasts, labeled as forecasts: morgan stanley puts autonomous agents at 10 to 20% of us e-commerce by 2030. bain says up to 25% if you count influence. today, the honest number for fully autonomous purchases rounds to zero. both things are true. that's what the beginning of a curve looks like.

the bottlenecks are real. agents are still unreliable on long tasks. liability law is half-built: europe withdrew its ai liability directive, and the card networks are effectively legislating faster than governments. trust between machines is a construction site.

but every transaction in the machine economy still starts upstream with a human wanting something. the first economy without humans in it will still be built for humans. the open question is whether it's owned by many of them or a few.