How AI agent payments are reshaping US stablecoin regulation and corporate treasury

Key takeaways

• Global stablecoin transaction volume hit $33 trillion in 2025, up 72% year-over-year,¹ though most of that is trading and internal transfers, not actual end-user payments.

• AI agent payments give autonomous systems their own identity, authorization, and settlement rails, so they can transact without a person clicking approve.

• Airwallex already covers the infrastructure side: multi-currency accounts, yield on idle balances, and programmatic virtual cards. Mastercard, meanwhile, is building the machine-speed rails those agents will transact over.

Commerce is moving from people paying each other to software paying software, and it's happening faster than most financial infrastructure was built to handle. Getting ready means knowing the federal rules now governing stablecoins, and the protocols agents actually use to move money.

What are AI agent payments?

The agentic AI boom has created transaction patterns the existing rails weren't built for. As AI shifts from experimental tooling to core infrastructure, software spend is getting harder to predict and control, driving demand for agentic payments, a category where autonomous systems act as payment entities in their own right, with identity, authorization, and accountability built into every transaction.

Traditional payment rails assume a person is around to approve a charge or clear a two-factor prompt. AI agents don't work that way: they run nonstop, with nobody in the loop. Grand View Research pegs the global AI agents market at $7.6 billion in 2025, growing to $182.9 billion by 2033 at a 49.6% annual clip, and supporting that kind of scale means rebuilding payment infrastructure, not bolting agents onto what's already there.

How a digital payments agent differs from autopay and checkout bots

Finance teams often lump payment agents in with autopay or checkout bots. Autopay triggers a fixed transaction on a schedule, like a monthly SaaS renewal, while checkout bots follow a script under preset conditions. Neither adapts, and both stall on anything unexpected.

A digital payments agent reasons instead. It can read a pricing table, negotiate a custom rate, and adjust its own spending within the limits a human set for it. Automation just follows the script; autonomy makes the call.

Agent payments vs. agentic commerce: which layer each one covers

Agentic commerce and agent payments get conflated constantly. Agentic commerce is the cognitive layer: product comparison, vendor selection, cart building, like a procurement agent that spots overlapping software licenses and recommends consolidating vendors. Agent payments are the execution layer underneath, the infrastructure that actually moves and settles the money, and without it an agent can build the perfect cart but still needs a human to type in a card number.

Keeping these layers separate lets agents shop freely while every dollar still moves through standardized, auditable rails.

The protocols agents use to negotiate and pay each other

Machines transacting with machines need standardized protocols for identity, communication, and value exchange, the same way humans rely on checkout pages and two-factor prompts. Building the financial execution layer for AI agents means solving both halves of the problem: negotiating the deal, then settling it.

Agent-to-Agent (A2A): how agents coordinate before money moves

Before any money moves, agents negotiate. Agent-to-Agent (A2A) protocols let autonomous systems establish trust and agree on terms before a transaction executes, through a three-step mandate flow:

• Intent Mandate: a human or corporate sponsor authorizes what the agent should find and what it's allowed to spend.

• Cart Mandate: once the agent finds what it needs, it locks in a structured cart with exact items and pricing.

• Payment Mandate: at checkout, the agent signs off on the payment, instructing the settlement rail to process it within agreed boundaries.

This keeps downstream rails from processing anything that wasn't pre-negotiated, preventing runaway spending loops.

Where MCP, x402, and commerce protocols fit

Once the handshake is done, execution takes over. The Model Context Protocol (MCP) defines how AI models discover and call external tools, linking agent logic to payment endpoints.

For moving money, developers increasingly reach for x402, an open standard backed by the x402 Foundation, which Coinbase and Cloudflare launched together. An agent requests a protected API, the server replies with an HTTP 402 status and a price, and the agent's wallet signs a stablecoin transfer to complete the exchange in one round trip. It's stablecoin-only, though, and doesn't handle physical goods.

For standard checkout flows, developers use agentic commerce protocols built to connect agents to existing card rails. The Agentic Commerce Protocol, built by OpenAI, uses SharedPaymentTokens for normal checkout forms. Google's AP2, developed with 60-plus partners including Mastercard and PayPal, extends Google's A2A and MCP work to manage trust across both fiat and crypto rails.

