When an AI agent spends your money, two questions sit underneath the purchase. Who decided the agent was allowed to do that, and what checked that it stayed inside the decision.

Almost everyone building for agentic commerce says they answer both, and most do, one credential at a time. A network scopes its own token, an issuer limits its own card, a signed mandate covers one cart at one merchant. That gets an agent through one checkout. That is the easy part. It is also not the point.

I carry three or four cards and a bank account like most people. I want an agent that buys what I asked for and pays the way that works best for me: the card with the better earn on that category, the one with the trip protection on that fare, the account with the cash when the bill is due. That only works if the agent can see and spend across every issuer I hold, under one budget I set. An agent that cannot do that is a faster checkout button with a little search help. An agent that can do it without any control is a liability. The promise of agentic commerce lives in the cross-issuer budget, and so does the hard problem, because nothing in consumer payments today controls an agent at that level. A limit on one card is not a limit when the agent holds three.

What’s interesting is we have an answer for two questions at the top and it’s been around for decades, for thousands of cardholders at a time. The answer is the commercial purchasing card program. A P-card program is a delegation system first and a card product second, and once you line it up next to how an agent would operate for a consumer, the parallels are hard to miss.

Then again, so are the gaps, and the biggest one is the budget above the cards.

How a P-card program delegates spend

For readers who have never run one. A program administrator issues cards to employees, each with a profile. A single-transaction limit. A cycle limit. Merchant category groups that are allowed or blocked. Sometimes a velocity limit on transactions per day, sometimes a lock to one merchant. A status the administrator can flip from active to suspended in a minute. For payables, a virtual card number issued for one supplier at one exact amount.

The part that matters for agents is where those rules get enforced. They are checked at authorization, on the network, using the merchant category the acquirer sends rather than whatever the merchant says about itself. A purchase outside the profile is declined before the money moves. The administrator sees the decline, decides whether to allow it once or change the profile, and the cardholder tries again.

That real-time layer is a coarse filter. It knows amount, merchant, category, count and status, and nothing else. Everything finer sits in the policy underneath it: the cardholder agreement the employee signed, manager approval, receipt matching after the cycle, and the fact that the employee can lose the card and the job. The program works because both layers exist. Keep that in mind, because only one of them transfers.

The parallels

I built a clickable demo called Steward to work this through. Steward is a fictional service that sits between a consumer and their agents, requests the agent tokens and one-time credentials, sets the rules on them, and receives the decline alerts. Lined up against a P-card program, the pieces map almost one to one.

The consumer is the program administrator. The agent is the cardholder, though an agent is closer to a ghost card used by a process than to an employee with a badge, and that distinction comes back later. Steward is the administrator console and the token requestor. The agent token is the card issued to the cardholder. A one-time credential capped at the purchase amount is the virtual card issued for one supplier at one amount. Category scope on the credential is the merchant category group. The network declining on the acquirer’s MCC is the real-time authorization decline, and the agent’s stated purpose gets exactly the trust an employee’s does, which is none. The decline alert to Steward is the administrator notification. Pausing the agent and suspending its token is card suspension.

A P-card program maps almost one to one onto agent spendCompany P-card programConsumer with StewardProgram administratorConsumerCardholder (employee)AI agentAdministrator consoleSteward, the token requestorCard or one-supplier virtual cardAgent token or one-time credentialMerchant category groupCategory scope on the credentialDecline at authorization on MCCNetwork declines on acquirer MCCAdministrator notificationDecline alert to StewardCard suspendedAgent paused, token suspended

Two of the demo scenarios are straight out of a program administrator’s week. In one, a bill-pay agent splits a $318 bill into two payments to get under a $300 cap. Steward links the two by invoice and holds them, which is the split-purchase control every administrator knows. In another, a travel agent gets a trip budget with rules attached: airlines only, either of two cards, allowed to book on its own only when the fare carries free cancellation. That is a travel card with policy rules, and the agent picks between the cards the way a well-briefed traveler would.

The agent never asks permission. The credential’s rules say no for it.

What the consumer version is missing

The parallels cover the real-time layer. The policy layer underneath is where the consumer version has nothing, and each gap changes how something has to operate. There are four.

01
Gap 1 of 4 · The cardholder agreement

Nobody signed anything

The first is the cardholder agreement. An employee signed something, and misuse can be recovered through a conversation with a manager or a payroll deduction. An agent signed nothing and cannot be held to anything. So the authority has to travel with the credential itself, as a signed scope the agent cannot exceed, and an agent transaction that arrives without a scoped credential has to be declined rather than processed as an ordinary card-not-present purchase. That is a change in how issuers treat agent traffic, not a feature. And the pieces to make it work are missing today. No network rule puts a field in the authorization message that tells every issuer an agent is on the line. FIS now passes know-your-agent data to its own issuers, which is a start, not a standard. And the auth flow has no step where anyone checks the signed scope before the issuer approves. Without both, the issuer sees an ordinary card-not-present purchase and approves it like one.

02
Gap 2 of 4 · The program hierarchy · The big one

Three cards, one budget, no boss

The second is the program hierarchy, and it is the one that matters most. A company has one program, one issuer and one administrator view across every cardholder. A consumer has cards from several issuers, a bank account, and a wallet or two. Token requestors can already set limits on an agent token today, by amount, merchant (not category, very important), and expiry. What no one holds is the limit across all of them, and without it the agent cannot be trusted to do the one thing that makes it worth having, which is choosing among them. Closing that gap means one console that sees every credential an agent holds across funding sources and holds the budget above them, which means issuers accepting a limit set outside their own walls and the token requestor, not the issuer, becoming the place the budget lives.

03
Gap 3 of 4 · Reconciliation and review

Who checks the checker?

The third is reconciliation and review. In a company, someone in finance matches receipts and reviews exceptions, and that person is not the cardholder. In a household the consumer is administrator, cardholder and reviewer at once and will rubber-stamp fastest of anyone. The consumer version has to be exception-only: declines and holds come to the consumer, routine spend inside scope does not. That requires decline advice to flow back to whoever provisioned the credential, which today it does not. It stops at the issuer.

04
Gap 4 of 4 · Rail coverage

Off the rails

The fourth is rail coverage. P-card controls live on the card network because that is where authorization happens. The bank account path has no network in the middle and no authorization step, so a console like Steward is the only control point there and agent access to the account has to run through it or not at all. The six banks’ principles paper names agents typing card details into websites as an unsafe practice. In a program like this, it becomes a decline reason. And an agent driving the consumer’s own logged-in browser session carries no agent credential at all, so none of this sees it. That one has no clean answer yet.

+1
Not a gap, a different world · Liability

Who eats the mistake?

There is a fifth difference that is not a gap so much as a different world. A company carries corporate liability and recovers from the employee. A consumer is covered by Reg Z and Reg E, and an agent that overspends inside its scope is an authorized transaction under both. The P-card world never had to solve agent error because the company was always on the hook. American Express put the first money behind it in April, crediting cardmembers when a registered agent errs against a shared intent. No network has followed with a rule.

What exists and what does not

Most of the plumbing exists. It just does not yet operate this way.

Agent tokens and one-time credentials are live or in pilot at both networks, scoped by amount, merchant, and expiry, not by category. Token domain restriction controls exist for channel and device, not for MCC. Passkeys are in every agentic pilot as authentication, not as the signature on a scope. Agent signatures and key directories identify the agent to the merchant, not to the issuer. Decline advice reaches the issuer and stops. Bank data access through FDX is read access under a rule that is currently enjoined, with no initiation controls at all.

Then there are the pieces that do not exist. A required indicator in the authorization message that tells every issuer an agent is on the line, rather than one processor’s version of it. Decision alerts back to the token requestor. A limit that holds across issuers and funding sources. A dispute path for agent error.

Every item on both lists has an ancestor the same banks already run for commercial clients. That does not make it a flag to flip. Commercial controls sit on commercial platforms, priced by commercial economics, and are not wired to consumer card processing.

The honest version is that the design work is largely done and the build is a migration, plus one new thing no issuer can supply alone: the view across issuers.

Who has the right to win

The console in the middle of all this has to do two things. It has to see and control every credential an agent holds, across issuers and across rails. And it has to be present where the agent spends.

That narrows the field. The network token services see every token they issue and what it is for, but only on their own cards. Issuers see their own cards and nothing else, and some will argue control belongs at authorization where it already sits; Mastercard’s new score estimating whether a transaction was agent-initiated is a bet in that direction, treating the agent as a risk signal rather than a credentialed party. Under the issuers sit the bank cores, FIS, Fiserv and Jack Henry, which run cards and deposit accounts together for thousands of banks and credit unions and already carry their trust. They see both rails, but one bank at a time. FIS has started passing know-your-agent data to its issuers at authorization, the first sign a core wants a seat. Independent services like the fictional Steward have the right design and have to be let in by everyone else.

The parties that already hold both halves are the wallets that are also token requestors at scale, because a wallet is the one place a consumer’s cards from different issuers already sit side by side. Google has Universal Cart built on Google Wallet and checkout through Google Pay. Apple Pay is the largest consumer token requestor and is adding agent capability to the same device the consumer already trusts. PayPal sees the card, the bank account and the balance for the same person and sits at checkout nearly everywhere online that matters. Stripe’s Link is starting to look like a PayPal for agents, holding saved payment methods across merchants and issuing single-use cards for around 200 agents. Each of these can see across issuers, provision a credential, scope it, and be standing there when it is used. Paze is the banks’ own version. It is owned through Early Warning by seven of the largest US issuers and already holds about 220 million of their cards, so it starts with the cross-issuer view and the bank trust the others have to earn. What it lacks is presence: it is card-only, is still working to sign the top hundred merchants after its relaunch, and has not announced an agent product.

Which brings it back to what the schemes and the banks keep saying. Both networks and the six-bank paper frame agentic commerce as a trust problem before anything else: disclosure that an agent is involved, a consumer who can see and manage every delegation, auditable records, liability landing where the error was introduced. The same paper names wallets, devices and networks as the parties most likely to steer payment choice and gatekeep access. Having the infrastructure does not earn the seat.

So draw it as two circles. One holds the small set of players with the plumbing and the wherewithal to run a budget across issuers and rails. The other holds the players consumers and banks already trust with delegated authority over their money, and are willing to keep trusting under the banks’ own principles. The winner sits in the overlap. There are not many candidates there, and none of them has claimed it yet. Whoever does will be the one that makes an agent worth having, because they will be the one that lets it choose.

The winner sits in the overlap, and nobody has claimed it yetHas the plumbingcan run a budget across issuers and railsHas the trusttrusted by consumers and banks to act for themThe seatunclaimedNetwork token servicesGoogle WalletStripe LinkApple PayPayPalCard issuersBanksAmexBank coresPazeIndependent consolesright design, have to be let inAgent platformsown the agent, not the money

The demo is at demo.confiniumadvisory.com. If you run a P-card program or you are building agent checkout, I would like to hear where this breaks.