AI agents can propose payments. Who is allowed to approve them?

An AI assistant can read an invoice, organise supporting information and prepare a payment proposal. That can save time. It does not establish that the proposed beneficiary is genuine or that the assistant has authority to move money.

The useful boundary is between preparing a request and accepting it for execution. An organisation needs to know which system enforces that boundary, what permission it checks and who can stop a questionable payment.

Recent announcements make the distinction timely

On 20 October 2026, Mastercard described an Agent Pay for Machines demonstration in which an agent purchased digital services within an approved budget. It also announced early access, rather than unrestricted availability.

On 21 October, Polygon described agent spend controls that check wallet requests against permissions, a stablecoin budget and an expiry. These are publisher descriptions of particular products, not assurance that every payment workflow is safe.

The defensive context matters too. Straiker's 21 October investigation reported AI-coordinated criminal activity against banking and payment targets, drawing on earlier evidence. That report does not establish that a particular approval checklist would have prevented the campaign.

Approve the exact payment

A useful review identifies the beneficiary, destination account, amount, currency and business purpose. Approval for one combination should not quietly become approval for another. A change to a supplier's bank details deserves its own verification through the organisation's approved process.

Consider a fictional assistant preparing an invoice payment. The amount fits the budget, but the request introduces a new account and demands immediate action. Staying below a numerical limit does not make that destination legitimate.

Independently contact the supplier or requester using a trusted contact already held by the organisation. Do not rely on the number supplied in the questionable message. A familiar name, confident explanation or convincing voice is not a substitute for that check.

Bound the execution authority

A spending limit is only one part of a mandate. Define the permitted task, allowed beneficiaries or actions, relevant limits, expiry and a way to revoke access. The payment system needs to enforce the approved terms before accepting an instruction. A prompt asking an agent to behave carefully cannot provide that enforcement by itself.

Human approval should sit at the points required by the organisation's risk assessment and payment policy. Some routine actions may operate within a previously approved, narrow mandate. New beneficiaries, changed terms, exceptions and uncertainty may require fresh human review. The person making that decision needs the exact proposal and supporting evidence, with time to challenge it before execution.

Resolve uncertainty before retrying

A timeout is not proof that a payment failed. Before another attempt, reconcile the transaction reference and outcome through the approved payment system. Record the original proposal, decision, current status and any permitted retry. Otherwise, an attempt to recover from an unclear response can create a duplicate payment.

Practise with a fictional treasury case

MICKAI® Training Academy's free course, Controlling payment proposals from financial AI agents, explores independent callbacks, exact beneficiary approvals and controlled retries using fictional cases. Study online or download the workbook to work through the decisions.

This is educational, non-accredited material. It is not a Polygon or Mastercard implementation guide, investment advice, a professional qualification or a guarantee against fraud. References to organisations do not imply their endorsement.

Primary sources

Product claims are attributed to their publishers. Publication dates below are the source dates, not the date of the activity described.

• Mastercard Agent Pay for Machines demonstration and early access Published 20 October 2026
• Polygon agent spend controls Published 21 October 2026
• Straiker CyberCartel investigation Published 21 October 2026