AI Regulation in Banking, 2026: What Comes Into Force in the EU, the US and the UK, and What Banks Now Owe Their Customers

AI Regulation in Banking, 2026: What Comes Into Force in the EU, the US and the UK, and What Banks Now Owe Their Customers

By Sergei Ponomarev β€’ 2026-11-18

Banks were among the first big users of AI: credit scoring, fraud detection, anti-money-laundering checks, chatbots and now AI agents. The rules for how they may use it are changing fast, and not in one simple direction.

In Europe, two things are happening at once. Some consumer rights are getting stronger: from 20 November 2026, national rules implementing the new Consumer Credit Directive are due to apply, and for consumer credit within its scope, people get rights to human intervention, an explanation and a review when their creditworthiness is assessed by automated processing of personal data. At the same time, parts of the AI Act have been postponed and simplified: its rules for AI credit scoring now apply from 2 December 2027.

In the United States, federal rules are becoming lighter. In April 2026 the bank regulators replaced the model risk guidance that had governed bank models since 2011, and the consumer protection bureau removed the effects test from its fair lending regulation. Some states are moving the other way, with Colorado's rules for consequential AI decisions, including lending, starting in January 2027.

The UK sits in between: no AI-specific rules, existing model risk and consumer duties applied to AI, and a regulator testing AI with firms in a supervised setting.

Here is what is in force, what is coming, and who owes what, both for banks and for the people who borrow from them.

The key dates at a glance

DateWhereWhat happens
17 January 2025EUDORA, the digital operational resilience regulation, applies; AI services bought from cloud and model providers are ICT third-party services
2 February 2025EUAI Act bans on prohibited practices and the AI literacy provision apply
17 April 2026USFed, OCC and FDIC issue revised model risk guidance, replacing SR 11-7
21 July 2026USRegulation B amendments take effect, removing the effects test and stating the CFPB's position on disparate-impact liability under the Equal Credit Opportunity Act
2 August 2026EUAI Act transparency duties for certain AI systems apply
20 November 2026EUNational measures implementing Consumer Credit Directive 2 are due to apply
2 December 2026EUAI Act deadline for watermarking AI-generated content from systems already on the market
1 January 2027US, ColoradoColorado's SB 26-189 takes effect, covering consequential AI decisions including lending
2 December 2027EUAI Act high-risk obligations apply to AI credit scoring of individuals

Europe: more rights for borrowers, more time for AI rules

The Consumer Credit Directive: human review for consumer credit

The new Consumer Credit Directive, (EU) 2023/2225, entered into force in 2023 and replaces the 2008 directive. Member states had to transpose it into national law, and those national measures are due to apply from 20 November 2026.

For consumer credit within its scope, its most important AI provision concerns automated assessments. Where a creditworthiness assessment involves automated processing of personal data, the consumer has the right to request:

  • human intervention by the creditor,
  • a clear and comprehensible explanation of the assessment, including the logic and risks of the automated processing,
  • the chance to express their point of view, and
  • a review of the assessment and of the decision.

Lenders must tell consumers about this right. It builds on the GDPR right not to be subject to purely automated decisions, applied specifically to consumer credit.

Three limits matter. The directive covers consumer credit agreements within its scope, with exclusions; most mortgage lending is governed separately, and business loans are not covered. It generally does not apply to agreements that already existed when the national rules came into force. And the exact rules in each country depend on how that country has transposed it.

What lenders must do now: be able to explain, in plain words, why a model assessed an applicant the way it did; have staff who can genuinely review and change a decision; and update application journeys and notices to tell consumers about the right.

The AI Act: credit scoring is "high-risk", from December 2027

The EU AI Act classifies AI intended to evaluate the creditworthiness of individuals or establish their credit score as high-risk (Annex III, point 5(b)). AI used to detect financial fraud is expressly excluded from that category.

Under the original timetable, high-risk duties would have applied from 2 August 2026. The AI Omnibus, Regulation (EU) 2026/1744, which the European Commission confirms has entered into force, moved stand-alone high-risk systems such as credit scoring to 2 December 2027.

The obligations are not the same for everyone. A bank that develops a credit-scoring system, or puts its name on one, is a provider. A bank that buys and uses a ready-made system is a deployer. A deployer can become a provider if it substantially modifies the system or changes its intended purpose.

ObligationIn plain wordsWho is responsible
Risk management systemidentify and reduce the risks the system poses, throughout its lifeprovider
Data governancetraining, validation and test data that are relevant, representative and examined for biasprovider
Technical documentationrecords that show how the system was built and worksprovider
Loggingthe system records events automatically; logs are keptprovider builds it in; deployer keeps the logs it controls
Instructions for useclear information for the bank staff who use itprovider
Human oversightpeople who can understand, monitor and override the systemprovider designs for it; deployer assigns competent people
Accuracy, robustness, cybersecuritytested performance, protected against manipulationprovider
Conformity assessment and registrationchecks and registration before useprovider
Use in line with instructions, monitoringuse the system as intended and report serious problemsdeployer
Fundamental rights impact assessmentfor deployers assessing creditworthiness, an assessment of the effect on people's rights before first usedeployer

Older systems. December 2027 does not mean every existing scoring model is treated the same way. Under the Act's transitional rules, high-risk systems already on the market before the application date are generally covered only if their design changes significantly afterwards. Banks should check which of their models fall where, rather than assume they are either fully in or fully out.

