Research/Fintech

AI Agents Can't Open Bank Accounts: The Regulatory Gap Blocking Autonomous Payments

No US, EU, or UK regulatory framework lets an AI agent hold its own payment account. How crypto wallets are filling the gap and what regulators think.

bcTanjiNov 22, 2026

AI agents can write code, book flights, and negotiate contracts. But they cannot open a bank account. No regulatory framework in the United States, European Union, or United Kingdom permits software to hold its own payment account, initiate card transactions in its own name, or pass the identity checks required to onboard as a financial customer. This is the regulatory gap blocking autonomous payments: a growing class of economic activity that existing financial law was never designed to accommodate.

The gap is not theoretical. In October 2026, JPMorgan Chase, Citigroup, Bank of America, and Charles Schwab all moved to restrict external AI agents from accessing customer accounts or moving money. Jamie Dimon described the threat of “rogue agents” on JPMorgan's October 13 earnings call. Meanwhile, autonomous agents using crypto wallets processed 6.4 million stablecoin payments in a single 35-day window, routing around traditional rails entirely.

Why AI Agents Cannot Use Traditional Payment Rails

The problem is structural, not technical. Financial regulations across every major jurisdiction define account holders, cardholders, and customers as either natural persons or legal entities. Software is neither. Three layers of rules create the barrier.

Identity verification requirements

Under the Bank Secrecy Act's Customer Identification Program (31 CFR 1020.220), US banks must collect a name, date of birth, address, and government-issued identification number before opening an account. These are properties of human beings and registered legal entities. Software cannot provide a date of birth. It has no passport. The KYC/AML framework assumes a verifiable identity behind every account, and AI agents do not have one.

The same pattern holds internationally. The EU's Anti-Money Laundering Directives require customer due diligence tied to natural or legal persons. The UK's Financial Conduct Authority enforces similar identity requirements under its money laundering regulations. No jurisdiction has carved out an exception for autonomous software.

Consumer protection definitions

US Regulation Z (Truth in Lending) defines a consumer as “a cardholder or natural person to whom consumer credit is offered or extended.” Regulation E (Electronic Fund Transfers) governs error resolution and unauthorized transfer liability using the same human-centric framing. When an AI agent initiates a payment, the question of whether the consumer “authorized” the transaction under Reg E has no clean answer. Did the user provide an “access device” to the agent? Does setting a spending policy constitute ongoing consent?

Card network rules

Visa and Mastercard's dispute resolution frameworks assume a human cardholder initiated the transaction. Chargebacks, fraud scoring, and liability assignment all flow from this assumption. When an AI agent buys the wrong product or overpays for a service, the existing chargeback process has no clear mechanism for resolving whether the agent acted within its authority. Industry analyst Chargebacks911 has warned that the infrastructure for managing disputes, assigning liability, and resolving contested charges in agent-initiated transactions remains largely unaddressed.

The core paradox: Financial law requires a “person” behind every payment. AI agents are not persons. Wrapping an agent in a human's credentials creates liability ambiguity. Giving the agent its own credentials requires a legal identity that does not exist.

How Major Banks Have Responded

Rather than adapting their systems, the largest US banks have chosen to restrict agent access. The wave of restrictions in October 2026 revealed how unprepared traditional finance is for autonomous economic actors.

InstitutionPolicyWhat remains allowed
JPMorgan ChaseBlocks external AI agents from acting on behalf of retail customersRead-only balance checks via Plaid
CitigroupBars third-party agents from logging in or transacting on digital platformsLimited read-only access through open-banking providers
Bank of AmericaRestricted agent access to consumer deposit accountsInternal AI tools for customer service
Charles SchwabBlocks outside agents attempting login with client credentialsProprietary research and analytics tools

The banks' concern is not purely regulatory. They worry that autonomous agents could shift idle deposits to higher-yield accounts at competitor institutions, optimizing away the low-cost funding that underpins bank profitability. The regulatory gap gives banks cover to protect a business model, not just consumers.

The Crypto Workaround: Self-Custodial Agent Wallets

While traditional finance debates whether agents should have accounts, crypto infrastructure has quietly built a working alternative. Self-custodial wallets do not require bank accounts, identity documents, or card network membership. A software agent can generate a cryptographic key pair, receive stablecoins, and spend them according to programmatic policies: all without passing KYC or opening an account with any financial institution.

The x402 protocol

The most significant implementation is x402, an open protocol that repurposes the long-dormant HTTP 402 “Payment Required” status code. When an AI agent requests a paid resource from an x402-enabled server, the server responds with a 402 status and a payment specification. The agent constructs a stablecoin payment, includes it in the HTTP headers, and retries the request. The server verifies the payment on-chain and delivers the resource.

This flow eliminates the need for accounts, credentials, or identity verification. The payment is atomic: either the stablecoin transfer confirms and the resource is delivered, or nothing happens. No chargebacks, no dispute resolution, no identity infrastructure. By March 2026, Coinbase reported that x402 had processed over 50 million transactions. A Yahoo Tech analysis found that more than 97% of on-chain agent transactions in Q2 2026 used the protocol.

