Tuesday, November 17, 2026

The Stablecoin Is Backed, But Who Authorized The Payment? – Analysis



Key Takeaways:

As AI agents begin initiating payments, a technically valid, fully reserved stablecoin transfer can still be one the customer never authorized, exposing a gap between issuer soundness and payment-instruction authority.

With GENIUS Act rules, MiCA, Hong Kong licensing and Bank of England infrastructure plans advancing, the article calls for portable, auditable mandate records that capture payer, beneficiary limits, amounts and validity periods so disputes can be traced across issuers, wallets and jurisdictions.

It urges regulators to test actual dispute scenarios—forged invoices, compromised credentials, conflicting instructions—rather than only token reserves, stressing that irreversible settlement and multi-provider chains require clear evidence standards before automated payments become routine.


As global regulators strengthen the reserves behind digital money, AI-driven payments raise another question: how will customers prove what they actually approved, and who must answer when that approval fails?

A payment can be backed by safe assets, signed with a valid cryptographic key and recorded correctly on a blockchain, and still be a payment the customer never intended to make. That is the distinction regulators must confront as artificial-intelligence agents begin arranging transactions rather than merely recommending them.

The timing is unusually good. On November 11, the Bank of England described future payment infrastructure that could support advanced delegated payments, programmable transactions and stablecoins alongside more familiar forms of money. In the United States, the Office of the Comptroller of the Currency finalized GENIUS Act rules in October, while the Federal Reserve is taking comments until November 30 on its own implementation proposal. Neither development makes autonomous stablecoin payments inevitable. Together, they expose a question that should be settled before the technology becomes ordinary: what proves that software had authority to spend a particular amount, for a particular purpose, on a particular recipient?

A sound reserve does not validate the instruction

The American GENIUS Act created a framework for payment stablecoin issuers, including reserve, redemption and other safeguards. Its implementation is substantial, not cosmetic. The OCC’s final rule, effective January 18, 2027, also covers information security, internal controls, independent testing and business continuity. It would be inaccurate to claim that Washington regulated reserves and forgot operational risk.

But issuer safety and payment authorization are different questions, often answered by different companies. An issuer can maintain sufficient reserves while a customer’s wallet, an intermediary or a business software agent sends funds to the wrong destination. The issuer might have complied with its own obligations; the customer might still face a loss. Equally, a failure in an issuer’s minting or administrative controls has a different reach from a mistaken payment out of one wallet. Supervisors should not pretend these are one failure with one accountable actor.


Consider a hypothetical importer using an AI agent to pay an overseas supplier in dollar-denominated stablecoins. A forged invoice changes the supplier’s payment address. The agent reads it, initiates a transfer and a legitimate signing service approves the transaction under credentials granted by the importer. The blockchain now records a technically valid transfer. What is missing is evidence that the importer approved the changed beneficiary. No reserve report, however accurate, answers that question.

There is no need to describe this as a sentient or “rogue” machine. NIST’s 2026 work on AI-agent security identifies risks that arise when language-model outputs interact with real software tools, including manipulated inputs and actions that exceed their intended purpose. Ordinary credential compromise and poor payment controls can produce similar consequences without AI. The new feature is the speed and reach of systems entrusted to turn instructions into payments.
Four financial rulebooks, one cross-border payment

The regulatory map is moving quickly, but not uniformly. The European Union’s MiCA regime governs important parts of stablecoin issuance and crypto-asset services. In an October 8 opinion, ESMA directed supervisors toward ending services involving stablecoins that fail to meet MiCA requirements. Hong Kong’s licensing regime began in 2025, and its monetary authority granted its first two stablecoin issuer licences in April 2026. Britain’s June 2026 policy statement laid out its approach to sterling-denominated systemic stablecoins, with further rules still being developed. These instruments have different scopes and are not a single global law.

