How the regulatory map for AI in banking actually fits together
Picture a mid-size bank rolling out an AI model that scores small business loan applications. A fintech vendor built the model. It was trained on data licensed from a credit bureau, runs on a cloud platform, and is used by loan officers in twelve states. Before it approves a single loan, it already sits inside the jurisdiction of at least four federal regulators and a patchwork of state laws. None of them wrote a rulebook specifically for "AI." Each stretches older authority to cover it, and since 2025 several of those authorities have shifted. The skill you need is knowing which lens applies, when, and how stable it is.
Why there is no single "AI law" for banks
The US has no comprehensive federal AI statute for financial services. Nothing like the EU AI Act applies to US banks. The EU Act classes creditworthiness scoring as high-risk. Its high-risk obligations were originally due in August 2026, but the EU's late-2025 Digital Omnibus proposal would push them back, so check the current date before relying on it.
US regulators instead apply existing banking, consumer protection and fair lending law to AI systems. The "AI rulebook" for a US bank is several overlapping rulebooks. Each is triggered by a different feature of the same model: what it does, who it touches, and where.
The federal banking regulators: safety and soundness lens
Three agencies supervise banks directly. Each treats AI mainly as a model risk issue: the risk that a model is wrong, misused or poorly understood, and that the error causes financial or reputational harm.
- The OCC (Office of the Comptroller of the Currency) supervises nationally chartered banks. It applies the model risk expectations in SR 11-7, joint Fed/OCC guidance from 2011. SR 11-7 predates modern machine learning, but examiners still use it as the default framework for AI models, covering validation, documentation and ongoing monitoring.
- The Federal Reserve supervises bank holding companies and state-chartered member banks. It co-authored SR 11-7 and asks boards whether they understand the AI embedded in vendor software they did not build.
- The FDIC supervises state-chartered non-member banks with similar model risk expectations. It pays particular attention to third-party AI tools at community banks that have no in-house data science team.
None of the three has issued binding AI-specific rules. They use the lens they applied to statistical credit scoring in the 1990s: can you explain it, validate it, and control it.
The consumer protection lens: CFPB and fair lending
The CFPB (Consumer Financial Protection Bureau) asks a different question. It does not care mainly whether the model is accurate. It cares whether the model treated a consumer unfairly or illegally.
Two statutes do the work:
- ECOA (Equal Credit Opportunity Act, 1974) and Regulation B prohibit discrimination in credit decisions. They also require creditors to give applicants specific, accurate reasons for denial, even when a machine learning model made the call. In 2023 the CFPB said explicitly that generic checklist reasons are not enough for complex models.
- FCRA (Fair Credit Reporting Act, 1970) governs how credit bureau data feeds automated decisions and what disclosures consumers receive.
The federal picture changed sharply in 2025. The CFPB lost most of its staff, dropped a large share of its open enforcement cases, and in May 2025 withdrew dozens of guidance documents. An April 2025 executive order told federal agencies to deprioritize disparate impact enforcement. The OCC then removed disparate impact from its fair lending exam procedures. In late 2025 the CFPB proposed amending Regulation B to stop recognizing disparate impact liability under ECOA.
Disparate impact means a model can be facially neutral (it never uses race as an input) and still produce discriminatory outcomes because a correlated variable, like zip code or education history, acts as a proxy. Federal examiners now press this far less. The risk has not gone away, for three reasons:
- The adverse action notice requirement comes from the ECOA statute itself, not from guidance. A vague "the model said no" is still a violation.
- Intentional discrimination (disparate treatment) remains illegal, and private plaintiffs can still sue.
- State attorneys general and state laws are filling the gap. Massachusetts reached a $2.5 million settlement with a student lender in July 2025 over AI underwriting that allegedly disadvantaged Black, Hispanic and non-citizen applicants.
Federal policy can also swing back with a new administration. A model deployed in 2026 may still be in use in 2030.
The FTC's overlapping claim
The FTC (Federal Trade Commission) polices unfair or deceptive practices under Section 5 of the FTC Act. It has authority over non-bank fintechs, data brokers and AI vendors, which bank regulators cannot supervise directly.
In 2023 the FTC put companies on notice that overselling what an algorithm can do is itself a deceptive practice. Marketing a credit model as "bias-free" without evidence is an example. This applies whether or not the model actually discriminates. The FTC followed with "Operation AI Comply" cases in 2024 against companies making inflated AI claims, and deceptive-claims cases have continued under the current commission.
Most AI credit models are not built entirely in-house. A bank partnering with a fintech underwriting vendor inherits FTC exposure through that vendor, even though the FTC cannot examine the bank the way the OCC can.
State regulators and the patchwork problem
State banking departments, insurance commissioners and attorneys general add another layer, and they are now the most active one.
- The Colorado AI Act (SB 24-205) treats AI used in lending decisions as high-risk. It requires impact assessments, consumer notices and a duty to avoid algorithmic discrimination. Its effective date was pushed to June 30, 2026, and the legislature has kept revising it, so check the current text.