Coinbase Agentic Wallets

Launched in February 2026 as part of Coinbase's Developer Platform, Agentic Wallets give AI agents their own on-chain accounts with programmable spending controls. The wallet owner (a human or business) sets funding limits, per-transaction caps, and session-level budgets. The agent operates autonomously within those constraints, signing transactions with its own key material.

The design mirrors how businesses already handle corporate credit cards: an authorized entity sets spending policies, and an agent operates within them. The difference is that the “card” is a crypto wallet, the “network” is a blockchain, and the “authorization” is enforced by code rather than card network rules. Wallets support gasless transactions on Base, Coinbase's Ethereum Layer 2, bringing per-transaction costs below one cent.

Why crypto works here: Permissionless blockchains do not require identity verification to hold or transfer value. This is precisely the property that makes them useful for AI agents: the same feature that regulators have historically viewed as a risk becomes an enabler for a new class of economic actor.

Google AP2 and agent-to-agent payments

Google's Agent Payments Protocol (AP2), announced in September 2025 as an extension of the Agent2Agent (A2A) coordination protocol, adds a payment layer to multi-agent workflows. AP2 uses signed digital “mandates” as proof of user authorization and supports multiple payment methods including credit cards, stablecoins, and real-time bank transfers. Coinbase's x402 functions as an AP2-compatible payment backend, allowing agents to settle with stablecoins when card rails are unavailable or impractical.

The Liability Question Nobody Has Answered

Even where agents can technically make payments, the question of who bears liability for errors remains unresolved. The scenarios are not hypothetical.

  • An AI procurement agent overpays for cloud compute by 10x due to a pricing API error. Who absorbs the loss?
  • An agent autonomously subscribes to a service that turns out to be fraudulent. Can the user file a chargeback?
  • Two agents negotiate a price for an API call. One agent's principal (the human or company behind it) claims the agreed price was unauthorized. Which entity is bound by the agreement?
  • An agent makes a payment that violates sanctions. Is the agent's operator liable, the developer who built the agent, or the platform that provided the wallet?

Card networks have offered initial positions. Mastercard says its standard chargeback protections apply if users dispute a transaction completed by an AI agent. Visa frames agent-initiated transactions as subject to the same dispute resolution processes consumers expect from card-based transactions. American Express went further, describing its Agent Purchase Protection as an “industry-first commitment” to back transactions made by registered agents. But all three approaches retrofit human-centric frameworks onto agent behavior without addressing the fundamental question: does the agent have authority to bind its principal?

How Regulators Are Responding (or Not)

Regulatory engagement varies dramatically across jurisdictions, from active experimentation in Hong Kong to near-silence in the United States.

United States: the GENIUS Act gap

The GENIUS Act, signed into law in 2025, establishes a licensing framework for stablecoin issuers but says nothing about how autonomous agents use those stablecoins. As Forkast has reported, agent payments were never in the draft. The law covers issuance, reserves, and redemption: it does not address authorization, liability, or consent for autonomous systems.

The OCC, FDIC, and Treasury released proposed stablecoin rules in spring 2026, but as of November 2026 the Federal Reserve has not published a standalone proposal. No federal agency has issued guidance specifically addressing AI agent payments. The Consumer Bankers Association published a white paper on agentic AI payments in January 2026, but it raised questions rather than proposing answers.

Hong Kong: Project SAGE

The Hong Kong Monetary Authority has taken the most proactive approach. On November 2, 2026, the HKMA launched Project SAGE (“Securing Agentic Finance with Governance and Enablement”) alongside an Agentic Finance Expert Group. The project covers three use case domains: consumer commerce, business procurement, and machine-to-machine payments. It explicitly addresses “Know Your Agent” (KYA) practices and authentication mechanisms, suggesting Hong Kong may develop an agent identity standard analogous to KYC.

European Union: PSD3 and the digital euro

The EU is approaching the problem through its payments regulation overhaul. PSD3 introduces consent dashboards that let customers see and manage permissions granted to third-party providers, laying groundwork for granular delegation of authority to AI agents. Permissions must be specific, time-limited, and revocable. An IMFS working paper by Grabowski and Costea found that typical agentic payment use cases are not currently listed as high-risk AI systems under the EU AI Act, but that a protocol-only payment model may amount to payment initiation services under PSD2, requiring licensing and strong customer authentication.

Separately, the ECB's digital euro program has opened its innovation platform to AI-enabled payment experiments, with prototypes expected in the first half of 2027.

United Kingdom: flagging the problem

The FCA has flagged that per-transaction payer consent requirements are incompatible with autonomous agent payments, but has not proposed a solution. The UK's approach thus far has been to identify the regulatory friction without resolving it.

