Most maps of the AI payments space are built around specific companies. This page does the opposite: it describes the layers of the stack so the map remains useful regardless of which players are active at a given moment. It is the reference model used by the rest of the AI Payments Map.
An AI-initiated payment passes through some or all of the following layers. A single company may operate in more than one layer; the point is to keep the functions distinct.
The layer that proposes a payment. Initiators include humans in checkout interfaces, automation tools, AI agents in consumer assistants, enterprise copilots, business systems such as ERP payment processes, and direct API calls under protocols such as HTTP 402.
A single enterprise may now have four or five parallel initiation paths. None of them were originally designed to route through one common downstream check.
The layer that evaluates whether a proposed payment should proceed, applies policy, and returns a final decision. In this map the outcomes are allow, hold, block, and escalate. See The payment decision layer.
Historically this layer was embedded inside other products. It is increasingly discussed as an independent layer as initiation fragments.
The layer that routes an approved payment between processors, acquirers, rails, and networks. Merchant-side examples include Stripe, Adyen, Checkout.com, and Airwallex; payment orchestrators include Primer, Spreedly, and Payrails. This layer decides how to execute a payment, not whether it should happen.
The rails themselves: card networks, bank rails (ACH, SEPA, SBP, FedNow and others), stablecoin networks, CBDCs, and internal ledgers. Execution is indifferent to intent; it moves money when instructed. See Execution infrastructure vs the decision layer.
Merchant-side fraud scoring (for example Signifyd, Riskified, Forter, Sift, Kount, Ravelin), bank-side AML and transaction monitoring (for example NICE Actimize, FICO, Feedzai), sanctions and PEP screening, wallet-risk tools, and crypto analytics. This layer produces scores and alerts rather than binding decisions. See Why fraud tools are not payment decisioning.
Agent identity schemes — Know Your Agent, Visa Trusted Agent Protocol, Mastercard Agent Pay, American Express ACE, Skyfire KYAPay, Cloudflare Web Bot Auth — plus user-consent formats such as AP2 mandates, and traditional KYC and KYB providers. At the moment, this is the fastest-moving layer.
A simplified flow for an AI-initiated enterprise payment:
1. An agent proposes a supplier payment — initiation. 2. The request reaches a decision layer that checks the mandate, counterparty, amount, sanctions, and operator rules — decisioning. Outcome: allow / hold / block / escalate. 3. If allowed, the request is routed to a suitable rail and processor — orchestration. 4. The rail moves the funds — execution. 5. Fraud, AML, and compliance checks produce signals that feed back into decisioning and monitoring — fraud / AML / compliance. 6. Identity and consent evidence — mandates, agent tokens — accompanies the request from initiation onward — identity / consent / verification.
Not a ranking. Not a vendor directory. It is a reference for functions. Companies move between layers; the layers themselves are comparatively stable.