PSD3 and Machine-Initiated Transactions: Four Key Shifts
-
MIT EXEMPTION MISFIT The machine-initiated transaction exemption in PSD3 inherits assumptions from standing-order and recurring-billing patterns authorized by a human, which do not map cleanly onto autonomous agent-triggered payments.
-
SCA DELEGATION GAP Strong Customer Authentication requirements were designed around human biometric or device-bound credentials; PSD3 does not yet supply a delegated-authority credential model for AI agents acting within pre-authorized scope.
-
LIABILITY ALLOCATION GAP Where a machine-initiated payment fails, is fraudulent, or violates a consumer's revoked consent, no PSD3 provision explicitly assigns liability among the PSP, the AI orchestrator, and the consumer.
-
DELEGATED ACTS SILENCE As of published draft text, no confirmed EBA mandate addresses automated or agentic payment workflows specifically. The counterpoint position, addressed in full below, holds that existing PSR provisions applied with interpretive flexibility may be sufficient without bespoke categories.
The Rise of Delegated Agency in Commerce
The payment authorization frameworks that PSD2 established, and that PSD3 refines, were designed around a specific sequence: a human customer forms an intention, authenticates that intention through a device or credential they control, and issues a payment instruction to a PSP. Every compliance obligation in the chain, from SCA to consent recording to unauthorized payment liability, presupposes that sequence.
Agentic commerce breaks that sequence at the instruction step. A consumer who configures an AI purchasing agent and defines a spending envelope has authorized a class of future transactions, rather than a specific transaction. The agent then executes within that class, selecting counterparties, amounts, and timing autonomously. The payment instruction that reaches the PSP was not issued by the consumer at the moment it arrives; it was issued by software acting within parameters the consumer set at an earlier, separate point.
This structural displacement is a material systemic concern. As AI orchestration layers become embedded in retail, travel, utilities, and subscription commerce, the proportion of payment instructions arriving at PSPs without a concurrent human act will grow. PSD3 must eventually define the credentialing, liability, and consent-withdrawal conditions under which delegated agency in payments is governed. Those conditions are not yet in the text [1][3].
How PSD3 Redraws Payment Initiation for Non-Human Actors
PSD3 and the companion Payment Services Regulation restructure the authorization chain by separating the framework directive (PSD3) from the directly applicable conduct rules (PSR), moving key technical standards into EBA-issued regulatory technical standards and anticipated delegated acts. Consent, authentication, and liability allocation each receive treatment, but the treatment proceeds from the same foundational assumption embedded in PSD2: that a natural person issues an instruction at or near the point of transaction.
For machine-initiated transactions, the existing MIT exemption requires that the payer has authenticated the initial mandate and that all subsequent payment orders execute within the parameters of that mandate. This model accommodates gym subscriptions and utility direct debits. It does not accommodate an AI purchasing agent that dynamically selects a supplier, negotiates a price, and initiates a payment within a pre-authorized spending envelope, because the specific counterparty and amount are not known at mandate creation. The payee-initiation condition, which requires that subsequent payment orders originate with the payee rather than the payer's own infrastructure, applies specifically in the card-based MIT context; in direct-debit arrangements, payer-side initiation within agreed parameters is permitted. An agentic orchestrator sits on the payer side, initiating from within the consumer's own systems, and does not satisfy the card-MIT payee-initiation condition; it also exceeds the direct-debit model because the counterparty and amount are determined by the agent at runtime rather than fixed in the original mandate.
Authentication under PSD3 carries forward from the RTS on SCA under PSD2 the same three-factor framework: possession, knowledge, and inherence, each bound to a legal person or a device in that person's control. PSD3 does not revise this inheritance in ways that create a statutory home for delegated-authority credentials. A delegated authority model, whereby a consumer pre-authorizes an agent to authenticate on their behalf within defined parameters, therefore has no explicit statutory basis in current PSD3 draft language. The practical implication is that a PSP processing a machine-initiated payment faces a binary choice: apply a SCA exemption whose conditions are not fully met, or require a full SCA step that interrupts the autonomous flow the agent was built to execute.
The liability allocation gap aggravates both the authentication and exemption deficits. PSD3 assigns unauthorized payment liability to the PSP unless the payer acted fraudulently or with gross negligence. An AI orchestrator sits outside this bilateral structure. When the orchestrator acts outside the consumer's instructed parameters, no provision in current draft text specifies whether the PSP, the orchestrator's operator, or the consumer bears the resulting loss. Delegated acts could address this, but no published EBA mandate confirms they will do so.
Key Materials
-
PSD3 legislative proposal and PSR text (European Commission, 2023): The primary legislative source for consent, SCA, and liability provisions; the structural basis for all extrapolation in this brief.
-
EBA roadmap on payment innovation and technical standards: Sets out the anticipated scope and sequencing of regulatory technical standards under PSD3, relevant to the delegated acts horizon.
-
Regulation (EU) 2024/886 on instant payments: Amends the SEPA Credit Transfer Regulation; its phased near-instant settlement mandate creates the fraud-risk window that agentic initiation will increasingly inhabit as obligations come into force across PSP categories.
-
Coche, Kolk & Dekker (2024) on EU data governance [3]: Documents structural misalignments between regulatory data-sharing mandates and the governance architecture intended to support them under the EU Data Governance Act and related instruments. The consent-architecture parallels drawn in this brief are the author's inference from structural analogy, not a direct finding of that paper.
