Agent authorization
Agent authorization is the explicit grant under which an AI agent acts on someone’s funds: which operations it may perform, on which accounts, up to what amounts. It is scoped per relationship (a platform’s agent serving many customers holds a separate grant from each), and it is revocable at any time without touching anything else the agent does. The grant is carried by a key that can pay and nothing else: it cannot create agents, mint keys, or read another agent’s transfers. Authentication proves which agent is calling; authorization defines what it may do.
By Krisan Nichani, Chief Compliance Officer
The pattern that has emerged is invite-and-approve: the party whose money will move is asked, sees exactly what is being requested (permissions, per-transaction caps, duration), and grants or declines. The grant is recorded, and every later payment traces back to it, which is what an auditor will ask for when software moved customer funds. Broad, standing, implicit authority is the anti-pattern; it turns the first bug into a question about every payment the agent ever made.
This is the agentic version of a problem payments has solved before: direct debit mandates and open-banking consents are authorization records for earlier kinds of automated money movement. On Infinite, moving money for customers is the third-party funds flow: customers are onboarded and verified, and the authority a platform or its agent acts under is documented as part of the flow. Revoking an agent is instant and total, and every key it holds dies with it.
Frequently asked questions
What is the difference between authentication and authorization for an AI agent?
Authentication is identity: the credential proves which agent is calling. Authorization is permission: the recorded grant saying what that agent may do with whose money, and up to what limits. A stolen credential fails at authorization if the grant is scoped tightly.
Can a customer revoke an agent’s authorization?
Yes, revocability is the point. A grant is per-relationship, so a customer can withdraw their agent’s authority without affecting other customers, and the platform can do the same from its side. Payments after revocation simply fail authorization, and revoking the agent itself goes further: every key it holds dies instantly.
What should an agent authorization specify?
Four things at minimum: the operations permitted, the accounts in scope, the amount caps per transaction and period that the network will enforce as the money moves, and how it ends, expiry or revocation. Anything the grant doesn’t name, the agent shouldn’t be able to do.
See it on your own flows.
A 30-minute walkthrough of a live account and the screening record behind each payment.
Ask an AI about Agent authorizationChatGPTClaudePerplexityGemini