When an AI Agent Pays on the Customer's Behalf
bankingNovember 18, 2026

When an AI Agent Pays on the Customer's Behalf

Specifying Mandates, Consent and SCA for Agentic Payments

Customers are starting to ask AI assistants to do more than recommend. They want them to book the train, renew the subscription, reorder the printer ink and pay for it. The card networks have launched agent-specific token programmes, open protocols for agent-initiated payments are maturing, and the first live agent-executed payments have already gone through European banks. For a bank, the question is no longer whether a third-party AI agent will try to start a payment from a customer's account. The question is which rules will govern it when it does. 

These rules require explicit definition and governance. Many conventional customer-initiated payment journeys rely on the payer reviewing the amount, identifying the payee and authenticating at checkout. Other established models, including recurring payments, already separate consent from individual execution. Agentic payments introduce a different delegation model: the customer may define an instruction in advance, and an agent may subsequently initiate one or more payments. The bank must specify the scope of that authority, applicable limits, revocation controls, authentication requirements, liability handling and evidence needed to reconstruct each decision. 


These are business decisions before they are technical ones. This article sets out the six that need to be made, from the perspective of the business analyst who has to turn them into requirements a delivery team can build. 


The regulatory starting point: no special regime, and a new one on the way 

There is currently no standalone EU regulatory regime specifically for agent-initiated payments. Depending on the payment instrument, initiation model and roles of the participants, existing PSD2 requirements and the Regulatory Technical Standards on strong customer authentication (SCA) and secure communication may apply. Dynamic linking requirements must be assessed where relevant; they cannot simply be assumed to have been satisfied by authenticating a broad mandate. An agent acting on behalf of a payer is not automatically equivalent to a merchant-initiated transaction (MIT), and MIT treatment should not be reused without legal and scheme-specific analysis. 


The regulatory landscape is evolving. The PSD3 and Payment Services Regulation (PSR) package reached political agreement in November 2025, with compromise texts published in April 2026. The Council adopted its first-reading position on 29 October 2026. Final adoption, publication and application dates should be checked against the official legislative record before publication. The proposed reforms address, among other matters, fraud prevention, liability and open-banking permission management; any implications for agent mandates require further assessment rather than automatic equivalence. 


Banks should design for applicable requirements today while allowing for future regulatory and scheme changes. Each requirement should identify whether it reflects a binding obligation, an approved legal interpretation, a scheme rule or an internal risk-control decision. Not every future regulatory change can be addressed through configuration alone; some may require changes to processes, integrations or customer journeys. 


Decision 1: what the agent is allowed to do 


A useful starting point is the mandate: a structured representation of the customer's delegated authority for a specified agent and defined payment activity. Where the bank controls or relies on that delegation, maintaining an authoritative mandate record, or an equivalent verifiable source of authority, is a strong architectural recommendation rather than a universal legal requirement. The bank must be able to determine what authority was granted and whether it remains valid when a payment is requested. 


A mandate specification needs at least the following: 

  • Which agent. The agent is identified and verified, not just "whoever presents a valid token". Agent identity is becoming a scheme-level and protocol-level capability, and the bank should decide which identity assurances it accepts. 
  • Which payment types. Card payments, SEPA credit transfers, instant payments, or a subset. A mandate for online shopping shouldn't silently cover account-to-account transfers. 
  • Which payees. A named merchant, a list, merchant categories, or "any". Each option carries a very different risk profile. 
  • Purpose. Recorded in the customer's own words and as a structured category. This is the reference point when a dispute asks whether a purchase matched the intent. 
  • Validity. A start date and a firm expiry. Open-ended agent mandates are the long-term risk most often overlooked. 


Emerging protocols use signed records of customer intent and transaction-specific purchase details. Banks should assess the assurance and interoperability of each protocol and preserve sufficient mandate information or verifiable references to avoid depending solely on an agent provider's uncorroborated assertion. 


Decision 2: spending limits 

Limits turn a mandate into an enforceable control. Although the requirements may appear straightforward, each requires an explicit policy decision: 


A per-transaction maximum, and a cumulative limit per day, week or month. 

Limits per payee or category within a broader mandate. For example, up to €200 a month on groceries, but no more than €50 at any single new merchant. 

Currency handling. Is a limit in euros applied to a payment in another currency at the bank's rate at the time of payment, or blocked altogether? 

How agent limits interact with the customer's existing account, instrument and available-funds constraints. Agent-specific limits should be additional restrictions, not a means of bypassing controls that would otherwise apply. 


What happens when a customer has several agents. Each agent may have its own limits, but the bank may also want an aggregate ceiling on all agent spending. 

Specifications should define a single hierarchy of applicable limits and identify the system responsible for enforcing each one. They should also address concurrent requests, atomic limit checks or reservations, settlement adjustments, reversals, refunds and currency conversion, so two simultaneous transactions cannot independently consume the same remaining allowance. 


