There is an old rule about where money is made, and it has almost nothing to do with how good your product is. Find a place where money already flows. Stand in it. Take a small slice of every unit that goes past.
You do not have to persuade anyone to use you twice. You do not have to win a feature war. When people use you, you get paid, and you get paid again tomorrow without doing anything.
Agentcard is days old, and it has gone straight for a spot in a flow of money that is only just starting to form.
What is actually known
Agentcard is building payment infrastructure for AI agents. Its website is agentcard.sh. The contact for the company resolves to Karen Serfaty.
That is all of it. There is no published product detail, no named customers, no published funding, and no numbers. Anyone writing confidently about what Agentcard has built is guessing, and this piece will not. What can be looked at is the position the company has taken, because a startup this early is mostly a bet on which spot is worth standing in.
How agents pay for things today
Badly, and by pretending to be somebody else. An agent that needs to buy an API credit, a dataset, a flight or a cloud server has three options, and all three are borrowed clothes. It can use a card belonging to a human, stored somewhere and replayed at the checkout. It can use a company card issued to a developer or a department, sorted out later by somebody reading a statement and trying to remember what happened. Or it can drive a browser and fill in the checkout form the way a person would, which is why a lot of agent shopping looks, to a merchant's fraud systems, exactly like fraud.
In every case the agent is wearing a human's identity, because the rails have nowhere else to put it. The card number says a person. The billing address says a person. The dispute process assumes a cardholder you can ring up and ask whether they recognise the charge.
Who else is in this field
Payments is not an empty category. Stripe, Adyen and Checkout.com turned moving money into a few lines of code. Marqeta, Lithic, Highnote and Unit made issuing your own cards straightforward. Ramp and Brex built clever controls on top of company spending. Privacy.com made throwaway virtual cards a consumer product. Plaid connected bank accounts to everything else.
Between them they have answered how money moves about as completely as it can be answered. Every one of them assumes a person or a company approved the spending. That assumption holds the whole thing up, and an agent spending on its own knocks it out.
The point: get into the flow of money early
Here is the argument, and it is worth stating simply.
The best place to make money has always been inside somebody else's transaction. You are not selling a thing that has to be chosen again each month. You are a toll on activity that was going to happen anyway, and the more the world uses you, the more you earn without lifting a finger.
That is why payments companies end up so valuable. Stripe did not win because developers loved its API more than they loved a competitor's. It won because it got into the flow, and once you are in a flow, growth is somebody else's job.
What is happening now is that a new flow is forming. Software is starting to spend money on its own. Nobody knows how big that flow gets, but it is clearly not zero and it is clearly growing. And right now, almost nobody is standing in it.
Agentcard has walked over and stood in it, very early, while the ground is still soft.
Be honest about the uncertainty, then look at the payoff
It is worth saying plainly what is not known, because pretending otherwise would be silly.
We do not know whether AI agents really need their own cards and their own financial identity. It is perfectly possible that agents stay firmly attached to their owner's account for a long time, that a virtual card with a tight limit turns out to be enough, and that the whole idea of an agent as a separate spender is something people decide they do not want.
But look at the shape of the bet.
If agents never become financially separate, this was an early idea that was too early, which is an ordinary outcome for a company days old, and the team will have learned a great deal about a part of the stack nobody else understands yet. And if agents do become financially separate, even partly, then whoever is already sitting in that flow with working infrastructure and merchants who accept it wins something very large, and they win it without having to take it from anyone.
That is a good trade. The cost of being wrong is one early company. The reward for being right is a position in a new payment flow. This is exactly the kind of bet worth making while the question is still open, and exactly the kind nobody can make once the answer is obvious.
It is also worth noticing what the hard part actually is. Moving the money is solved. What is not solved is proving that a particular agent was allowed to spend, up to what limit, for whom, and who carries the loss when that turns out to be wrong. Chargebacks exist because a human can say they did not approve something. Fraud models are trained on how humans buy. Limits are set by one person for another. An agent spending by itself breaks all three, and fixing that is the work.
Where else can you cut into a flow of money?
This next part is my opinion, not something the company has said.
The AI build-out is creating brand new flows of money the way the crypto boom did, and the people who got rich in that boom were mostly not the ones with the best technology. They were the exchanges, the custodians, the on-ramps and the settlement layers. They stood in the flow and took a slice.
The same opportunity is opening now, in more places than most people have noticed. Agentcard is one good example, and here are others where money is about to start moving and nobody owns the toll booth yet:
- Agent to agent payments. When one company's agent buys something from another company's agent, somebody has to clear it. That is a settlement business, and it barely exists.
- Paying publishers when an agent reads them. A vast amount of content is being consumed by machines for free. The moment that becomes a licence fee per read, somebody has to meter and distribute it.
- Metering and reselling compute. Inference is billed by the token today and by nobody in particular tomorrow. Whoever sits between the buyer and the model provider and counts accurately gets paid on every call.
- Escrow for work an agent completed. If a machine did a job, somebody has to hold the money until both sides agree it was done. That is an old business with a brand new customer.
- Insurance for what agents do. A genuinely new class of risk with no actuarial history. Premiums are a flow, and being first to price a risk nobody else can price is an enormous advantage.
- Tax on automated transactions. Sales tax, VAT and withholding all assume a human transaction with an address. Machine commerce will need someone to calculate and remit it, and that someone touches every payment.
- Royalties on training data. If data owners ever get paid per use, the plumbing that tracks and distributes those payments is a toll on the entire industry.
- Cross-border micropayments. Agents will make enormous numbers of tiny international payments. The existing rails are far too expensive for that, which means a gap.
- Energy and capacity settlement for data centres. Power contracts, cooling, unused capacity resold. All of it is trading with money moving and very little of it is standardised.
Each one has the same shape. A new kind of activity, real money attached to it, no established player positioned to take the slice, and a window that closes as soon as the flow is big enough for everyone to see. The people who move now are paying almost nothing for a spot that will be expensive later.
That is the general lesson worth carrying out of this file, and it is why a company only days old is worth writing about at all.
Why it is bigger than it sounds
The short description is payments for agents. What it really is, is an early claim on a new flow of money.
- Being in the flow is the best business model there is. You get paid when people use you, and the growth is somebody else's job.
- The uncertainty is honest and the payoff is lopsided. Nobody knows whether agents need their own financial identity, but if they do, whoever is already standing there wins something very large.
- Being first to a spot nobody is fighting over is cheap. The same position will cost a fortune the moment the flow is big enough for everyone to see.
What to watch
One thing settles whether Agentcard is real: whether a merchant or a card network accepts an agent's own credential as permission in its own right, rather than as a human's card in disguise.
Everything else is a wrapper. Virtual cards with tight limits, spending rules for agents and tidy audit trails are all useful, and they are all features on top of the human-approved model, which means the big players can ship them. The new thing would be somebody on the other side agreeing that an agent's permission counts as permission, with the question of who carries the loss written down. Until that exists, agent payments are a human's card with better logging. After it exists, they are a new rail, and whoever built it is standing in the flow.
Agentcard is days old and the evidence is not in. The spot it has chosen to stand in is the right one.