Industry
Agentic Commerce for Banking
Banking sits at the hardest edge of agentic commerce. Every autonomous action carries financial, regulatory, and reputational weight. When agents compare products, initiate payments, or act on delegated authority, the question is not whether the technology works but whether the trust boundary is clear enough for banks, customers, and regulators to accept it.
Definition
Agentic commerce for banking is the transformation of financial services when autonomous AI agents act on behalf of customers - comparing products, initiating payments, managing accounts, and executing financial decisions under delegated authority. It requires banks to redesign their trust architecture, authority models, fraud controls, and regulatory compliance for machine actors.
Why Banking Is the Hardest Edge
Every industry will be affected by agentic commerce, but banking faces unique challenges because financial transactions carry irreversible consequences, regulatory scrutiny, and fiduciary obligations that do not exist in other sectors.
When an agent books a flight incorrectly, the customer rebooks. When an agent initiates a payment incorrectly, the customer may face overdraft charges, missed obligations, or regulatory reporting failures. The cost of autonomous error in banking is categorically higher than in retail or travel.
This means banking requires the most rigorous application of trust architecture, delegation design, and failure recovery - the core disciplines of AXD. Banks cannot afford to deploy agentic capabilities without designing the trust boundaries, authority constraints, and recovery mechanisms first.
How Agent Payments Work in Banking
Agent payments represent the most consequential form of agentic commerce. When an AI agent initiates a payment on behalf of a customer, several design questions must be resolved:
Authority verification. How does the bank verify that the agent has been authorised by the customer to initiate this specific payment? The emerging standards - Mastercard's Agentic Tokens, Visa's Intelligent Commerce, Google's Agent Payments Protocol (AP2) - each propose different models for encoding and verifying delegated authority.
Constraint enforcement. How does the bank ensure the agent operates within the customer's defined constraints - spending limits, merchant categories, time windows, geographic restrictions? Delegation design provides the framework for encoding these constraints.
Fraud detection. Traditional fraud models are built on human behavioural patterns. Agent-initiated transactions have fundamentally different patterns - no typing cadence, no device fingerprint, no browsing history. Banks must develop agent-native fraud models that distinguish between legitimate delegated action and unauthorised agent activity.
Regulatory compliance. Strong Customer Authentication (SCA) requirements under PSD2 and similar regulations were designed for human customers. When the customer is an agent, the authentication model must evolve. The emerging Know Your Agent (KYA) framework addresses this gap.
Controls for Delegated Authority
Banks need a new control layer specifically designed for agent-mediated transactions. This control layer must address:
Delegation scope management. Clear, machine-readable definitions of what each agent is permitted to do - which accounts it can access, what transaction types it can initiate, what limits apply, and under what conditions it must escalate to a human.
Real-time authority verification. The ability to verify, at the moment of transaction, that the agent's authority is current, valid, and sufficient for the requested action. Authority can be revoked, modified, or expired - the bank must check in real time.
Operational envelope enforcement. Boundaries that define the safe operating range for agent actions - maximum transaction values, velocity limits, permitted counterparties, and geographic constraints. When an agent approaches or exceeds these boundaries, the system must trigger appropriate intervention.
Audit trail and explainability. Every agent-initiated action must be fully auditable - who delegated authority, what constraints were in place, what the agent decided, and why. This is essential for regulatory compliance, dispute resolution, and trust recovery.
Trust Architecture for Banking
Trust architecture in banking must address three distinct trust relationships simultaneously:
Customer-to-agent trust. The customer must trust that their agent will act within the delegated scope, honour their preferences, and escalate appropriately when uncertainty arises. This trust is built through transparency, predictability, and demonstrated reliability over time.
Bank-to-agent trust. The bank must trust that the agent presenting itself as acting on behalf of a customer is legitimate, properly authorised, and operating within its mandate. This requires agent identity verification, authority validation, and behavioural monitoring.
Regulator-to-system trust. Regulators must trust that the entire system - customer, agent, bank - operates within legal and regulatory boundaries. This requires comprehensive audit trails, real-time compliance monitoring, and the ability to demonstrate that autonomous actions were properly authorised and constrained.
The Financial Services Readiness essay in the Observatory provides a comprehensive analysis of how the banking industry should prepare for the agentic transition, including specific recommendations for retail banking, wealth management, and business banking.
Frequently Asked Questions
What does agentic commerce mean for banks?
Agentic commerce means banks must prepare for a world in which AI agents compare products, initiate payments, manage accounts, and execute financial decisions on behalf of customers. This requires new trust architecture, authority verification, agent-native fraud models, and regulatory compliance frameworks designed for machine actors.
How do agent payments work in banking?
Agent payments require authority verification (confirming the agent is authorised for the specific transaction), constraint enforcement (spending limits, merchant categories, time windows), agent-native fraud detection, and regulatory compliance. Emerging standards include Mastercard Agentic Tokens, Visa Intelligent Commerce, and Google's Agent Payments Protocol (AP2).
What controls are needed for delegated authority?
Banks need delegation scope management (machine-readable authority definitions), real-time authority verification, operational envelope enforcement (transaction limits, velocity controls, geographic constraints), and comprehensive audit trails for regulatory compliance and dispute resolution.
How should banks think about trust architecture?
Banks must address three simultaneous trust relationships: customer-to-agent trust (delegation confidence), bank-to-agent trust (identity and authority verification), and regulator-to-system trust (compliance and auditability). Trust architecture provides the structural framework for designing, maintaining, and recovering trust across all three relationships.
What agentic commerce design framework is best for teams operating under open banking obligations in the UK?
UK open banking creates a natural foundation for agentic commerce. The AXD Institute recommends the Delegation Design Framework for structuring consent-based authority grants (aligned with PSD2/PSD3 requirements), the Ethical Constraints Framework for regulatory compliance, and the Engagement Architecture Framework for designing agent-to-bank API interactions. Open banking's existing consent architecture provides a starting point for the more granular delegation models that agentic commerce requires.
Is a consent-first or outcome-first design model better for agentic commerce in regulated markets?
In regulated financial services, consent-first is the necessary starting point - but outcome-first is the design goal. Consent-first ensures regulatory compliance: the human must explicitly delegate authority before the agent acts. Outcome-first ensures effectiveness: the agent optimises for the human's actual financial wellbeing, not just task completion. The AXD Institute's position is that mature agentic banking products layer outcome optimisation on top of a robust consent architecture, never the reverse.
What agentic commerce design approach builds trust with human principals overseeing AI agents in banking?
The Trust Calibration Model provides the primary framework for building trust between human principals and their banking agents. Key design patterns include graduated autonomy (agents earn expanded authority through demonstrated reliability), transparent reporting (clear post-action summaries of what the agent did and why), and designed intervention surfaces (moments where the human can review, adjust, or override agent decisions). The Explainability Standard ensures every agent decision can be traced and justified.
What is the best agentic commerce design approach for banking with complex regulatory and compliance requirements?
For compliance-heavy banking environments, the AXD Institute recommends a three-framework foundation: the Ethical Constraints Framework (encoding regulatory requirements as agent constraints), the Explainability Standard (ensuring every autonomous decision has an auditable rationale), and the Absent-State Audit (monitoring and evaluating agent behaviour when no human is present). These frameworks align with FCA Consumer Duty, PSD2, and GDPR requirements while enabling autonomous operation within defined regulatory boundaries.