Decision 3: how permission is withdrawn 


Customers should have a clear, accessible way to revoke delegated authority, preferably within the bank's own app where the bank manages that authority. Revocation should prevent new payment requests from being authorised under the revoked mandate as soon as it becomes effective in the relevant systems. Payments already initiated must be handled according to processing status and applicable irrevocability rules; revocation does not automatically reverse completed or irrevocable transactions. Requirements must also cover scheduled requests, propagation to connected channels and notifications to agent providers.

Automatic expiry and suspension deserve equal attention: examples include reaching the validity end date, repeated declines, fraud indicators or changes in account status. A customer-facing dashboard showing active agents, scope, limits and spending is a valuable control to consider from the first release. The design should also specify mandate lifecycle states, versioning and how in-flight requests are handled during a status change. 


Decision 4: when additional authentication is required 

SCA is among the most consequential design questions because the authentication treatment depends on the payment type, initiation model and applicable rules. Banks should distinguish authentication when a mandate is created or materially changed from authentication of subsequent transactions. They should assess whether an exemption or another compliant mechanism is available for each flow, including any relevant dynamic-linking requirements, rather than assuming that initial mandate authentication automatically covers future payments. 


Potential risk-based step-up triggers include a new payee within the permitted scope, an unusual spending pattern, a higher-risk transaction or a material mandate change. A payee outside the mandate's scope is not merely a step-up case: the request must be rejected under that mandate unless the customer separately authorises the transaction or updates the authority through an appropriate flow. The bank should define secure customer approval channels and a fail-safe outcome when the customer cannot be reached; real-time contact cannot be guaranteed. 


Decision 5: who is liable in a dispute 

Agentic payments can generate several distinct dispute types. Requirements should separate disputes about payment authorisation from disputes about the underlying purchase or the agent's performance: 


The agent misunderstood the customer. It bought the wrong item, the wrong quantity or the wrong date, but stayed within the mandate. 

The agent was manipulated. A malicious website or product listing steered it into a purchase the customer never intended. 

The merchant failed to deliver, exactly as in any other purchase. 

The customer says they never authorised the agent at all. 

Under PSD2, whether the payer consented to a payment is central to determining authorisation. A payment executed within the scope of a valid mandate may be authorised if the applicable consent, authentication and regulatory conditions have been met, but the mandate alone does not conclusively prove authorisation in a dispute. Evidence and the payment service provider's statutory obligations remain important. Responsibility may involve the bank, agent provider, merchant or other payment service providers depending on the dispute, payment instrument, contractual arrangements and applicable law. The bank should define customer remediation, investigation ownership, evidence collection and any contractual recovery rights without implying that contractual terms remove mandatory customer protections. 


Decision 6: what evidence is kept 


These disputes require reliable, proportionate evidence. For each agent-initiated payment, the bank should be able to retrieve: 

the mandate identifier and version in force at the time; 

the verified agent identity; 

the instruction context the agent presented, ideally as a signed record of the specific purchase; 

every authentication event and risk decision, including step-ups triggered and their outcomes; 


the limit checks performed and the remaining balance on each limit; 


any revocation, suspension or change affecting the mandate, together with timestamps, transaction and correlation identifiers, decision reason codes and relevant system events. 


Retention must comply with applicable recordkeeping and data-protection obligations, rather than relying on a single generic payment-record period. The bank needs proportionate evidence of the authority and transaction, not necessarily the customer's entire conversation with the agent. Access controls, integrity protection and a clear retention schedule are essential. 


Turning decisions into requirements 


Once the six decisions are made, they can be translated into testable requirements with explicit triggers, expected outcomes and exception handling: 

When an agent-initiated payment would cause cumulative spending to exceed the remaining monthly limit of the active mandate, the system shall reject the request, record the reason and notify the customer through the configured channel. 

 

Upon mandate revocation, the system shall prevent new payment requests from being authorised under that mandate. Pending and scheduled payments shall be assessed according to their processing status and applicable cancellation and irrevocability rules. When an agent submits a payment request for a payee outside the mandate's authorised scope, the system shall reject the request under that mandate, record the validation failure and return an appropriate response to the initiating party. 


Each requirement should trace to a policy decision, accountable owner and applicable regulatory or scheme basis. Product may own mandate scope, fraud may own risk triggers, legal and compliance should validate the regulatory interpretation, and operations should own dispute handling. Implementation also requires coordination across payment orchestration, authentication, fraud monitoring, card or account processing and audit services. Idempotency keys, concurrent requests, retries, timeouts and mandate-version checks should be specified so the same controls are enforced consistently across channels. 


Preparing banks for agent-initiated payments 

The banks that manage agentic payments effectively will not necessarily be the first to support them. They will be those that define delegated authority, limits, exceptions, evidence and ownership clearly, then implement those decisions consistently across their systems. From a business analysis perspective, the actor is new, but the discipline remains familiar: establish what it may do, enforce the rules at execution time and retain evidence of what occurred.