How does the U.S. Treasury regulate AI in banking?

Treasury's AI work runs on three tracks. First, as the Sector Risk Management Agency for financial services, its Office of Cybersecurity and Critical Infrastructure Protection (OCCIP) produced the March 2024 report on managing AI-specific cybersecurity and fraud risks (written under Executive Order 14110) and then, with the FBIIC and the Financial Services Sector Coordinating Council, stood up the Artificial Intelligence Executive Oversight Group (AIEOG). The AIEOG's six workstreams — AI Lexicon and taxonomy, a Financial Services AI Risk Management Framework, explainability, data 'nutrition labels', AI-enhanced fraud, and identity and authentication — were announced complete on February 18, 2026, with the Lexicon and FS AI RMF published the next day as non-binding resources.

Second, Treasury's Domestic Finance office ran the June 2024 Request for Information on the uses, opportunities and risks of AI in financial services and published its findings on December 19, 2024, recommending that regulators clarify supervisory expectations, close gaps in existing frameworks, and that firms review every AI use case for compliance with existing law before deployment. Third, Treasury chairs the Financial Stability Oversight Council: FSOC first named AI an emerging vulnerability in its 2023 annual report, catalogued explainability, data, performance, third-party and bias risks in 2024, and in its 2025 report pivoted to 'harnessing AI to promote financial stability', creating a standing AI Working Group and the 2026 AI Innovation Series of public-private roundtables.

Under the current administration Treasury's posture is explicitly pro-adoption: its 2026 outputs are framed as implementing Executive Order 14179 ('Removing Barriers to American Leadership in Artificial Intelligence', January 2025) and the July 2025 AI Action Plan. Treasury also practices what it recommends — its Office of Payment Integrity credited machine-learning check-fraud detection with $1 billion of the $4 billion in fraud and improper payments it prevented or recovered in fiscal 2024.

What is an AI treasury risk assessment, and which Treasury frameworks apply?

'AI treasury risk assessment' means two different things, and both route through the same documents. For the U.S. Department of the Treasury it is the sector-level assessment it has run since 2024: the March 27, 2024 report on AI-specific cybersecurity and fraud risks, built on 42 institutional interviews and naming a capability gap between large and small banks and a 'fraud data divide'; the December 19, 2024 report synthesising 103 responses to its AI request for information; FSOC's annual AI vulnerability assessments; and, since February 19, 2026, the Financial Services AI Risk Management Framework (FS AI RMF), which adapts NIST's Govern–Map–Measure–Manage cycle to banks and is the closest thing to an official template. For a bank's own treasury function — liquidity and cash-flow forecasting, asset-liability management, funding, hedging and payments — an AI risk assessment is the use-case-level review both Treasury reports recommend before deployment, carried out under the model risk (SR 26-2), third-party (SR 23-4) and risk-data (BCBS 239) frameworks that already govern treasury models, with fraud and cyber controls layered on for payments.

The U.S. Treasury's own assessment has a structure worth copying because the sector's regulators helped write it. The March 2024 report, produced by the Office of Cybersecurity and Critical Infrastructure Protection under Executive Order 14110, sorted AI risk into a capability gap (large institutions build in-house, small ones lack data and expertise), a fraud data divide (too little cross-firm fraud data to train models), data supply-chain opacity (nobody can say what trained a vendor model or how customer inputs are reused), explainability of black-box and generative systems, and regulatory fragmentation. Each of those became a workstream of the AIEOG public-private group, and two of its six deliverables — the AI Lexicon and the FS AI RMF — were published on February 19, 2026. The remaining four (explainability, data 'nutrition labels', AI-enhanced fraud, identity and authentication) are still to come; a bank's assessment will need updating when they land.

The FS AI RMF is the practical template. It keeps NIST's four functions and adapts them to what a financial institution has to prove: under Govern, who is accountable for the AI use case and how it fits the risk appetite and existing regulatory obligations; under Map, what the system does, what data it uses, who is affected and what could go wrong; under Measure, how performance, robustness, bias, security and explainability are tested; under Manage, how the residual risk is accepted, monitored and retired. It is explicitly scalable — the AIEOG's stated aim was resources that small and mid-sized institutions can use — and explicitly non-binding. Binding expectations still arrive through the prudential regulators, which is why the requirement stack above pairs each Treasury document with the supervisory rule that examiners actually cite.

