Regulators Catch Up With AI Agents That Spend: Banks Push Liability to Customers

The OCC issued its first agentic AI supervisory framework, Wells Fargo is shifting the risk of AI agents onto customers, and Block is paying $764 million for Cash App’s weak AML checks. Machine-driven payments are being repriced use case by use case.

On November 4, the OCC published a 22-page “super intelligence” supervisory framework, telling bank examiners for the first time how to review agentic AI. The same day, Block disclosed in its quarterly filing that it had reached a deal in principle with the DOJ over Cash App’s anti-money-laundering failures, setting aside $764 million. Earlier, Wells Fargo updated its online banking agreement: from November 24, customers bear full responsibility for AI agents given access to their accounts, including losses from actions an agent takes beyond its authority. Three institutions, three tools, one direction: the speed at which machines move money is being repriced, use case by use case.

These three events do not look like coincidence. After Meta’s Muse personal assistant launched, the trend of consumers handing banking credentials to AI agents started to be taken seriously. Apollo Global Management chief economist Torsten Slok warned that AI agents would automatically sweep household balances from checking accounts averaging roughly 0.1% yield into alternatives yielding 3.3% to 5%, potentially costing banks $47 billion a year in low-cost deposits. Regulators and banks alike realize that what stands between agents and the plumbing is now just contracts and examination manuals.

I. The OCC Framework: Graded by Consequences, Not Model Sophistication

OCC Bulletin 2026-54 groups 16 components under six pillars: Purpose, Risk, Control, Oversight, Resilience, and Assurance. The framework grades AI use cases by “consequences” rather than technical sophistication. An agent authorized to initiate payments (Tier 1) poses greater risk than a complex model used for drafting documents; real-time credit decisioning, fraud-hold release, and market access are all classified as critical tiers requiring tighter controls.

The framework defines “super intelligence” as the technologies previously called AI, explicitly covering generative and agentic systems excluded by the April interagency model-risk guidance (Bulletin 2026-13). The OCC stresses that this is a resource for examiners, does not establish supervisory expectations for banks, and does not require banks to define a separate AI risk appetite. The Federal Reserve and FDIC are likely to mirror similar frameworks.

For banks, the priority is that every AI use case must answer six questions: what the system is authorized to do, what happens if it fails, who is watching it, whether it can recover from failure, whether controls remain effective, and how third-party vendor accountability is allocated. The OCC also notes that banks can use contracts to decide who pays when an agent gets things wrong, a sentence pointing directly at the fuzziest link in the agentic AI liability chain.

II. Wells Fargo’s Choice: Let Agents In, Put the Risk on Customers

Section 15 of Wells Fargo’s updated Online Access Agreement, titled “Third-Party Artificial Intelligence Agents or Agentic AI,” takes effect November 24. The definition is broad: any software, application, or program that takes autonomous, semi-autonomous, or non-autonomous action on behalf of a person or entity counts as an AI agent. When a customer gives an agent login credentials, every instruction, action, or transaction the agent initiates is deemed to have been initiated and authorized directly by the customer, and the bank has no obligation to investigate the agent’s authority.

The core of the clause is “AT YOUR OWN RISK”: customers are responsible for the agent’s unintended actions, actions caused by incomplete, inaccurate, or manipulated inputs, and the agent exceeding its authority, unless law provides otherwise. Customers who want to revoke an agent must do so “properly and timely,” retain evidence of the revocation, and provide it at the bank’s request; responsibility continues until the bank is notified. Wells Fargo also reserves the right to require customers to register agents, change or discontinue access methods, and terminate agent access at any time.

This contrasts with the approach of JPMorgan, Bank of America, and Citi. Within one week (October 13-20), the four big banks each took a position: JPMorgan, Charles Schwab, and Citigroup chose to block third-party agents from logging into customer accounts with credentials, while Wells Fargo chose to let them in but shift all risk onto customers. Behind both postures is the same judgment: credential-based access makes an agent indistinguishable from the customer, defeats MFA by design, and leaves no separate audit trail.

III. Block’s $764 Million: AML Is the Entry Ticket to the Machine Age

Block disclosed in its November 3 quarterly filing that it reached a deal in principle with the DOJ over the Cash App AML probe, setting aside $764 million. The DOJ says Block failed to maintain an effective anti-money-laundering program as the Bank Secrecy Act requires. The charge would be held for three years and dropped if Block meets the deal’s terms. The same day, Cash App’s gross profit grew 27%.

The probe traces back to a 2023 short-seller report questioning Cash App’s identity checks. In July, Block paid $45 million to settle state claims that Cash App failed to protect users from fraud. The growth built on light identity checks is now being paid for in the hundreds of millions. For crypto and AI payments, the signal is more direct: FinCEN and DOJ scrutiny of AML for “machines spending on behalf of customers” will only tighten, because agents can naturally automate around identity-verification flows.

IV. Deep Analysis: Three Fault Lines in the Liability Chain

V. Investment Perspective

Circle’s stock rose to $77 after earnings but remains down 35% year to date. The market is skeptical that one-off ARC presale revenue can support 2027 growth. USDC accounts for 79% of stablecoin transaction volume, yet circulation is stalled, showing that “fast turnover does not equal high balances”, a fundamental challenge to issuers’ valuation logic.

After the midterms, Democrats control both chambers, with Elizabeth Warren positioned to lead the Senate Banking Committee and Sherrod Brown returning to the Senate. TD Cowen expects the CLARITY Act to make little progress over the next two years, shifting industry focus from legislation to regulatory implementation. The OCC examination framework is the first concrete signal of that shift: the rules are not gone, they just have a different author.

VI. Conclusion: The Compliance Cost of Agentic AI Is Becoming Calculable

For practitioners, this week’s three events add up to one judgment: accountability for machines that spend is moving from a technical problem to a contract-and-examination problem. The OCC supplied the examination manual, Wells Fargo supplied the liability-allocation template, and the DOJ supplied the fine scale. The differentiator for agent-payment startups will shift from “how fast can it run” to “who can explain who is responsible when things go wrong.”