These protocols let a business move away from static API keys and pre-paid credit toward payment infrastructure that works on demand.

US regulation for AI agent payments infrastructure

AI agents have no legal personhood: they can't open a bank account, hold property, or sign a contract in their own name. They operate on credentials held by a human sponsor or company, which puts real compliance weight on stablecoin issuers and treasury teams.

The GENIUS Act: what stablecoin issuers must do before agents can spend on-chain

The Guiding and Establishing National Innovation for U.S. Stablecoins Act, known as the GENIUS Act, became law on July 18, 2025² as the first comprehensive federal stablecoin framework. It requires any entity issuing a payment stablecoin in the US to get a federal trust bank charter from the OCC, creating a path for Permitted Payment Stablecoin Issuers (PPSIs) to operate alongside traditional banks. PPSIs have to back every token 1:1 with liquid reserves, think US dollars and short-term Treasuries, and publish a reserve report every month.

Section 4(a)(11) is the part getting the most attention: it bars issuers from paying interest, yield, rewards, or bonuses to holders, directly or indirectly⁵, which keeps stablecoins focused on payments instead of competing with banks for deposits. The OCC backed that up in December 2025, conditionally approving national trust bank charters for five digital asset firms, including Circle and Paxos, and clearing the way for compliant, bank-grade issuance.

All this is happening against a backdrop of serious growth. Total stablecoin supply, a much smaller number than the raw transaction volume above, is on pace to hit $420 billion by the end of 2026, up 56% year-over-year⁴.

AML, sanctions, and Customer Identification Program rules for stablecoin issuers

The GENIUS Act also puts every issuer under anti-money laundering and sanctions rules. On April 8, 2026, FinCEN and the Office of Foreign Assets Control jointly proposed a rule treating PPSIs as financial institutions under the Bank Secrecy Act. Since software can't undergo KYC, issuers run Customer Identification Program (CIP) and due diligence checks on the human or company funding an agent's wallet instead, and need the ability to freeze, seize, or burn tokens on request to stay compliant with OFAC sanctions lists.

How to pay agent endpoints safely under current guardrail requirements

For any company deploying agents, compliance starts with never giving an autonomous system direct access to the main corporate bank account. Guardrails belong at the payment layer instead, built on cryptographic identity checks, scoped spend limits, and automated circuit breakers. Card networks and fintechs are already building for this:

• Mastercard Agent Pay for Machines (AP4M): launched in June 2026 with more than 30 partners including Coinbase and Cloudflare,³ it permissions and settles machine-initiated micropayments, some just fractions of a cent, credentialing agents through “Verifiable Intent” across both cards and stablecoins.

• Visa Intelligent Commerce: Visa's Trusted Agent Protocol, built with Cloudflare, lets consumers delegate Visa credentials to vetted AI shopping agents.

• Cryptographic identity passports: Vouched's Model Context Protocol – Identity (MCP-I) gives agents a verifiable identity, so companies can audit behavior and revoke access if compromised.

The treasury impact of AI agent payments

A 24/7 machine economy doesn't fit how corporate treasury has always worked. Treasury has always run on banking hours, weekend closures, and batch cycles. AI agents don't care about any of that, they fire off high-frequency microtransactions around the clock, and that means cash management, yield, and liquidity planning all need a rethink.

Why idle agent wallets create a float management problem

Agents need pre-funded wallets to transact on demand. Multiply that across hundreds of agents and real money sits idle in corporate float. Because the GENIUS Act bans stablecoin issuers from paying interest directly, that idle cash earns nothing, a drag on margins that grows with every agent deployed.

That's real money left on the table, and it's pushing treasurers to find more flexible ways to manage liquidity.

How corporate treasury teams are adapting: yield, stablecoin issuers, and 24/7 liquidity

Treasury teams are moving to just-in-time funding models that fund agent wallets programmatically, only as needed, freeing cash to sit in yield-bearing instruments until it's deployed. Some park reserves in compliant, yield-bearing stablecoin wraps like sUSDe, converting back to payment stablecoins only when they need the liquidity.

Because machine payments never stop, treasury teams are also shifting to digital-first accounts with real-time FX and immediate local bank transfers, since legacy clearing windows don't work for a system running all night.

How Airwallex supports compliant AI agent payments infrastructure

Navigating agent-economy treasury and compliance takes a financial partner built for both. Airwallex operates its own payment network across dozens of markets, reducing reliance on the correspondent banking chain that typically slows cross-border payments.

Multi-currency accounts and yield for globally distributed agent spend

AI agents don't respect borders; they call APIs and buy resources from vendors across dozens of markets. McKinsey estimates B2B stablecoin payments alone reached $226 billion in the past year, up 733% year-over-year⁶. To avoid the international card fees and forced FX conversions of legacy rails, businesses can use Airwallex Global Treasury to open local currency accounts across 70+ countries and hold balances in 20+ currencies natively, so agents collect and pay locally without conversion eating into the balance.

To address the float problem directly, Airwallex Yield lets businesses earn competitive returns on USD balances with no lock-up period, turning idle stablecoin float into working capital instead of dead weight.

API-ready virtual cards for programmatic issuance

Airwallex Issuing lets finance teams generate virtual cards programmatically through a developer API and assign them to specific agents, backed by real-time controls:

• Programmatic spending caps: set daily, weekly, or lifetime limits per card so a runaway agent loop can't rack up charges nobody approved.

• Merchant category restrictions: lock a card to specific categories, cloud providers, API gateways, whatever fits the use case.

• Real-time API controls: if an agent fails a compliance check or starts acting oddly, freeze or kill the card instantly through the API.

Airwallex Spend Management syncs receipts two-way with NetSuite, QuickBooks, and Xero too, so machine expenses reconcile themselves instead of landing on someone's desk. From there, businesses can open an Airwallex Business Account to issue virtual cards programmatically, manage multi-currency balances, and set real-time spend guardrails.

Frequently asked questions about AI agent payments

How does a digital payments agent transact money autonomously?

It connects a programmatic wallet to payment protocols through APIs. When it hits a paywall, it reads the terms, signs the transaction cryptographically, and executes the transfer without a human in the loop.

What does the GENIUS Act require of stablecoin issuers used for agent payments?

Issuers must back tokens 1:1 with liquid reserves, disclose composition monthly, register under the Bank Secrecy Act, run AML and sanctions compliance programs, and maintain the ability to freeze or seize tokens. Section 4(a)(11) bars paying direct interest to holders.

How do you set compliant spending controls for agent payments?

Layer them. Verify each agent's identity cryptographically, issue merchant-restricted virtual cards with hard spending caps, and run real-time monitoring with circuit breakers that freeze credentials the moment something looks off.

Why do idle AI agent wallets create a treasury management problem?

Idle wallets require pre-funding, trapping cash in non-interest-bearing accounts. Since the GENIUS Act bans stablecoin interest, treasurers must fund wallets on demand or park reserves in yield-bearing products in the meantime.

Who is held legally responsible if an AI agent makes an unauthorized purchase?

AI agents have no legal personhood, so they can't sign contracts or hold liability themselves. That responsibility falls on whoever deployed and funded the wallet, the human or the company, which is exactly why spend controls and real-time auditing aren't optional.

What is the Agent-to-Agent (A2A) protocol and how does it relate to payments?

A2A is the open standard for how autonomous systems find each other, verify who they're dealing with, and negotiate terms before any money moves. A buying agent hands a selling agent signed mandates, then the payment layer, x402 or a card network, takes over to execute and settle.

Sources

1. https://nevermined.ai/blog/stablecoin-payments-ai-agents-statistics

2. https://www.whitehouse.gov/fact-sheets/2025/07/fact-sheet-president-donald-j-trump-signs-genius-act-into-law/

3. https://www.mastercard.com/us/en/news-and-trends/press/2026/june/mastercard-launches-agent-pay-for-machines.html

4. https://www.spark.money/research/stablecoin-supply-420-billion-growth

5. https://www.federalregister.gov/documents/2025/09/19/2025-18226/genius-act-implementation

6. https://www.mckinsey.com/industries/financial-services/our-insights/stablecoins-in-payments-what-the-raw-transaction-numbers-miss