Inside AMEX’s agentic commerce stack: How intent contracts and single-use tokens enforce AI transactions

Imagine a world where your personal AI assistant doesn’t just remind you to buy groceries—it actually goes ahead, finds the best organic avocados at the lowest price, negotiates a discount, and pays for them—all without you lifting a finger. This isn’t science fiction. It’s the emerging reality of agentic commerce, a new frontier where artificial intelligence agents autonomously conduct transactions on behalf of users. At the forefront of this revolution is American Express (Amex), quietly building a sophisticated infrastructure designed to make AI-driven shopping both seamless and secure. But beneath the sleek interface lies a complex web of trust, control, and validation—one that Amex is tackling with a unique blend of innovation and caution.

Unlike traditional e-commerce, where humans make decisions and initiate payments, agentic commerce shifts control to AI systems. These digital agents are programmed with user preferences, budgets, and goals, allowing them to act independently. While the promise is immense—personalized shopping, time savings, and hyper-efficient transactions—the risks are equally profound. Who ensures the AI doesn’t go rogue? How do merchants know they’ll get paid? And how can users trust that their financial data isn’t being misused?

Amex’s answer lies in its Agentic Commerce Experiences (ACE) developer kit, a proprietary system designed to bring structure, security, and accountability to AI-driven transactions. But as with any emerging technology, the devil is in the details—especially when those details remain partially hidden.

The Rise of Agentic Commerce: A New Era of Autonomous Transactions

Agentic commerce represents a fundamental shift in how we interact with digital marketplaces. Instead of manually browsing, comparing prices, and checking out, users delegate these tasks to AI agents trained on their preferences, spending habits, and ethical boundaries. These agents can operate across platforms, negotiate deals, and even manage subscriptions—all in real time.

This concept isn’t entirely new. Early forms of automation, like browser extensions that auto-fill forms or apply coupon codes, laid the groundwork. But today’s AI agents go far beyond simple automation. They use machine learning to interpret intent, adapt to changing conditions, and make decisions in dynamic environments. For example, an AI might detect that your favorite brand of coffee is running low, compare prices across multiple retailers, apply available discounts, and complete the purchase—all while staying within your monthly grocery budget.

However, this autonomy introduces significant challenges. Without proper safeguards, AI agents could make impulsive purchases, fall victim to phishing scams, or even be manipulated by malicious actors. This is where Amex’s ACE system steps in—not as a standalone solution, but as a critical layer in the evolving agentic commerce stack.

Amex’s Unique Position: Issuer and Network in One

What sets American Express apart in the financial ecosystem is its dual role: it is both a card issuer and a payment network. Unlike Visa or Mastercard, which operate solely as networks and rely on banks to issue cards, Amex controls the entire transaction flow from end to end. This closed-loop architecture gives it a rare advantage in the world of agentic commerce.

In traditional payment systems, a transaction involves multiple intermediaries: the cardholder, the issuing bank, the payment network, the acquiring bank, and the merchant. Each step introduces latency, complexity, and potential points of failure. Amex’s integrated model streamlines this process, enabling faster validation and greater control over how transactions are processed.

This control is especially valuable when AI agents are involved. Because Amex can validate transactions directly within its network, it can implement real-time checks that ensure agents are acting within predefined boundaries. For instance, if an AI attempts to purchase a luxury item that exceeds the user’s typical spending pattern, the system can flag it for review or block it outright.

Moreover, Amex’s participation in broader initiatives like Google’s Agent Pay Protocol (AP2) highlights its commitment to interoperability. While AP2 focuses on creating common standards for AI-to-AI payments across platforms, Amex’s ACE kit goes a step further by embedding transaction control directly into the payment layer. This means that even if an AI agent operates across multiple ecosystems, Amex can enforce rules and validate actions at the moment of payment.

The Black Box Problem: Trust and Transparency in AI Transactions

Despite its technical sophistication, Amex’s ACE system has drawn criticism for its lack of transparency. While the company emphasizes that it performs validation at the payment layer, it remains vague about how that validation actually works. This opacity creates a “black box” effect—where users and developers know the system works, but can’t see inside to understand why certain decisions are made.