What connects them is the boundary between money, service provider and user. A token could be issued in one jurisdiction, held through a service in another and used by an AI-assisted business process somewhere else. A customer disputing a transfer might confront an issuer, custodian, wallet provider, software integrator and merchant, none of which holds the complete record of the instruction.


The Financial Stability Board warned in October that global stablecoin arrangements cut across borders, sectors and supervisory authorities, making vulnerabilities difficult for any single regulator to monitor. This is not evidence that AI agents have already caused a global stablecoin crisis. It is a reason to design authorization evidence and dispute handling across the payment chain while the ecosystem is taking shape.
Give delegated payments a record of authority

A practical improvement would be a portable, privacy-conscious record of the payment mandate, produced by the institution authorizing an agent to act. It would not be a new public ledger of customer identities. It would be an auditable record, available to the customer and appropriate investigators, connecting the transaction to the permission that existed when it was initiated.

For consequential automated payments, the record should establish the authorized payer, the intended beneficiary or controlled beneficiary list, the relevant asset and network, the maximum amount or exposure, the validity period, and the service responsible for approving exceptions. If a beneficiary or spending limit changes, the system should require a fresh authorization under the applicable rules. An AI assistant may draft the instruction, but its explanation or apparent confidence should not itself expand the mandate.

For customers, the key question is not whether a wallet address had a valid signature. It is whether the transfer matched the mandate they gave. For regulators, the question is which institution possessed the evidence needed to prevent, detect or investigate a departure from that mandate. This division is compatible with different technical designs and need not require every payment to pass through a new central authority.

There is a serious counterargument: card networks, banks and payment firms already handle authentication, fraud and disputes. Stablecoins should not be burdened with a parallel bureaucracy merely because an AI interface is involved. That objection is right about the need for proportionality. Small, low-risk transactions and supervised business payments do not warrant identical procedures. The case for a shared evidence standard is strongest where agents can move consequential value across multiple providers and legal systems, particularly when ledger settlement does not offer a built-in chargeback.
Test the dispute, not just the token

Regulators should ask payment firms to demonstrate how a customer can challenge an agent-initiated payment. The exercise is straightforward: change an invoice recipient, compromise a delegated credential, submit two conflicting payment instructions or interrupt service during settlement. Can the payer establish which mandate was active? Can the responsible provider identify an unauthorized alteration? How rapidly can further transfers be stopped, and what remedy is actually available under the relevant law and contract?


Recovery must not be sold as guaranteed reversal. Some transfers may be irreversible in practice; some issuers or service providers may have lawful powers to freeze assets in defined circumstances. A backed stablecoin is not automatically a bank deposit with the same protections, and the existence of a control function does not establish a customer’s right to its use. These differences should be clear before money moves, not first explained after it disappears.

The next step does not require a world regulator. Financial authorities can clarify the duties of institutions already under their supervision; payment providers can preserve compatible mandate evidence; and international bodies can compare what consumers and businesses actually receive when disputes cross borders. The Federal Reserve’s current rulemaking is an immediate place to ask how the relevant supervisory duties interact with automated payment services, without claiming that GENIUS Act powers automatically cover every independent wallet or AI developer.

Digital money may improve the speed and reach of global commerce. So may software that pays suppliers or manages recurring transactions without a person reviewing every routine step. But the promise should be judged by more than the quality of the reserves or the speed of settlement.

A stablecoin needs enough assets to honor its value. An AI-driven payment also needs evidence that someone was entitled to send it.


About Burak Oktenli
Burak Oktenli holds an MBA and a Master of Professional Studies in Applied Intelligence from Georgetown University. His research addresses the governance of authority in autonomous and AI-enabled systems, and his writing has appeared at the Modern War Institute at West Point, RUSI, RealClearDefense, RealClearMarkets, and Geopolitical Monitor. He is the author of Authority Architectures for Autonomous Systems, a ten-volume series on how authority in autonomous systems is delegated, monitored and recovered, at authority-architecture.me.
View all posts by Burak Oktenli →

No comments: