The Argument
The identity schism is the fundamental philosophical divide in agentic payments, defined by whether an AI agent is treated as a new, visible entity or as an invisible extension of the human user. This is not a technical detail but the foundational design decision of agentic commerce, determining how trust, accountability, and control are structured. Mastercard’s Agent Pay model treats the agent as a distinct, registered participant with its own identity, fostering mechanical trust through verification. In contrast, Visa’s Intelligent Commerce treats the agent as a delegated capability of the human, relying on relational trust through the existing cardholder relationship. Neither model, however, adequately addresses the comprehensive trust architecture required for mature agentic systems, particularly concerning failure recovery and dynamic consent.
The Evidence
Mastercard's Agent Pay framework establishes the AI agent as a first-class citizen in the payment ecosystem. By issuing a unique Agentic Token, the agent becomes a visible, governed participant. This architectural choice anchors trust in the agent's own verified identity, a model of mechanical trust. Every party to a transaction - issuer, acquirer, and merchant - can see that an agent initiated it, allowing them to calibrate their trust and risk models accordingly. This approach is suited for higher levels of agent autonomy (Stripe's Levels 4-5), where an agent operates independently, making it more of a distinct market actor than a simple tool. The visibility of the agent provides a clear target for accountability and the potential for agent-specific dispute resolution frameworks.
Conversely, Visa's Intelligent Commerce model treats the agent as an extension of the human, using the human's tokenized credentials under the Trusted Agent Protocol (TAP). This approach builds on relational trust, where authority flows from the human's existing, trusted card relationship. The agent is functionally invisible at the transaction level, appearing as a standard card payment. This creates a trust attribution gap, as merchants cannot distinguish between a human-initiated purchase and an agent-initiated one, complicating risk assessment and failure analysis. This model aligns well with lower levels of agent autonomy (Stripe's Levels 1-3), where the agent acts as an assistant under close human supervision, but struggles to account for the agent's independent judgment in more autonomous scenarios.
Neither payment giant has adequately designed for recovery when an agentic transaction fails. The existing chargeback system, built for human disputes, is ill-equipped to handle cases where an agent acts within its technical constraints but against the user's nuanced intent - a category of failure for which there is no precedent. Mastercard’s visibility offers a potential foundation for agent-specific recovery processes, but these remain undeveloped. Visa's invisibility makes applying even the current, flawed chargeback model more difficult, as the dispute is not about fraud but about the agent's judgment. This oversight highlights a critical gap: both have engineered for successful execution but have not designed the failure architecture essential for building lasting user trust.
The Implication
The identity schism forces practitioners to design for a hybrid reality. Since both Mastercard's entity-based and Visa's extension-based models will coexist, any robust trust architecture must function with both. This requires building an independent visibility layer that records agent actions and decisions, creating an audit trail separate from the payment rails. Product leaders and designers must prioritize failure architecture over execution, designing recovery pathways for when an agent inevitably makes a poor judgment call. This means moving beyond static, set-and-forget permissions to implement dynamic consent mechanisms, allowing users to observe and adjust an agent's authority in real time, treating the Consent Horizon as a living boundary.
Ultimately, organisations must prepare for an identity transition. An agent may begin as a supervised extension of the user but evolve into an autonomous entity as it earns trust. Systems must be designed to manage this evolution gracefully, shifting from a relational to a mechanical trust architecture without creating friction. The core task for designers is to recognize that the payment networks are solving a payment problem, not a trust problem. The true work of Agentic Experience Design is to build the comprehensive trust relationship - encompassing delegation, observation, intervention, and recovery - within which these payment solutions operate. The future of agentic commerce will be defined not by the payment rails, but by the trust architecture built on top of them.