This is a recurring challenge in AI-driven systems. Machine learning models, especially deep neural networks, often operate as black boxes, making it difficult to trace how inputs lead to outputs. In finance, where accountability is paramount, this is a major concern. If an AI agent makes a fraudulent purchase or violates a user’s spending limits, who is responsible? The user? The AI developer? The payment provider?

Amex claims its system uses a combination of deterministic checks—such as verifying card validity and transaction limits—and semantic evaluation, which interprets the intent behind an agent’s actions. For example, if an AI books a flight to Paris, the system might assess whether this aligns with the user’s travel history, calendar events, or stated preferences. But because this process is abstracted, external parties can’t audit or verify its fairness.

Luke Gebb, Amex’s EVP and global head of innovation, argues that trust and security are the missing pieces in agentic commerce. “Some of what is missing so far is the perspective of a company like ours,” he told VentureBeat. “We feel that trust and security are critical to advancing this space. This is really the first time that an issuer is coming to the table.”

Intent Contracts and Single-Use Tokens: The Mechanics of Control

At the heart of Amex’s ACE system are two key innovations: intent contracts and single-use tokens. These mechanisms work together to enforce boundaries and ensure that AI agents act in accordance with user intentions.

An intent contract is a digital agreement that defines the scope of an AI agent’s authority. It specifies what the agent can buy, how much it can spend, which merchants it can transact with, and under what conditions. For example, a user might grant an AI agent permission to purchase office supplies up to $200 per month, but only from approved vendors. The contract is cryptographically signed and embedded in the transaction flow, making it tamper-proof.

When the agent attempts a purchase, the system checks the intent contract in real time. If the transaction aligns with the agreed terms, it proceeds. If not, it’s blocked. This provides a layer of programmable control that goes beyond traditional spending limits.

Complementing this are single-use tokens, which act as temporary, limited-access credentials for each transaction. Instead of using a static card number, the AI agent receives a unique token that is valid for only one purchase. This reduces the risk of token reuse, replay attacks, and unauthorized access. Even if a token is intercepted, it becomes useless after the transaction is complete.

Together, intent contracts and single-use tokens create a robust framework for secure delegation. Users retain ultimate control, while AI agents gain just enough autonomy to be useful. It’s a delicate balance—one that Amex is refining through real-world testing and developer feedback.

The Road Ahead: Challenges and Opportunities

While Amex’s ACE system represents a significant step forward, it’s not a panacea. The broader agentic commerce ecosystem still faces hurdles, including regulatory uncertainty, interoperability gaps, and public skepticism.

Regulators are beginning to take notice. In the U.S., the Consumer Financial Protection Bureau (CFPB) has signaled interest in AI-driven financial services, particularly around transparency and consumer protection. The European Union’s AI Act also imposes strict requirements on high-risk AI systems, which could include autonomous shopping agents.

Interoperability remains another challenge. While Amex is working with Google’s AP2 and other protocols, there’s no universal standard yet. Different platforms may use conflicting validation methods, making it hard for AI agents to operate seamlessly across networks.

Still, the potential benefits are too great to ignore. Imagine a world where elderly users with mobility issues can rely on AI agents to manage their shopping, or where small businesses use autonomous systems to optimize supply chains in real time. Agentic commerce could democratize access to commerce, reduce friction, and unlock new efficiencies.

Amex’s approach—centered on control, validation, and trust—may well become the gold standard. But for agentic commerce to reach its full potential, the industry must move beyond black boxes and toward auditable, explainable systems that users can understand and trust.

As Luke Gebb puts it, “This isn’t just about making transactions faster. It’s about building a new foundation for digital commerce—one where humans and AI can collaborate safely and effectively.” The journey is just beginning.

This article was curated from Inside AMEX’s agentic commerce stack: How intent contracts and single-use tokens enforce AI transactions via VentureBeat