AI agents are now authorized to spend money on their own, but the legal framework for handling mistakes doesn't exist yet. Three major payment networks have launched agentic commerce infrastructure in the past 16 months, enabling autonomous agents to authenticate, spend stablecoins, and complete transactions. However, existing payment law assumes a simple binary: either a person authorized a transaction or they didn't. When an agent operates exactly as programmed but still makes a bad purchase, current dispute rules have nowhere to route the complaint.

What Is Agentic Commerce and How Does It Actually Work?

Agentic commerce refers to autonomous AI agents making purchases on behalf of businesses or users. The infrastructure crystallizing around this capability includes four distinct layers: a payment protocol (x402), a settlement asset (USDC stablecoin), wallet and identity controls, and a settlement blockchain. Agents authenticate using cryptographic signatures, receive spending allowances and merchant whitelists from their operators, and execute transactions without human intervention for each purchase.

The demand for this capability is substantial. Autonomous agents consume approximately five times more tokens than human users, with over 85 percent of that consumption coming from cached prompts, according to OpenRouter data circulated in late August. McKinsey projects agentic commerce will reach a market value of $3 trillion to $5 trillion by 2030, and Deloitte forecasts that 25 percent of global e-commerce sales will be enabled by AI agents by 2030.

How Are Payment Networks Building Agent Wallets and Controls?

Visa, Mastercard, and Cloudflare have each launched agent payment infrastructure with built-in guardrails. Cloudflare Wallets, launched August 4, 2026, uses a two-tier design with separate Account Wallets for humans and Virtual Wallets for agents, enforcing allowance limits, merchant whitelists, maximum transaction sizes, and human review prompts when anomalous spending is detected. Coinbase launched similar agent wallet tools in June 2026.

Visa's rollout began with Intelligent Commerce on April 30, 2025, allowing agents to browse and buy within consumer-set limits. By December 2025, Visa had completed hundreds of agent-initiated transactions across its partner ecosystem. In June 2026, Visa integrated tokenized credentials directly into ChatGPT with real-time authorization and fraud monitoring.

• Mastercard Agent Pay (April 29, 2025): Agentic tokens built on Mastercard's existing tokenization system, with agents registered and verified through One Credential and on-device biometrics before they can act.
• Visa Intelligent Commerce (April 30, 2025): Tokenized credentials letting agents browse, select, and buy on a consumer's behalf, inside limits the consumer sets.
• Visa Trusted Agent Protocol (October 14, 2025): Cryptographic signatures letting a merchant confirm which agent it is dealing with and block impersonators.
• Cloudflare Wallets (August 4, 2026): Stablecoin-funded wallets for agents built on Cloudflare's x402 payment standard with spend caps and allow-lists.

Why Does the Liability Gap Matter?

The critical problem emerges when an agent operates perfectly within its mandate but still makes an unwanted purchase. Consider a real scenario: an agent with a spending allowance, approved merchant list, and maximum transaction size authenticates correctly and renews a data subscription the business had already cancelled. The agent did exactly what it was authorized to do, yet the result is a purchase nobody wanted.

Current payment law cannot address this situation. Regulation E, which governs debit and bank transfers, defines unauthorized transfers as those "initiated by a person other than the consumer without actual authority." An agent furnished an access device by its operator was given authority by construction, so the liability framework never activates. Credit card rules follow the same pattern: unauthorized use requires someone acting "without actual, implied, or apparent authority." An agent operating inside its mandate always has authority, regardless of the outcome.

"Where it gets murkier is everything short of outright fraud. Once an agent authenticates properly, fraud liability follows existing rules, but for an agent that simply misjudged an instruction, no liability shift exists yet," stated Reshmi Suresh, head of agentic commerce at Worldpay.

Reshmi Suresh, Head of Agentic Commerce at Worldpay

The billing-error categories under Regulation Z (the federal credit card rule) map to merchant failures or genuine authority failures. They cover unauthorized charges, unidentified charges, items not accepted or not delivered as agreed, payment posting errors, and computational errors. None of these categories fit an authorized agent's bad purchase decision.

What Settlement Infrastructure Is Actually Being Used?

USDC stablecoin has emerged as the dominant settlement asset for agent payments. Circle's Q2 2026 financial results, published August 5, 2026, reported that 99.3 percent of x402 agent-payment volume settled in USDC. This dominance reflects USDC's structural advantages: gasless EIP-3009 authorization eliminates the need for agents to hold volatile ETH or SOL for transaction fees, deep liquidity from $73.3 billion in circulation, and regulatory clarity under the emerging GENIUS Act.

Circle's Arc L1 blockchain is positioned to become the institutional settlement layer. Arc launches public mainnet on September 16, 2026, with a founding validator cohort of 11 institutions including BlackRock, DTCC, Galaxy, Global Payments, ICE, Mastercard, MoneyGram, and SBI Group. The x402 protocol itself uses HTTP's 402 "Payment Required" status code in a 12-step challenge-response flow where an agent signs a gasless USDC authorization, retries with a payment header, and a facilitator verifies and settles the transaction on-chain.

Circle reported USDC in circulation at $73.3 billion, up 19 percent year-over-year, with onchain volume of $14.8 trillion, up 151 percent year-over-year. The Circle Payment Network reached $14.7 billion in annualized volume with 175 financial institutions participating, up 29 percent quarter-over-quarter.

How Can Businesses Implement Agent Payment Controls Today?

While the liability framework remains undefined, operators can implement multiple layers of control to minimize risk. The emerging best practice involves separating authorization from execution, using allowance limits, maintaining merchant whitelists, and implementing human review triggers for anomalous activity.

• Spending Allowances: Set maximum transaction sizes and daily or monthly spending caps for each agent, enforced at the wallet level before any transaction is submitted.
• Merchant Whitelists: Restrict agents to pre-approved vendors and service providers, preventing purchases from unknown or untrusted sources.
• Anomaly Detection: Configure human review prompts to trigger when spending patterns deviate from normal behavior, such as unusually fast transaction velocity or purchases outside typical categories.
• Gasless Authorization Separation: Use EIP-3009 gasless signatures so agents don't need to hold native tokens, reducing the surface area for accidental overspending on transaction fees.

The x402 SDK ecosystem is primarily built in TypeScript (50.5 percent), Python (29.3 percent), and Go (19.1 percent), making it accessible to most development teams. Cloudflare's Monetization Gateway sells content and APIs headlessly using x402 micropayments, allowing agents to discover and purchase tools through integrations.

What Happens Next in Agentic Commerce Regulation?

The silence from payment networks and regulators on agent liability is inherited from payment law written for a different problem. Regulation E assumes binary authorization: a transaction was either authorized or it wasn't. An agent operating inside its mandate always answers yes to the authorization question, whatever it actually bought.

Visa's Jack Forestell, chief product and strategy officer, framed the initial launch plainly: "Now, with Visa Intelligent Commerce, AI agents can find, shop and buy for consumers based on their pre-selected preferences." By December, Visa's Rubail Birwadker was already describing the shift as settled: "This holiday season marks the end of an era. In 2026, AI agents won't just assist your shopping, they will complete your purchases".

By December

The infrastructure is moving faster than the legal framework. Three of the world's largest payment networks are competing to own the identity and authentication layer for agentic commerce, but none has published a liability framework for the transaction that clears every check and is still wrong. As agent-initiated transactions scale from hundreds to millions, the gap between technical capability and legal clarity will become impossible to ignore.