Jurisdictional Comparison

JurisdictionAgent account statusActive initiativeTimeline
United StatesNo framework; banks actively blockingNone at federal levelNo proposal pending
Hong KongExploring KYA standardsProject SAGE + Expert GroupLaunched Nov 2026
European UnionPSD3 consent dashboards (not agent-specific)Digital euro AI experimentsPrototypes expected H1 2027
United KingdomFCA flagged consent incompatibilityNo formal initiativeUndetermined
Crypto (global)Permissionless: no identity requiredx402, AP2, Agentic WalletsLive since early 2026

One proposed solution is to give AI agents legal form through new entity structures. If an agent can be wrapped in a legal entity, it gains the capacity to hold accounts, enter contracts, and bear liability.

The Delaware AIC proposal

In July 2026, Delaware's Secretary of State and Norm Ai proposed an “Artificial Intelligence Company” (AIC): a new entity type whose day-to-day operations are managed by an AI agent rather than a human. The AIC would be a separate legal entity that can sue, be sued, hold property, and incur obligations. Each AIC would require a human or corporate member to fund it, and it must maintain a log of all activities. The proposal contemplates a regulatory sandbox, not general-purpose deployment.

The LLC workaround

Legal scholar Shawn Bayern has argued that existing LLC law already allows software to exercise effective control. Under this theory, a human forms an LLC, drafts an operating agreement delegating management to software, and the LLC (not the software) holds the legal identity. The agent operates the entity; the entity holds the account. This requires no new legislation, but it stretches LLC law beyond its intended purpose and has not been tested in a financial-services context.

DAO LLCs

Several US states (Wyoming, Tennessee, Vermont) have enacted DAO LLC statutes that allow decentralized autonomous organizations to register as limited liability companies. A March 2025 Oxford University Press chapter argued these DAO LLCs could serve as a precursor to AI legal personhood, though no AI-managed DAO LLC has been formed to date.

The entity approach has limits: Even if an AI agent can hold an account through a legal wrapper, the entity still needs to pass KYC, maintain beneficial ownership records, and comply with anti-money laundering obligations. The legal wrapper solves the identity problem but not the compliance problem.

Why Micropayments Make the Gap Visible

The regulatory gap matters most for micropayments, precisely the transaction type AI agents generate most frequently. When an agent pays $0.003 for an API call, $0.01 for a data lookup, or $0.05 for a compute burst, traditional payment rails are not just inconvenient: they are unusable.

  • Card networks impose minimum transaction amounts, typically $0.15 to $0.30 depending on the processor
  • Interchange fees on a $0.01 transaction would exceed the transaction value
  • ACH transfers take hours to days and are designed for batch settlement, not real-time per-request payments
  • Wire transfers carry fixed fees that make sub-dollar payments economically irrational

This is where stablecoin payment rails become essential. On low-fee Layer 2 networks, stablecoin transfers can settle for under one cent, making sub-dollar agent payments viable. The x402 protocol was specifically designed for this use case: HTTP-native payments small enough that card economics do not apply.

Stablecoin Rails and Bitcoin Settlement for AI Agents

The emerging stack for AI agent payments pairs stablecoin denomination with crypto settlement. Agents hold and spend dollar-denominated stablecoins for predictable value, while the underlying settlement occurs on blockchain infrastructure that requires no identity verification or institutional relationship.

For agents that need Bitcoin-settled transactions, Layer 2 protocols offer a path to instant, low-cost transfers. Spark, for example, supports programmable payments with dollar-denominated stablecoins like USDB settled on Bitcoin through its statechain-based architecture. The combination of programmatic spending policies and sub-cent transaction costs maps directly to the requirements of autonomous agents: no accounts to open, no identity to verify, and no minimum transaction size. Developers building agent payment infrastructure can explore the Spark SDK for self-custodial stablecoin integration.

What Happens Next

The regulatory gap will not last. AI agents are generating real economic activity, and the volume is growing fast enough that regulators cannot ignore it. The likely trajectory involves three phases.

First, sandbox experimentation. Hong Kong's Project SAGE and the ECB's digital euro innovation platform represent this phase: testing agent payment frameworks in controlled environments before writing permanent rules.

Second, delegated authority frameworks. Rather than granting agents their own accounts, regulators will likely formalize the delegation model: a human or entity authorizes an agent to act within defined parameters, and the authorizing party retains liability. PSD3's consent dashboards are an early version of this approach. This model works for agent payment authorization but requires new standards for expressing, verifying, and revoking delegated authority.

Third, purpose-limited legal identity. Delaware's AIC proposal hints at this endpoint: a legal form designed specifically for autonomous software, with built-in constraints, logging requirements, and mandatory human oversight. This is the furthest out and the least certain.

In the interim, crypto rails will continue to serve as the de facto infrastructure for agent payments. They work today, they require no regulatory approval to deploy, and they align with the permissionless, programmable properties that agents need. For developers building agentic payment systems, the architecture is already live: explore wallets like General Bread to see how Spark-powered self-custodial stablecoin payments work in practice.

This article is for educational purposes only. It does not constitute financial or investment advice. Bitcoin and Layer 2 protocols involve technical and financial risk. Always do your own research and understand the tradeoffs before using any protocol.