AI Agent Compliance in the US: A Guide for Fintechs and Banks

In the US, AI agents at fintechs and banks are governed by the rules that already cover the work they do, not by a single AI law. That means bank supervisory guidance on models and third parties, consumer protection law, data security rules, and a growing set of state AI laws, with the NIST AI Risk Management Framework as the shared voluntary reference.

This is a guide to those rules for compliance, risk, and engineering leaders at US payment companies, fintechs, and banks: what each one asks, when it applies, and how Runtime's controls produce the evidence. It is the US companion to AI agent risk management, which covers the controls in depth. For other regions, see the UK guide and the EU guide.

What it covers. The federal and state rules that matter for agents that investigate payouts, answer disputes, review onboarding files, or prepare responses to a partner bank.

What it is not. Legal advice. Every material improvement comes from a human questioning a premise. Plan for that, and have counsel check scope for your charter and your states.

At a glance

Is there a federal AI law?

No. Congress has not passed a general AI law. Executive Order 14365, signed December 11, 2025, directs the Attorney General to form an AI Litigation Task Force to challenge state AI laws, asks Commerce to identify state laws it considers onerous, and asks for a federal framework that would preempt them. It does not itself override any state law. Until a court or Congress acts, plan to comply with the state laws below.

The federal reference point is the NIST AI Risk Management Framework (January 26, 2023) and its Generative AI Profile, NIST-AI-600-1 (July 26, 2024). Both are voluntary, and both are what partner banks and examiners most often use to organize AI questions.

What it asks. Four functions: Govern (policies, roles, accountability), Map (context and risks of each system), Measure (assess and track risk), and Manage (prioritize and act). The generative AI profile adds risks such as confabulation, data privacy, and information security.

How Runtime helps. Runtime's control mapping already follows the four functions: org-wide rules, agent owners, RBAC and SSO for Govern; the agent inventory for Map; per-run risk scores and evals for Measure; approval gates and tiered guardrails for Manage. The mapping exports with the evidence.

Federal banking rules

Does SR 26-2 apply to AI agents?

On April 17, 2026, the Federal Reserve, OCC, and FDIC issued revised model risk guidance as SR 26-2, OCC Bulletin 2026-13, and FDIC FIL-15-2026. It replaces SR 11-7 (2011) and SR 21-8 (2021).

The guidance is explicit: generative AI and agentic AI models "are not within the scope of this guidance." It still applies to traditional statistical models and non-generative AI. It also says a bank's own risk management and governance practices should guide the controls for tools it does not cover. It is most relevant to banks above $30 billion in assets, and it is not enforceable on its own, though unsafe or unsound practices still draw supervisory action.

What it asks. For agents: nothing specific. In practice, banks govern agents under their broader risk and governance programs, and many reuse model risk habits such as inventory, testing before changes, and ongoing monitoring.

How Runtime helps. An inventory of every agent with an owner and scoped access, evals that re-check an agent whenever its model, skills, or tools change, a change history with who approved each change, and per-run risk scores for ongoing monitoring.

What does a sponsor bank have to oversee?

The Interagency Guidance on Third-Party Relationships: Risk Management, final June 6, 2023, says banks must manage the risk of third parties, including fintech partners, across planning, due diligence, contracts, ongoing monitoring, and termination. Your agents are part of how you deliver the bank's product, so they are part of what the bank oversees.

What it asks. Evidence that your controls work, access for the bank to review them, and notice when something material changes.

How Runtime helps. A record of every run, approvals in the dashboard or Slack, and an evidence packet you can export for any agent and period. Agents can run in your own AWS, GCP, or Azure account or self-hosted, which makes the bank's data questions simpler.

Consumer protection

Do adverse action rules apply when AI is involved?

Yes. Under ECOA and Regulation B, a creditor that denies credit must give the specific principal reasons, whatever technology made the decision. The CFPB had spelled this out for complex algorithms in Circular 2022-03 and Circular 2023-03. On May 12, 2025 it withdrew both, along with 65 other guidance documents. The regulation itself did not change.

The same goes for unfair, deceptive, or abusive acts or practices under the Dodd-Frank Act and unfair or deceptive practices under Section 5 of the FTC Act. An agent that tells a customer something misleading creates the same exposure as a person who does.

What it asks. Keep agents out of credit decisions unless you can explain the reasons, and review what agents say to customers.

How Runtime helps. Approval gates so nothing reaches a customer without a person, a run record that shows what the agent looked at and why it recommended what it did, and risk scores that flag runs touching regulatory decisions.

Data and security

What does GLBA require for AI agents?

The FTC Safeguards Rule (16 CFR Part 314) applies to non-bank financial institutions, which covers most fintechs. Banks follow the banking agencies' equivalent standards. It requires a written security program, a qualified individual in charge, risk assessments, access controls, encryption, MFA, monitoring of user activity, and oversight of service providers. Since May 13, 2024, you must notify the FTC within 30 days of discovering a security event involving the unencrypted information of at least 500 consumers.

PCI DSS v4.0.1, published June 2024, applies if your agents could see card data. Its future-dated requirements became mandatory on March 31, 2025.

What it asks. Agents count as users: limit what they reach, log what they do, and keep card numbers and personal data out of places they don't need to be.

How Runtime helps. Each agent gets its own identity and scoped API keys rather than a teammate's login. Card numbers, SSNs, and secrets are masked before any prompt or log, which keeps card data out of agent sessions. Agents run in isolated sandboxes wiped at session end, egress allowlists limit where data can go, and Runtime never trains on your data. Runtime is SOC 2 audited.

States

What does NYDFS expect for AI?

23 NYCRR Part 500 applies to DFS-licensed entities, including many money transmitters, lenders, virtual currency businesses, and state-chartered banks. The second amendment took effect November 1, 2023, and its last phase, MFA for anyone accessing information systems and a complete asset inventory, applied from November 1, 2025.

DFS has explained how Part 500 applies to AI in two letters. The October 16, 2024 letter covers AI-enabled social engineering, AI-enhanced attacks, exposure of nonpublic information, and third-party dependencies. The May 21, 2026 letter on frontier AI models asks for testing and human oversight of AI-generated code before production, and for logging and alerting that keep up. Neither adds new requirements.

How Runtime helps. Agents appear in the inventory with an owner, access is scoped and gated by RBAC and SSO, production deploys wait for a named approver, and every run is logged.

California

The CPPA's regulations on ADMT, risk assessments, and cybersecurity audits were approved September 23, 2025 and took effect January 1, 2026. Risk assessment duties started January 1, 2026, with first submissions due April 1, 2028. ADMT rules for significant decisions, which include financial or lending services, apply from January 1, 2027. Cybersecurity audit certifications are due April 1, 2028, 2029, or 2030 depending on revenue. Personal information covered by GLBA is exempt from the CCPA, so scope depends on the data an agent touches.

California's SB 53, signed September 29, 2025, applies to developers of frontier models, not to companies deploying agents built on them.

Colorado

SB 26-189, signed May 14, 2026, repeals and replaces the 2024 Colorado AI Act (SB 24-205) before its obligations took effect. From January 1, 2027, deployers of covered automated decision-making technology that materially influences consequential decisions, including financial services, must give notice at the point of interaction, explain the technology's role within 30 days of an adverse decision, offer human review on request, and keep records for three years. The Attorney General enforces it; there is no private right of action.

Texas and Utah

Texas TRAIGA (HB 149) took effect January 1, 2026. For private companies it mostly bans a short list of intentional harmful uses, such as deploying AI with intent to unlawfully discriminate, and the Attorney General enforces it.

Utah's AI Policy Act, as amended by SB 226 (2025), requires businesses to disclose generative AI use when a consumer clearly asks, and requires regulated occupations to disclose it up front in high-risk interactions, such as collecting financial data or giving personalized financial advice. It is set to sunset in July 2027.

How Runtime helps across states. Approval gates give you the human review step, the run record shows what an agent did in any decision, risk scores flag runs that touched a regulatory decision, and the export answers a record request in one step.

What a sponsor bank will ask

A 30-day plan

1. Week 1: Write your agent policy. One page: which actions agents may never take, which need a person, and who owns each agent. This step is human.
2. Week 1: Inventory every agent. Give each one its own identity, an owner, and scoped access, including agents running on a laptop.
3. Week 2: Scope the rules. With counsel, mark which agents touch credit decisions, consumer communications, card data, or New York, California, or Colorado customers.
4. Week 2: Tier guardrails. Read-only, customer-facing, and money or production, with approvals on the last two.
5. Week 3: Set evals. Build an eval set from resolved cases for each agent so model, skill, and tool changes get re-checked.
6. Week 4: Run a mock exam. Export an evidence packet for one agent and one month, and walk your sponsor bank's questions against it.

The point

US rules for AI agents are scattered across banking guidance, consumer law, security rules, and a few states, but they ask for the same things: know your agents, limit what they touch, keep a person on sensitive decisions, and keep the record. Do that once and every examiner, auditor, and sponsor bank gets the same answer. The AI agent risk management guide covers how to run those controls day to day.