Abstract:
This article examines the EU-law challenges arising from the use of artificial intelligence for payment initiation and execution ("Payment Agents") under European Union law. It focuses on Payment Agents that support purchasing workflows and are capable of initiating payments in both human-in-the-loop and human-out-of-the-loop environments. The article conceptualises Payment Agents as agentic systems and develops three regulatory models: (i) a protocol-only model, (ii) a model based on the involvement of a licensed payment service provider under PSD2, and (iii) a contract-based model with separated roles for a Payment Agent Provider and a Credentials Provider. It concludes that Payment Agents will, as a rule, qualify as "AI systems" within the meaning of the AI Act, whereas payment protocols should be understood as transactional infrastructure rather than general-purpose AI models. Typical agentic payment use cases are not currently listed as high-risk AI systems under Annex III of the AI Act. Under PSD2, the protocol-only model may amount to payment initiation services when it initiates transactions from a payment account, which implies licensing and strong customer authentication. This configuration resembles screen scraping. Agentic payments are not equivalent to merchant initiated transactions (MIT), where the payee initiates the transaction within scheme processing. In the contract-based model, the Credentials Provider can often be aligned with a pass-through wallet and a technical services provider. Depending on design, the arrangement may resemble a payment scheme and may trigger outsourcing under the EBA outsourcing framework and, for ICT services, DORA. The article concludes that existing legal institutions address the risks only partially. It argues that mitigation would be stronger if agentic payments were expressly recognised as a high-risk use case under the AI Act.