What is AI in banking, and how is it used?

AI in banking is used mostly for back-office work such as process automation and software engineering, not for customer-facing decisions. The Cambridge Centre for Alternative Finance's 2026 report found process automation at 79% and software engineering at 75% of surveyed financial services firms, with four of the top five use cases overall being back-office functions. Customer support leads front-office applications at 74% at pilot stage or beyond, fraud detection sits at 58%, and credit risk modelling at 54%. The pattern is that AI in banking concentrates where no customer decision and no regulatory disclosure depends on the output.

What are the benefits of AI in banking?

The documented advantages of AI in banking concentrate in cost and cycle time rather than revenue. Back-office automation reduces handling time on work that was already measured, fraud and monitoring models improve detection on a supervisory path that already exists, and document retrieval cuts the time staff spend finding the right policy or file. The benefits of AI in banking that survive scrutiny are the ones attached to a number finance already reports — which matters, because the Cambridge research found 40% of firms reporting increased profitability from AI against 43% reporting no change at all. What separates those two groups is mostly whether anyone agreed what to measure before the work started.

What are some examples of AI in banking?

Ordered by how commonly they appear in the Cambridge data, the clearest AI examples in banking are process automation in internal operations (79% of surveyed firms), software engineering support (75%), customer support handling and drafting (74% at pilot stage or beyond), fraud detection and transaction monitoring (58%), and credit risk modelling (54%). Less established but widely discussed are agentic workflows that act without a person approving each step, at 52% adoption and currently outside the scope of US model risk guidance.

Does the new model risk guidance cover AI?

SR 26-2 covers traditional models and basic AI applications but explicitly excludes generative and agentic AI. The guidance was issued on 17 April 2026 and replaced SR 11-7, which had governed US bank models since 2011. The exclusion is stated by both agencies: the OCC’s Spring 2026 Semiannual Risk Perspective puts generative and agentic AI outside the revised guidance’s scope, and the Federal Reserve’s Vice Chair for Supervision said the same on 1 May 2026. Those systems are expected to be governed by other risk management and governance practices, which have not yet been specified, and the agencies have signalled a request for information on the subject.

Can a bank use AI to decline a loan application?

Only if it can state the specific reasons accurately. The Equal Credit Opportunity Act requires a statement of specific reasons for adverse action, and CFPB Circular 2022-03 confirmed that using a complex algorithm does not reduce that obligation — a creditor's lack of understanding of its own model is not a defence. In practice this pushes AI toward ranking and triage, with the declining decision made on a basis that can be written on a notice.

Why do banking AI projects stall?

Usually on explainability or on ownership of validation rather than on model performance. The Cambridge research found 78% of regulators treating explainability as critical or important against 50% of industry having adopted explainable AI methods, and roughly two-thirds of industry respondents not monitoring for bias. A model that performs well and cannot be explained is not a deployable system in this sector.

Does the $30 billion threshold in SR 26-2 mean smaller banks are exempt?

No. The $30 billion figure marks where SR 26-2 is most relevant, not an exemption for smaller banks. The OCC’s version of the same guidance, Bulletin 2026-13, states that it is applicable to all community banks subject to the limitations set out in it, and that it may also apply to smaller institutions with significant model risk exposure. What smaller banks get is proportionality — the approach is explicitly risk-based and tailored. Obligations sitting outside model risk guidance do not scale with size either: consumer protection requirements apply regardless, and for most smaller banks the AI in use arrives embedded in a third-party product, where the risk is still the bank’s.

How accurate are published AI in banking statistics?

Treat widely circulated figures carefully. Claims that upwards of 90% of banks now use AI, along with large market-size and ROI multiples, appear mainly on aggregator pages citing one another rather than any identifiable study. Material from the Federal Reserve, the OCC, the CFPB and the Cambridge Centre for Alternative Finance is traceable, and consistently more sober.

Do the UK rules treat banking AI the same way?

No. The PRA’s SS1/23 applies to all models “regardless of technology”, while US guidance excludes generative and agentic AI. SS1/23 goes further than silence on the subject: one of its sub-principles deals specifically with identifying and managing the risks of using artificial intelligence in modelling techniques. The practical effect for a group operating in both markets is that a single generative AI deployment can fall outside the US framework and inside the UK one. SS1/23’s reach is narrower than it first looks, though — it applies to UK firms holding internal model approval for credit, market, or counterparty credit risk rather than to every UK bank.

Where should a bank start with AI?

With internal operations, where the governance burden is lowest and the data and integration work carries over. Fraud detection and transaction monitoring are a reasonable second, because the supervisory path for predictive models of that type is established. Customer-facing applications work when designed so a person reviews the output. Credit decisioning and agentic workflows belong last, once a validation practice exists and has been tested on lower-stakes work.