Agentic payments are payments that an AI agent initiates, prepares or manages on behalf of a person or business, within a mandate that sets what it may pay, to whom, how much and when. The agent decides that a payment should happen and carries out some or all of the steps; a person's authority sits behind each one.
Why do agentic payments matter?
Payments are where an agent's actions become hardest to undo. Once money has settled, getting it back depends on the other side's cooperation. Every control around payments was also built with a person in mind: authentication proves the customer is present, fraud models learn from human habits, and payee checks assume someone reads the warning.
An agent that pays changes all three. Under PSD2, strong customer authentication exists to prove the payer is who they say they are; an agent holding a customer's credentials defeats its purpose. Supervisors in every market expect the payment firm to stay responsible for what its systems do, which is why compliance teams such as those covered in AI agents for compliance teams in payments in Germany are asking where agents fit.
How does an agentic payment work in practice?
| Step | What the agent may do | What stays with a person |
|---|---|---|
| Mandate | Read and apply it | Set and change it |
| Payee | Use payees from verified records | Approve a new or changed payee |
| Amount | Pay within limits | Anything above the customer's limit |
| Authentication | Request it from the customer | Complete it |
| Execution | Submit routine payments, at depth 4 if certified | Payments outside the mandate |
| Record | File the verdict and its reasons | Review exceptions |
The payee step carries the most risk. An external claim, such as bank details in an emailed invoice, cannot authorise a payment. The agent may pay only to details that came from a record the business controls.
What does that look like in an example?
A distributor's accounts payable agent pays supplier invoices under 1,000 when each one matches a purchase order and a goods receipt in the purchasing system. An invoice for 800 arrives by email, matching both records, with the supplier's bank details already on file. The agent pays it. A second invoice for 800 arrives with different bank details in the footer. The agent ignores those details, pays nothing, and holds the invoice for the finance manager with a note that the account in the invoice differs from the account on record.
How is it different from agentic commerce?
Agentic commerce is the whole purchase: choosing, ordering, paying and handling what comes after. Agentic payments are only the money movement, which can happen with no shopping at all, as in the invoice example. They also differ from ordinary automated payments. A standing order executes a fixed instruction on a fixed date. An agentic payment involves the agent deciding that a payment is due, which is why it needs bounds rather than a schedule.
Which related terms should you know?
- Agentic AI: systems that plan and take actions, not only answer.
- Mandate: the recorded authority under which the agent pays.
- Merchant onboarding: the checks a payments firm runs before accepting a business, discussed in can an AI agent onboard merchants for a payments company?
The guide to AI agents in banking sets out where payment actions sit among the other things an agent may do.
Where Pryme fits
Through the Pryme Context Protocol, every fact behind a payment arrives stamped with where it came from, and an external claim, such as bank details in an emailed invoice, cannot authorise a payment. Paying money out sits at depth 4, which an agent reaches only on evidence from Attest, and every payment verdict is filed.
Questions people ask
Can an AI agent complete strong customer authentication?
It should not complete it on the customer's behalf. Authentication exists to prove the customer is present, so payments that need it should go back to the customer to authenticate.
Are agentic payments regulated differently?
In most markets there are no separate rules. Existing payment, conduct, data and financial crime rules apply in full, and the payment firm remains responsible for payments its agents make.