Should AI agents make payment decisions independently?

An IMF note by Sonja Davidovic and Hervé Tourpe argues that payments are undergoing a structural shift from "click-to-pay" — transactions a human explicitly initiates — toward "decide-to-pay," where "software agents operating under delegated mandates" anticipate payment needs, evaluate options, and coordinate execution "at machine speed and across multiple layers of the payment value chain." The note's central claim is architectural: because "core payment infrastructures are built on deterministic logic, requiring predictability, auditability, and legal enforceability at each step," while agentic AI systems rely on probabilistic reasoning and adaptive decision making, which can yield different outcomes under similar conditions, the two cannot simply be merged. The authors propose a three-layer model "to assess the appropriate scope for integrating agentic AI into payment workflows," so machine-speed agent judgment and rule-bound settlement each operate where they are suited.

The reasoning places agent autonomy in a specific, bounded role rather than treating it as a replacement for existing rails. The note frames "Layer 1" as "the locus for adaptive orchestration and delegated initiation, producing structured intent that must still pass through the deterministic authorization and settlement layers." Concretely, an agent instructed to "buy the latest book on topic x when it becomes available on any platform, at the best price" decides when and where to search and compares merchants, but the note observes that "humans typically remain involved in reviewing and finalizing purchases, as the mix of probabilistic decision making and deterministic execution can complicate accountability if issues arise." The authors state the stakes plainly: "the central issue is not whether AI should be used in payments... but how to keep uncertain, probabilistic decision making separate from automatic payment execution."

This gives an infrastructural rationale for a pattern Do consumers want to approve AI shopping purchases before they happen? documents as consumer preference, and the IMF note explains why that checkpoint exists at the architecture level, not just the UX level — probabilistic agent reasoning is deliberately kept from directly triggering deterministic, irreversible settlement. The three-layer model is also a structural cousin of the sequencing logic in Must AI monitoring precede incentive alignment as agents evolve?: both treat an AI system's growing autonomy as something that must be gated by an earlier-stage mechanism (verification layers, or deterministic authorization) before the system is trusted to act on its own judgment, though Humberd and Latham address organizational agency and the IMF note addresses payment-rail architecture.

The excerpt does not test the three-layer model against a working payment system or report any data on how often "structured intent" produced by an agent is rejected, modified, or passes unchanged through authorization. Vendor pilots it cites — Visa's Intelligent Commerce, Mastercard's Agent Pay, Google's UCP — are described only as early experiments, not evaluated outcomes. The note also flags algorithmic herding, flash crashes, and prompt injection as systemic risks without analyzing how the three-layer separation would contain them specifically. What the excerpt supports is a design principle — keep probabilistic judgment and deterministic execution in separate, bridgeable layers — not evidence that any current implementation achieves it.