- California's privacy regulator finalized automated decision-making technology (ADMT) rules under the CCPA in 2025. Businesses using ADMT for significant decisions, including financial services, must provide notices and opt-out or access rights, with compliance phased in from 2027. Some bank data covered by federal privacy law (GLBA) is exempt, but banking-adjacent products often are not.
- New York's Department of Financial Services issued AI guidance for insurers in 2024, and its cybersecurity rules reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → AI used by banks it licenses.
The practical effect: a bank operating in 12 states may face several different disclosure, assessment or opt-out requirements for the same model, while federal bank regulators require only one model risk framework.
Putting the mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → together: one model, four lenses
Go back to the small business loan model. Here is how the same deployment gets pulled in different directions:
| Regulator | Question it asks | Legal hook |
|---|---|---|
| OCC / Fed / FDIC | Is the model validated, documented, monitored? | SR 11-7 model risk guidance |
| CFPB | Did a denied applicant get a specific, accurate reason? | ECOA / Regulation B |
| FTC | Did anyone oversell the model's fairness or accuracy? | FTC Act Section 5 |
| States | Is an impact assessment, notice or opt-out required? Is the outcome discriminatory under state law? | Colorado AI Act, CCPA ADMT rules, state fair lending and UDAP laws |
A model can pass the OCC's validation checklist and still violate ECOA if its adverse action notices are vague. It can be technically fair and still trigger an FTC action if the marketing overstated it. It can clear every federal check and still breach Colorado's rules. Compliance with one lens is not compliance with all of them.
Knowledge check
1. Why does an AI model used by a US bank fall under multiple regulatory regimes at once, rather than one dedicated AI law?
2. How do the federal banking regulators (like the OCC) primarily frame the risk posed by AI models used in banking?
3. SR 11-7 was originally written in 2011, before modern machine learning existed. What does its continued use for AI model oversight illustrate about US financial AI regulation?
4. Select ALL correct answers about why a single bank AI model can trigger oversight from multiple federal regulators simultaneously.
Select all the correct answers.
5. Select ALL correct answers describing how the EU AI Act differs from the US regulatory approach to AI in banking, as described in the lesson.
Select all the correct answers.
Practical guardrailsguardrailsRules and controls that keep an AI system inside safe, legal and on-brand boundaries, blocking outputs and actions that cross the line.View full definition → before deployment
A defensible pre-deployment checklist for an AI credit or risk model looks like this:
- Model documentation good enough for SR 11-7 style validation. Record what data trained the model, its known limitations, and who owns ongoing monitoring.
- Adverse action reasoning that produces specific, individualized denial reasons, not just an overall score. This usually means explainability tooling (SHAP values, for example) on top of the model.
- Disparate impact testing before launch and periodically afterward as the model or applicant population drifts. Keep doing this even though federal examiners ask less often. State laws and private litigants still rely on it, and federal policy may change again.
- Vendor due diligence that treats third-party AI as carrying the bank's own exposure. The 2023 interagency guidance on third-party risk management (OCC, Fed, FDIC) is clear that outsourcing the model does not outsource the liability.
- A state-by-state map of where the model is used, flagging states like Colorado that require impact assessments or notices.
- Marketing and disclosure review, so that no claim about the model's fairness or accuracy goes beyond what was validated.
A simplified fair lending check on model outputs might look like this:
# Simple disparate impact ratio check (illustrative only)
approval_rate_group_a = approvals_group_a / applications_group_a
approval_rate_group_b = approvals_group_b / applications_group_b
di_ratio = approval_rate_group_b / approval_rate_group_a
# Ratios below 0.80 (the "four-fifths rule") are commonly
# flagged for further investigation. It is a rule of thumb,
# not a safe harbor under ECOA or state lawThe four-fifths rule comes from EEOC (Equal Employment Opportunity Commission) employment guidance and was adapted informally to lending. Use it to screen, not to prove compliance.
🎬 [VIDEO: "How AI Bias Happens in Lending" - https://www.youtube.com/results?search_query=AI+bias+lending+regulation - search for recent explainer content from Brookings, state regulators, or major law firms on algorithmic fair lending, useful as a visual companion to this lesson]
Key Takeaways
- No single US law governs AI in banking. Bank safety and soundness (SR 11-7), consumer protection (ECOA, FCRA), FTC Section 5 and state rules all apply at once.
- Federal fair lending enforcement pulled back in 2025. Disparate impact was deprioritized, the OCC dropped it from exams, and the CFPB proposed removing it from Regulation B. The statutory duty to give specific denial reasons did not change.
- States are now the most active layer. Colorado's AI Act, California's ADMT rules and state attorney general settlements create real exposure.
- A model that passes one regulator's test can still fail another's. Vendor-built AI does not reduce the bank's liability.
- Keep running disparate impact tests and keep the documentation. Federal priorities change between administrations, but models stay in production for years.
Related articles
Recent articles from the blog that build on this lesson.