Operational Consequences for Merchants and Payment Service Providers
For payment service providers, the immediate operational consequence is a compliance classification problem. Every machine-initiated transaction in an agentic commerce flow must be assigned to an existing exemption or authentication pathway. Where the MIT exemption is applied to transactions that do not satisfy its payee-initiation or pre-agreed-amount conditions, the PSP carries the unauthorized payment liability exposure that PSD3 assigns when a valid exemption is not in place. Fraud detection systems built on behavioral biometrics and device-bound signals will receive transactions with no human behavioral signature; the anomaly models these systems rely on will require architectural revision, not parameter adjustment.
For merchants, the consent recording obligation becomes both more demanding and structurally ambiguous. A merchant receiving a payment from a consumer's AI agent must satisfy itself that the original mandate covers the specific transaction. Where the agent has exercised discretion on amount or counterparty selection, the audit trail linking the payment to the consumer's original authorization is attenuated. If a consumer disputes the transaction on the grounds that the agent exceeded its instructed scope, the merchant's position in the chargeback process under PSD3's shifted liability framework is untested.
Transaction monitoring obligations under anti-money laundering rules and PSD3's fraud-reporting requirements attach to the PSP in the first instance. When the initiating actor is an orchestration platform operating across many consumers simultaneously, the monitoring unit of analysis shifts: a single suspicious behavioral pattern may present as distributed micro-transactions across thousands of accounts. The aggregation and escalation logic built for individual account monitoring does not accommodate this architecture without deliberate redesign.
Regulation (EU) 2024/886 introduces instant payment obligations in phased timelines: euro-area PSPs are required to receive instant credit transfers from January 2025 and to send them from October 2025, with non-euro-area timelines following later. As each phase comes into force, the ten-second settlement window progressively closes the time available to intervene before funds move. That compression arrives precisely when the payment initiation model is most opaque, and PSPs whose agentic transaction volumes grow in step with the rollout schedule will face the narrowest fraud-detection window at the point of highest operational uncertainty [2].
PSD3 arrives with a consent and authentication architecture calibrated for a natural person at a checkout. The commerce infrastructure it will govern is being built for agents that select counterparties without human instruction, authenticate without biometric or device-bound credentials tied to a natural person, and settle within windows that precede any manual intervention. Each of those three displacements, from human intent to parameterized delegation, from personal credential to agent-held token, from deliberate review to sub-second clearing, generates a distinct compliance surface that the current text does not address. The delegated acts process is the remaining legislative instrument capable of closing those gaps, and the absence of a confirmed EBA mandate to do so is the most consequential open variable in the PSD3 horizon for any operator building on agentic payment infrastructure.
The Case Against Prescriptive Machine-Initiated Rules
The strongest case against introducing PSD3-specific rules for machine-initiated transactions at this stage rests on the pace of architectural change relative to the legislative cycle. Agentic commerce infrastructure is still being assembled; the orchestration models, credentialing conventions, and delegation governance patterns that will dominate in five years are not yet settled. A delegated act that assigns liability categories or defines authentication standards for AI agents now risks encoding one particular architectural pattern as the regulatory reference point, making subsequent technical evolution a compliance problem rather than a commercial choice.
Proponents of this position would further observe that the existing PSR framework, applied through EBA technical standards with sufficient interpretive flexibility, can accommodate machine-initiated flows without bespoke categories. A well-scoped delegated authority mandate, issued and recorded by the consumer through a PSP interface, maps onto existing consent frameworks. The liability gaps that appear acute today may resolve through contract law and commercial terms between PSPs and orchestration platform operators, as occurred with acquirer-processor relationships under PSD2, without requiring statutory intervention.
The specific mechanism that gives this concern weight is accreditation asymmetry. Detailed rules for agent payment initiation that reference specific technical architectures become barriers to entry for smaller PSPs and fintech operators who lack the compliance infrastructure to implement them. Large platform operators with dedicated regulatory teams absorb the requirements and convert them into accreditation advantages that smaller entrants cannot match. Prescriptive rules designed to govern concentration in agentic commerce may, through this mechanism, reduce the number of operators capable of meeting them and concentrate the market further [2][3].
Unresolved Questions on Agent Authorization
-
Whether PSD3 or the PSR introduces any statutory category for machine-initiated payment orders distinct from the MIT and recurring-transaction exemptions, or whether those exemptions are expected to carry the full weight of agentic initiation by interpretive extension, lacks authoritative confirmation in published draft text.
-
The mechanics by which a consumer withdraws consent from an active AI agent mid-authorization period, and how that withdrawal is communicated to the PSP before the next transaction executes, remains unaddressed in current regulatory technical standard mandates.
-
Which supervisory reporting obligations attach to an AI orchestration platform that initiates payments across a large consumer population is contested, given that orchestrators do not hold PSP licenses under the current framework.
-
Whether the phased rollout of near-instant settlement under Regulation 2024/886 creates a graduated obligation on PSPs to develop non-behavioral fraud signals adequate for machine-initiated flows awaits guidance that has not been issued.
Sources
[1] Borgogno, O., & Colangelo, G. (2019). Data sharing and interoperability: Fostering innovation and competition through APIs. Computer Law & Security Review.
[2] Westermeier, C. (2020). Money is data: the platformization of financial transactions. Routledge.
[3] Coche, E., Kolk, A., & Dekker, M. (2024). Navigating the EU data governance labyrinth: A business perspective on data sharing in the financial sector. Alexander von Humboldt Institute for Internet and Society.