When Transaction Models Become Payment Infrastructure

The argument

Payment risk systems have long followed one pattern: a model scores upstream and a rule decides downstream. Visa Advanced Authorization and Visa Risk Manager are the familiar example. The model upstream is now changing. Transaction foundation models learn general representations of financial behaviour that transfer across tasks which once needed separate pipelines. Revolut’s PRAGMA, pre-trained on 24 billion user events, reported large gains in credit scoring, fraud recall and engagement against its own task-specific baselines. Mastercard is training a Large Tabular Model on billions of anonymised transactions, and its On-Demand Decisioning lets issuers run their own rules on the network.

The essay names the layer this creates: the conditioning layer. A score, segment or embedding-derived feature sits upstream of the rule and changes what the rule sees without changing the rule. An agent’s reasoning is scoped to one transaction and one principal. The conditioning layer persists across transactions and can be shared by institutions that never interact.

Deterministic authorisation, the design principle in the IMF’s April 2026 note on agentic payments, keeps the final gate auditable. The inputs to that gate are increasingly probabilistic. When a shared upstream model is retrained, for example after a new mule-account pattern appears, every issuer’s thresholds stay exactly as written and approval rates fall anyway. Each institution sees its own decline curve and explains it locally. None can easily see that others are moving the same way for the same reason.

Model-risk supervision is organised around the regulated entity, through model inventories, validation and vendor assessment. Correlated movement across several compliant institutions sits between those units. Existing tools can make the dependency visible: disclosure that an authorisation signal is produced outside the institution, by whom and on what retraining cadence; authorisation-impact testing after major upstream model updates; and monitoring for aligned movement across institutions that share a provider.

Jurisdictions that restrict cross-border pooling of payment data, India among them, meet the same choice early: concentrate model capability in private transaction corpora, or build it through consented, supervised and interoperable data infrastructure.

Key points

• A model can shape authorisation outcomes without approving any payment, by changing the inputs rules act on.
• The conditioning layer persists across transactions and can be shared across institutions.
• A shared model update can move approval, decline and step-up rates at many institutions with no rule change anywhere.
• Each institution’s view is locally correct, which makes the common cause hard to see.
• Disclosure of external signals, impact testing after upstream updates and cross-institution correlation monitoring would make the shared dependency visible.

Evidence and sources

• IMF, How Agentic AI Will Reshape Payments, April 2026
• Visa Advanced Authorization and Visa Risk Manager
• Revolut, PRAGMA foundation model paper (arXiv, 2026)
• Mastercard, Large Tabular Model announcement (March 2026) and On-Demand Decisioning (September 2025)
• RBI draft guidance on model risk management (June 2026) and RBI FAQ on storage of payment system data
• FSB, Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector, October 2025

Limits

• PRAGMA results are internal benchmarks reported in a research paper, without supervisory validation.
• No public source says Mastercard’s Large Tabular Model feeds On-Demand Decisioning; the essay’s point is adjacency.
• The comparison with pre-2008 value-at-risk convergence is limited to correlated decisions from shared model logic.