Agentic Commerce: Mastercard’s AI-Driven Tokens and Real-Time Fraud Defense

Mastercard’s latest push into artificial intelligence marks a clear inflection point for payments: the company has moved from embedding AI into detection systems to enabling agentic commerce—where AI agents can negotiate, authorize and execute purchases on behalf of users—backed by new developer tools, token innovations and expanded fraud-fighting capabilities.

Background​

Mastercard’s announcements in 2025 formalize a strategy that pairs two long-running themes in payments: tokenization and real-time fraud decisioning. The company has layered a new concept—Agent Pay—on top of tokenized credentials, making tokens usable and controllable by registered AI agents. At the same time, Mastercard has amplified its fraud-detection stack with generative-AI-enhanced models (branded as an evolution of Decision Intelligence) and bolstered threat-intelligence capabilities through recent acquisitions and partnerships.
This is not merely a convenience play. It is a structural effort to keep issuers, merchants and consumers secure as commerce shifts to conversational and autonomous interfaces. The combination of agent-aware tokens, on-device authentication, verifiable credentials and network-level monitoring is designed to let AI agents act for users without exposing raw card data or reducing issuer visibility.

---

What Mastercard announced — the essentials​

Agent Pay and agentic tokens​

• Agent Pay is Mastercard’s program to enable AI agents (from chatbots to orchestration engines) to transact on behalf of users.
• Agentic Tokens are tokenized payment credentials bound to registered agents and governed by rules and permissions that limit what an agent can do.
• The model requires agent registration and verification, and it embeds mechanisms for user consent and intent verification before a payment completes.
• Mastercard has launched developer tooling—an Agent Toolkit, Agent Sign-Up and Insight Tokens—to help AI platforms, merchants and banks integrate securely.

Expanded fraud and security capabilities​

• Mastercard has rolled out an AI-enhanced fraud stack that builds on its Decision Intelligence platform; the upgraded offering uses generative models and advanced transformer architectures to analyze transactions in real time.
• Separately, Mastercard completed the acquisition of a major threat-intelligence firm to augment its ability to detect and respond to cyber risks and leaking credentials, folding that intelligence into fraud scoring and issuer services.
• The company emphasizes responsible, human-centered AI and proposes standards and verifiable credentials to create traceability and consumer control for agentic transactions.

Partnerships and rollout scope​

• The program is explicitly designed to interoperate with major cloud and AI platforms, with announced integrations and collaboration frameworks to support scaling.
• Mastercard signaled a U.S.-first enablement window, expanding to cardholders broadly before a wider global rollout slated to follow.

---

Why this matters: tokenization meets autonomy​

Tokenization has been a backbone of digital payments for years: substituting card numbers with cryptographic tokens drastically reduces card-present and card-not-present exposure. Mastercard’s new approach extends tokenization into the agentic world by adding two critical elements:

• Agent-binding: Tokens are no longer just device- or merchant-scoped credentials; they can be scoped to a verified AI agent. That means tokens carry metadata about agent identity, permissions and intent boundaries.
• Policy and consent enforcement: Tokens and the surrounding protocols carry rules that enable immediate, automated enforcement of consumer-defined constraints (spending limits, merchant white-lists, one-time approvals).

The result is a practical path for AI agents to act with real-world purchasing power while giving issuers and consumers the hooks they need for governance and dispute resolution.

---

Fraud detection: generative AI and real-time decisioning​

Mastercard’s fraud stack evolution is a central pillar of this strategy. The move to generative and transformer-based models has three material effects:

• Richer contextual signals: Transformer models can ingest wider context—conversational metadata, agent behavior patterns, device telemetry—and correlate signals across time to better distinguish anomalous activity.
• Faster decisioning at scale: Real-time scoring with these models enables near-instant decisions, critical when agents are negotiating prices, completing multi-step purchases or making split-second authorization calls.
• Lower false positives: Early deployments show material reductions in falsely declined transactions, improving customer experience while preserving detection efficacy.

These improvements are complemented by the integration of threat intelligence and behavioral biometrics, creating a layered defense that ties network scoring to external indicators of compromise.