What already applies. Since February 2025, certain practices, such as social scoring, are banned, and the AI literacy provision applies. After the Omnibus, it requires providers and deployers to take context-appropriate measures to support the development of AI literacy among relevant staff, taking into account their role and how the AI is used. That is not a requirement to run a particular course or certification. Since August 2026, transparency duties apply, for example telling people when they are interacting with an AI system such as a chatbot.

What European supervisors say

The European Banking Authority concluded in late 2025 that it saw no immediate need for new AI-specific guidelines, because many AI Act duties for credit scoring overlap with what banks already owe under capital rules, DORA, the consumer credit and mortgage directives and existing EBA guidelines. Its focus for 2026 and 2027 is consistent supervision and cooperation with the new AI market-surveillance authorities.

The European Central Bank, which supervises the largest euro-area banks, has looked at how they use AI and, according to summaries of its work, is widening its attention from credit scoring and fraud detection to generative AI in its 2026 to 2028 priorities.

DORA already applies, in two layers. For all ICT third-party services, including AI bought from cloud and model providers, banks must manage the risk: keep a register of arrangements, apply their risk policy, and have contracts that meet DORA's basic requirements. Additional requirements, such as stronger contractual terms, exit strategies and closer testing, apply where a service supports critical or important functions. Not every AI tool is a critical supplier; a credit-scoring engine or a core fraud system is more likely to be.

United States: lighter federal rules, active states

New model risk guidance replaces SR 11-7

On 17 April 2026, the Federal Reserve, the OCC and the FDIC issued Revised Guidance on Model Risk Management. The Fed's letter, SR 26-2, says it supersedes and replaces SR 11-7, the 2011 guidance that shaped bank model governance for fifteen years, and SR 21-8 on anti-money-laundering models. The Fed says it is most relevant to banking organisations with more than $30 billion in assets.

The guidance keeps the core disciplines: testing, independent challenge, monitoring, governance and validation of vendor models. A footnote in the guidance itself places generative and agentic AI outside its scope. That is not an exemption from other requirements: banks remain subject to their broader risk management, consumer protection and third-party obligations when they use those technologies. It does mean there is still no specific federal model risk guidance for banks' use of generative AI.

Fair lending: the effects test removed, reasons still required

On 22 April 2026 the Consumer Financial Protection Bureau published a final rule amending Regulation B, effective 21 July 2026. It removes the effects test from Regulation B and sets out the bureau's position that the Equal Credit Opportunity Act does not provide for disparate-impact liability. That is the CFPB's interpretation, not a final ruling by the courts, and the rule is being challenged in court. Intentional discrimination and the deliberate use of proxies for protected characteristics remain prohibited, and other anti-discrimination laws continue to apply.

What has not changed is the duty to tell a rejected applicant the specific reasons for a credit denial. The bureau withdrew its AI-specific guidance on this in 2025, but the requirement sits in the regulation itself. For a bank using a complex model, "the model said no" is still not an acceptable reason.

States: Colorado from January 2027

Colorado replaced its 2024 AI Act with a new law, SB 26-189, effective 1 January 2027. It is a notice-and-rights law, enforced by the state attorney general, covering consequential decisions made with AI, including lending. Banks operating in Colorado should prepare for its notice requirements and follow any legal challenges.

At federal level, a December 2025 executive order, EO 14365, directs federal agencies to challenge state AI laws and calls for a national framework. An executive order does not by itself override state law; whether particular state laws are preempted would depend on legislation or the courts.

United Kingdom: principles, testing and existing duties

The UK has chosen not to write AI-specific rules for financial services. Instead:

  • The Prudential Regulation Authority's model risk principles, SS1/23, apply to AI and machine learning models, including vendor models, at banks and building societies with permission to use internal models for regulatory capital. Other firms are not formally in scope, though many use the principles as a benchmark. The PRA has been discussing with chief risk officers how to apply them to AI.
  • The Financial Conduct Authority relies on existing rules, including the Consumer Duty, and runs AI Live Testing, where firms test AI with regulatory support. Its second cohort, announced in April 2026, includes Barclays, Scottish Widows and UBS, with an evaluation due in early 2027.

What this means for bank customers

In the EU, from 20 November 2026: if you apply for consumer credit covered by the new rules and the assessment is automated, you can ask the lender for human intervention, an explanation you can understand and a review. If you are rejected, use it. Ask in writing and keep a copy. Mortgages and business loans follow different rules.

