What the Corporate Card Already Solved

Payments spent fifty years building a delegated-authority architecture for the one kind of authority we got serious about early: money. An instrument issued for an approved purpose. An approval bound to what the approver was shown, eventually by law. Delegation that only narrows. A network that authorizes every transaction against current state. Endings, disputes, and statements that close the loop.

The claim of this series is not that the card is a good metaphor. It is that enterprise finance independently discovered the governance architecture that agent authorization now requires, and proved it at global scale. The card is not the hero of that story. It is only the credential the governance system issues, and the one sentence that carries everything:

Enterprise spend is not controlled because cards are secure. It is controlled because every card is continuously bound to a business purpose.

Then payments did it a second time. In the last decade, spend platforms rebuilt the card itself: approval mints the instrument, policy rides the authorization, and the market rewarded the upgrade. The pattern in this series is not a proposal. It is the direction the money world has now moved twice.

To be precise about the title: the corporate card did not solve AI agent authorization. It solved the control pattern, issue for a purpose, authorize each use, narrow every delegation, and make endings real, for the one kind of authority where we demanded nothing less.

AI agents need the same architecture for authority that is not money, in a world with no payment network and no honest cardholder. This series teaches the mental model with no protocol in sight, and it is written to be sent to the people who will never read a spec. Part 1 is the flagship: the whole expense-governance loop end to end in about fifteen minutes, and the one post to send if you send only one. The four parts after it each take one control in that loop and go deep.

Every post names where the analogy breaks. An analogy this convenient should be distrusted on schedule, and the breaks are the build list: each one marks something the agent stack cannot borrow from the expense world and has to build.

The failure mode is not autonomy. It is the blank check. The through-line is simple: do not give the agent a credential and hope the rest of the system remembers why it exists. Record the approved purpose, project it into a bounded instrument, check every meaningful use against live state, narrow every delegation, and make the ending reach the places where the authority was projected.

The whole mapping fits in one table, and each row is one post:

And each post distills to one rule:

1. Credentials are projections, not authority.
2. The approval binds the disclosure, not the intent.
3. Delegation narrows, or it contaminates.
4. Authorize the action, not the instrument.
5. Ending the credential is not ending the arrangement.

The Corporate-Card Test

The rules compress into a diagnostic. By the end of the series you should be able to look at any agent platform and ask five questions:

1. What approved purpose is this authority projected from?
2. What exact disclosure did the approver see?
3. Can delegated authority only narrow?
4. What authorizes each consequential action at the moment of use?
5. What actually stops when the work ends?

If the answers would be unacceptable for a corporate card program, they are unacceptable for an agent. Part 1 closes with the long form of the test.

The Five Controls

1. Agents Need a Corporate Card, Not a Blank Check. The whole loop. Nobody hands a new hire the company checkbook. The corporate-card pattern maps, control by control, onto what agent authorization is missing, with the five global breaks named.
2. You Approve What You Were Shown. The approval. Requests are proposals, reviewers narrow more often than they deny, approvals take days, and the thing that binds is the disclosure the approver actually saw. Payment regulation made that last part explicit.
3. The Contractor Gets Their Own Card. The delegation. Lending your card is the convenient failure everyone recognizes. The right answer is a narrower instrument with its own name on it, one that dies with the project.
4. The Network Approves Every Transaction, Not the Card. The enforcement. Possession was never the control. Every swipe is authorized against current state with the parameters bound into the authorization, declines are the system working, and the ATM is the escape hatch every honest program names.
5. Canceling the Card Doesn’t Stop the Charges. The ending. Freezes, cancellations, the subscription that bills the new number you never gave it, charges caught mid-air between authorized and settled, disputes as governed undo, and the statement that joins it all.

The protocol work is not inventing a new control model. It is translating one enterprises already trust for money into authority agents can safely carry. This series stands on its own, and the translation exists: From the Card to the Architecture maps each rule here onto the architecture, and the Mission-Bound Authorization series carries it from there. And if the analogy is wrong somewhere, the breaks sections are where to aim, and issues on the draft repository are where the argument lands.