Inside a bank, the treasury function is where AI use cases and regulatory frameworks overlap most densely. Cash-flow and liquidity forecasting models are squarely inside the 2026 model risk definition and are among the most material models a bank runs, so they get full validation and outcomes analysis. Asset-liability and hedging models that reach into markets pick up the Fed's financial-stability concerns about correlated AI-driven strategies. Payments are the operational-resilience critical operation and the primary target of AI-enabled fraud — FinCEN's deepfake alert is addressed to exactly the controls a treasury team runs. Generative-AI assistants drafting funding memos or summarising counterparty documents sit outside model risk guidance and inside NIST AI 600-1's confabulation and data-leakage risks. And almost all of it arrives through vendors — treasury management systems, bank portals, cloud-hosted models — which makes SR 23-4 the first framework in the sequence, not the last.

A defensible AI treasury risk assessment therefore has five steps, each anchored to a document. Inventory every AI-enabled system in the function, including AI features switched on inside vendor platforms (SR 23-4; ESAs' July 2026 call for asset inventories that include AI components). Tier each by materiality — the FSB's Sound Practice 5 and the 2026 model risk guidance both make intensity of control follow materiality. Map each tier to its requirement stack: model risk, third-party, BCBS 239 data lineage, operational resilience, fraud and cyber, and consumer law where customers are affected. Measure against the FS AI RMF's Measure function — validation, robustness and adversarial testing, explainability for the decision-makers who rely on the output, drift monitoring. Manage: a named owner, a human able to override, thresholds at which automation stops, vendor exit plans and a re-assessment date, since the Treasury's December 2024 report asks for periodic re-evaluation rather than a one-time review.

What changes next is scheduled. The remaining AIEOG deliverables will add explainability and data-labelling expectations to the Map and Measure steps. The interagency request for information on AI and model risk management, promised in the April 2026 guidance, is the first federal document that will address generative and agentic AI in bank models directly. The FSB's final sound practices are due in October 2026 as a G20 deliverable, and FSOC's 2026 annual report in December will be the first full-year output of its AI Working Group. A treasury AI assessment written today should carry those three dates as review triggers.

WHAT THIS MEANS IN PRACTICE

• Start the assessment with the vendor inventory, not the model inventory: most AI in a bank treasury function arrives as a feature inside a treasury management system or bank portal, and SR 23-4 is the framework that reaches it.
• Use the FS AI RMF's four functions as the section headings of the assessment document. It is the template Treasury and the sector regulators wrote together, and it maps one-to-one onto NIST, which examiners and auditors already recognise.
• Classify liquidity and cash-flow forecasting models as high-materiality under the 2026 model risk guidance and validate them accordingly; a generative-AI assistant that drafts the commentary around them is a different tier with a different control set.
• Put deepfake-resistant payment controls in the AI risk assessment, not only in the fraud programme: FinCEN's red flags, phishing-resistant MFA and live re-verification of counterparties are the mitigations for the AI risk most likely to cost a treasury team money this year.
• Document data lineage for every input to an AI treasury model. BCBS 239 is the standard supervisors cite when they ask how the data feeding AI is controlled, and the Basel Committee said in January 2026 that lineage is where banks still fall short.
• Define the human override for every tier — who can stop an automated funding, hedging or payment action, on what evidence, and how fast — and test it, because the FSB, the OCC and the EU AI Act all treat human oversight as the control that scales with autonomy.
• Schedule the re-assessment now: the remaining AIEOG deliverables, the interagency AI and model risk RFI, the FSB final report (October 2026) and FSOC's 2026 annual report (December 2026) are the four events that will change the requirement stack.

What has the U.S. Treasury actually published on AI?

Which of the 100 largest US banks answer to the U.S. Treasury on AI?

2 of the 100 bank pages on this site name the U.S. Treasury among the authorities their AI programme answers to. Each page lists the documents that apply and why.

• Publication and uptake of the remaining AIEOG deliverables (explainability, data nutrition labeling, AI-enhanced fraud, identity and authentication) and whether examiners begin referencing the FS AI RMF
• FSOC's 2026 annual report (expected December 2026): the first full-year output of the AI Working Group and any follow-through on the 'regulatory impediments' identified in the AI Innovation Series
• Whether Treasury or FSOC translate Innovation Series feedback into concrete asks of the banking agencies — e.g., harmonized AI guidance or safe harbors for AI-enabled fraud and cyber defense
• Treasury's own AI deployment (AI Transformation Office, Chief AI Officer appointed June 2025) and updated fraud-prevention figures from the Bureau of the Fiscal Service