---

Strategic strengths — what Mastercard brings to the table​

• Network scale and trust architecture
• Mastercard operates at very large transaction volumes, giving it access to cross-market signals that are uniquely useful for fraud detection and model training. Network-scale telemetry accelerates threat detection and helps build robust, generalizable models.
• Established tokenization expertise
• Mastercard’s token infrastructure already powers mobile wallets, card-on-file credentials and programmable payments; extending that to agentic tokens is evolutionary, not revolutionary, which shortens integration and safety testing cycles.
• Ecosystem partnerships
• The approach intentionally relies on collaborations with cloud and AI platform providers, payment-enablers and checkout players—allowing rapid scaling across developer ecosystems rather than attempting to own every layer.
• Security posture and acquisitions
• Bringing threat-intelligence capabilities into the fold strengthens Mastercard’s ability to detect compromised credentials or coordinated attacks, improving both prevention and rapid response.
• Standards-first framing
• By building toward verifiable credentials and participating with standards bodies, the company aims to create interoperable, auditable frameworks for agentic commerce that many stakeholders (issuers, merchants, platforms) can implement.

---

Risks and open questions​

While the technical and strategic direction is sound, several real and emergent risks deserve scrutiny.

1. Privacy and data governance complexity​

Agentic systems will require permissioned access to personal context (shopping preferences, subscription details, budgets). Effective privacy controls, consent revocation flows and robust data minimization are essential. Missteps could trigger regulatory scrutiny and consumer backlash.

2. New attack surfaces​

Agentic tokens and agent registration introduce novel vectors:

• Compromised agents (or rogue agent instances) could attempt automated abuse at scale.
• Supply-chain attacks on AI platforms or third-party integrators could lead to credential leakage.
• Automated negotiation agents may be vulnerable to adversarial manipulation (price inflations, offer spoofing) if marketplace integrity controls are weak.

3. Algorithmic bias and fairness​

Generative and transformer models can bake in historical biases from training data. If decisioning models inadvertently discriminate (e.g., higher friction for certain geographies or demographics), that raises both ethical and regulatory risks. Proactive bias testing and human oversight are required.

4. Operational dependency and vendor concentration​

Heavy reliance on major cloud and AI providers could produce concentration risk. If an AI platform outage or a change in policy blocks agentic access, user experiences could degrade suddenly.

5. Legal and regulatory uncertainty​

Agentic commerce touches identity, consumer protection, payments regulation and data privacy. Jurisdictions will vary widely on what constitutes adequate authentication and what consumer consent entails in autonomous transactions. Regulators may demand more stringent logging, explainability and liability assignments.

6. Human factors and trust​

Consumers must understand when an agent acts on their behalf and how to revoke or audit authority. Poor UX or opaque behaviors will erode trust, regardless of technical safeguards.

---

Technical architecture — what’s under the hood​

Core building blocks​

• Agentic Tokens: Cryptographic credentials that include agent identity, permitted merchant domains, spend limits and transaction scopes.
• Agent Registry/Sign-Up: A verification layer that registers agents, issues identity assertions and enforces governance.
• Model Context Protocol (MCP) & Agent Toolkit: Machine-readable API docs and protocols to let agents discover capabilities and interact with payment APIs programmatically.
• Insight Tokens: Permissioned tokens that let agents access contextual insights (e.g., loyalty status, merchant offers) with consumer consent.
• Verifiable Credentials: Standards-based assertions for intent, amount and merchant to provide traceable, auditable evidence of authorization.
• Real-time Decisioning: Transformer-based scoring engines that evaluate transactions in milliseconds, using behavioral, contextual and external threat signals.
• On-device authentication: Integration with passkeys, biometrics and strong device-bound authentication to verify human approval when required.

Typical transaction flow (simplified)​

• An AI agent proposes an intent (e.g., “book flight X”).
• The user grants permission within pre-configured rules (one-time approval, spending limit).
• The agent requests an Agentic Token scoped to that transaction from the issuer’s provisioning service.
• Mastercard’s network validates the token, applies real-time decisioning and risk scoring, referencing threat intelligence.
• If authentication is required, the flow triggers an on-device passkey or biometric prompt.
• Transaction executes; telemetry and verifiable credentials are logged for auditing and dispute resolution.

---

Practical implications for banks, merchants and platforms​

For issuers (banks and credit unions)​

• Need to adapt issuer-side token management and consent platforms to support agent-specific scoping and revocation.
• Must integrate AI-based decisioning outputs into underwriting and fraud workflows, balancing automated approvals with human review thresholds.
• Should prepare for increased API traffic and new telemetry requirements; capacity planning and SLA considerations are essential.

For merchants and checkout providers​

• Merchants must recognize agentic transactions and surface clear receipts and intent metadata to consumers.
• Checkout platforms will need to accept agentic tokens and produce verifiable credentials to confirm the relationship between agent, user intent and the merchant.
• There is an opportunity to leverage agent-provided contextual signals to personalize offers, but doing so requires stringent privacy compliance.

For AI platforms and agent developers​

• Developers must implement clear opt-in, consent revocation, and explainability features for end users.
• Agents should be designed to support granular spending rules, multi-stage approvals and friction points to ensure user control.
• Security best practices—ephemeral credentials, secure key storage and tamper-resistant logging—are mandatory.

---

Plausible fraud and abuse scenarios — and mitigations​

• Scenario: A compromised agent makes repeated, high-volume micro-purchases across merchants.
• Mitigations: Rate limiting on agent tokens, anomaly detection for burst patterns, immediate token revocation and issuer throttling.
• Scenario: A malicious marketplace fakes offers to induce agents to purchase at inflated prices.
• Mitigations: Merchant reputation scoring, verifiable credential checks for offer authenticity, agent-side validation of merchant credentials.
• Scenario: Adversarial queries manipulate agent behavior to bypass consent (prompt injection).
• Mitigations: Prompt safety filters, intent verification steps, human-in-the-loop fallback for high-risk actions.
• Scenario: Third-party integrator leaks agent credentials.
• Mitigations: Zero-trust architecture, short-lived tokens, mandatory encryption-at-rest, continuous supply-chain monitoring and threat intelligence correlation.

---

Governance and policy recommendations​

• Agent identity and provenance
• Maintain a tamper-evident registry with periodic re-verification and reputation scoring for agents.
• Transparent consent UIs
• Provide clear, machine- and human-readable consent receipts that list scope, duration and revocation paths.
• Explainability & logging
• Mandate auditable logs that record agent intent, decisioning rationale and the signals used for each authorization.
• Bias testing and independent audits
• Regular third-party audits of decisioning models and datasets to find and remediate disparate impact.
• Interoperable standards
• Invest in verifiable credential standards and cross-industry protocols to reduce fragmentation and increase auditability.
• Regulatory engagement
• Work proactively with regulators on classification of agentic transactions, liability assignment and consumer remediation frameworks.

---

Business opportunities and revenue models​

• Programmable, contextual payments: Merchants can monetize agent-driven upsells, subscriptions and tailored bundles.
• Value-added security services: Issuers and network operators can charge for advanced AI decisioning, risk intel feeds and dispute assistance.
• Developer ecosystems: Tooling like Model Context Protocol access, Agent Toolkit subscriptions and consulting services can become new revenue streams.
• Cross-sell with threat intelligence: Integrated offerings that combine real-time fraud scoring, threat feeds and identity services are strong enterprise propositions.

---

What to watch next​

• Adoption metrics: Track how quickly issuers, major merchants and AI platforms adopt agent-aware tokens and whether consumer opt-in rates meet expectations.
• Interoperability standards: The emergence of widely accepted verifiable credential schemas and a consensus on the Model Context Protocol will be pivotal for global scale.
• Regulatory responses: Expect proposals around consent, liability and auditability—particularly in jurisdictions with strict privacy laws.
• Security incidents: Any high-profile compromise involving an agent or agent token will be a major test of the architecture and the remediation playbooks.
• User experience signals: Consumer understanding and the UX for granting, reviewing and revoking agent permissions will determine the practical viability of agentic commerce.

---

Conclusion​

Mastercard’s pivot toward agentic commerce and the simultaneous hardening of fraud defenses illustrate how payments networks are preparing for AI-driven client experiences. By fusing tokenization, real-time transformer-based decisioning, developer tooling and threat intelligence, the company is building an end-to-end framework that could allow AI agents to transact reliably on behalf of people and businesses.
The upside is substantial: convenience, highly personalized commerce, and materially improved fraud detection at scale. The downside is equally significant: new attack surfaces, complex privacy trade-offs, algorithmic fairness issues and regulatory ambiguity.
The next 12 months will be decisive. If agent registration, consent controls, verifiable credentials, and robust threat intelligence work together as designed, agentic commerce could be a secure extension of today’s digital payments. If any of those pillars lag—especially governance, explainability and cross-industry standards—the landscape risks fragmentation and fresh avenues for fraud and consumer harm.
For banks, merchants and developers, the prudent path is clear: adopt early, prioritize governance and privacy by design, and insist on auditable, standards-based implementations. Agentic commerce is coming; the outcome will depend on whether the ecosystem treats trust and control as first-class design requirements rather than afterthoughts.

---

Source: WebProNews Mastercard Advances AI in Payments for Fraud Detection and Security