In the US: if you are denied credit, you are still entitled to the specific reasons. Ask for them if they are vague. If you live in Colorado, new notice rights for AI decisions arrive in January 2027.

Everywhere: in the EU you are entitled to know when you are interacting with an AI system such as a chatbot, and in many places you can ask for a person. On the related question of what you should never give an AI agent, see is it safe to give an AI agent your bank login or card.

What this means for banks

PriorityWhy now
Explainable credit decisions and a real human-review routeEU consumer credit rules from 20 November 2026; US adverse-action reasons remain mandatory
An inventory of every AI system, including vendor AIneeded for AI Act classification, DORA registers and model risk management
Decide, for each credit-scoring system, whether you are provider or deployerthe obligations differ, and the December 2027 deadline still requires documentation, data work and impact assessments
Map which models count as already on the markettransitional rules depend on whether a system's design changes significantly
Tier AI suppliers under DORAbasic duties apply to all ICT services; stronger ones to those supporting critical or important functions
A governance line for generative AI and agentsUS model risk guidance leaves them out of scope; other obligations still apply
Measures to support staff AI literacyan AI Act provision since February 2025, proportionate to role and use

The common thread is that regulators are not inventing a new discipline for AI. They are applying the old ones, model risk, consumer protection, outsourcing and fair treatment, to a new technology, and adding specific customer rights where automated decisions affect people's access to credit. For smaller lenders and fintechs without a compliance team, a practical starting point is AI governance for small companies.

The honest take

What strikes me about this map is that neither side of the Atlantic is moving in a single straight line. Europe is giving consumer borrowers new rights over automated decisions while giving companies more time on the AI Act's heaviest rules. The US federal agencies are loosening model and fair lending rules while some states add new duties. A bank operating in both has to run several regimes at once.

For customers, the practical message is simpler than the legislation. You have more rights than you probably use: to know when AI is involved, to get reasons for a rejection and, for EU consumer credit from this week, to have a person look again. For banks, the rules point to one expectation: if you cannot explain what your model did, you should not be using it to decide who gets credit.

Frequently asked questions

Is AI credit scoring regulated in the EU?

Yes. The AI Act classifies AI used to assess individuals' creditworthiness or set their credit score as high-risk. Under the AI Omnibus, those obligations apply from 2 December 2027, with different duties for providers and deployers. Separately, national rules implementing the Consumer Credit Directive give consumers rights over automated assessments for consumer credit from 20 November 2026.

Can I ask for a human to review an automated loan decision?

In the EU, for consumer credit within the scope of the new Consumer Credit Directive, yes: from 20 November 2026, when a creditworthiness assessment involves automated processing of personal data, you can request human intervention, an explanation and a review. Mortgages and business loans are covered by other rules. In the US there is no general right to human review, but lenders must give specific reasons for a credit denial.

What replaced SR 11-7?

On 17 April 2026 the Federal Reserve, OCC and FDIC issued Revised Guidance on Model Risk Management. The Fed's SR 26-2 supersedes SR 11-7 and SR 21-8. A footnote in the guidance places generative and agentic AI outside its scope, which does not exempt banks from other requirements.

Did the US remove fair lending rules for AI?

No. A Regulation B final rule effective 21 July 2026 removed the effects test and set out the CFPB's position that ECOA does not provide for disparate-impact liability; the rule is being challenged in court. Intentional discrimination and proxies for protected characteristics remain prohibited, and lenders must still give specific reasons for denials.

Does the EU AI Act apply to bank chatbots?

Chatbots are not high-risk in themselves, but transparency duties apply: people must be told when they are interacting with an AI system. If an AI system is used to assess creditworthiness, the high-risk rules would apply to that use.

Does the UK have AI rules for banks?

Not AI-specific rules. The PRA's model risk principles, SS1/23, apply at banks with permission to use internal models for regulatory capital, and the FCA's existing rules, including the Consumer Duty, apply to AI. The FCA also runs AI Live Testing with firms including Barclays, Scottish Widows and UBS.

Sources: EUR-Lex: Consumer Credit Directive (EU) 2023/2225; European Commission: consumer protection in financial services; EUR-Lex: AI Act, Regulation (EU) 2024/1689; EUR-Lex: AI Omnibus, Regulation (EU) 2026/1744; European Commission: AI Omnibus enters into force; EUR-Lex: DORA, Regulation (EU) 2022/2554; Federal Reserve: SR 26-2; Federal Reserve: Revised Guidance on Model Risk Management (attachment); CFPB: Regulation B, version effective 21 July 2026; Husch Blackwell: CFPB finalizes Regulation B overhaul; Colorado General Assembly: SB26-189; Bank of England: SS1/23 model risk management principles; FCA: second cohort of AI Live Testing; Bird & Bird: AI and financial institutions.

Share this article: