AI Fraud Controls Face $8 Trillion Test as Agentic Payments Reshape Transaction Risk

The $8 Trillion Question: Can AI-Driven Fraud Controls Keep Pace With Agentic Commerce?

The expansion of online commerce is accelerating pressure on payment networks to close a widening gap between fraud prevention and legitimate transaction approval. With global online spending forecast to surpass $8 trillion by 2028, according to Visa, the financial stakes attached to getting that balance wrong — in either direction — have rarely been higher.

Two Costs, One Problem

The industry debate has long framed fraud prevention as a trade-off: tighten controls and decline more legitimate transactions; loosen them and absorb more losses. Primary data from Visa illustrates the scale of both sides of that equation. Global losses from false declines — genuine customers wrongly turned away — are projected to exceed $264 billion by 2027. On the other side, Visa has cited one study finding that up to 30% of customers may leave their bank following a fraud event, a figure that underlines the reputational and retention cost of allowing fraudulent transactions through. Readers should note that Visa attributes that 30% figure to a single unnamed study, so the statistic should be treated as indicative rather than sector-wide consensus.

Both Visa and J.P. Morgan have published material describing how AI and machine learning are being applied to this challenge — analysing transaction signals in real time to reduce false declines while improving fraud detection. The goal, as framed by both organisations in their primary communications, is to move beyond static rules toward adaptive systems capable of distinguishing genuine customers from fraudsters even as attack patterns evolve.

Payment networks are deploying AI and machine learning to analyse transaction signals in real time — with the dual objective of reducing the $264 billion false-decline problem while preventing the customer attrition that follows a fraud event.

The Agentic Frontier

The emergence of AI payment agents — software systems capable of initiating purchases autonomously on a consumer's behalf — introduces a new dimension to this challenge. A typical agentic payment stack, as commonly described by industry practitioners and reflected in published commentary from both Visa and J.P. Morgan, may involve multiple parties between a consumer's authorisation and the final settlement of a transaction, creating a chain in which accountability for an erroneous or fraudulent outcome is not straightforwardly allocated under frameworks originally written with human-initiated payments in mind.

J.P. Morgan has published a direct argument on this point: AI payment agents require enforceable controls — not merely prompt-based instructions — to safely operate in payments. The firm's published position is that a prompt can be overridden or manipulated in ways that a technical control embedded in the payment architecture cannot, though the precise attack vectors J.P. Morgan envisages — including the possibility of agents transacting outside their authorised scope — are an inference drawn from that published position rather than a separately enumerated factual claim in the same document.

Tokenization as a Scoping Mechanism

Both Visa and J.P. Morgan have discussed tokenization as a mechanism for scoping AI agent payment permissions without exposing full card credentials. Under this model, an agent would operate with a token that carries a defined scope — constraining what the agent can authorise — rather than holding the underlying payment credential. Whether Visa or J.P. Morgan have explicitly specified the parameters of that scope (such as merchant categories, spending amounts or time windows) in their published primary sources could not be fully confirmed from the materials reviewed; accordingly, this article describes the constraint as "a defined scope" without enumerating specific parameters.

Why It Matters

For European issuers and payment infrastructure providers operating under existing strong-customer-authentication and liability frameworks, the agentic payment model raises questions that current regulation did not anticipate. Issuers facing the prospect of up to 30% customer attrition after a fraud event — a figure Visa attributes to one study — have a direct commercial incentive to resolve the accountability and control questions before agentic commerce scales. The technical architecture decisions being made now, including whether enforceable token-level controls become standard practice, will shape who bears liability when an AI agent transacts incorrectly.