Machine learning in banking rarely makes headlines the way chatbots do. Bank of America’s Erica assistant passed 3.2 billion lifetime interactions by March 10, 2026, with 20.6 million users generating nearly 700 million interactions last year. Behind that visible layer sits a much larger set of models scoring loan applications, flagging suspicious transactions, and predicting which customers are about to leave, quietly and without a chat window.
You can read more about how frontline staff use AI in this guide to AI in banking customer service. For a broader view of how banks apply AI beyond retail products, see artificial intelligence in commercial banking.
Credit Scoring and Underwriting with Machine Learning in Banking
Credit scoring is where this practice started, decades before anyone called it AI. Logistic regression and gradient boosting models rank borrowers by default risk using income, repayment history, and increasingly, alternative data such as telecom payments and utility bills. Lenders use these scores to approve applications in seconds instead of days, without a loan officer rereading every file by hand.
Underwriting teams also use these models to reach borrowers with thin credit files, people who pay rent and phone bills on time but have no traditional credit history. The model finds the signal in that alternative data instead of rejecting the application outright, widening the pool of approved borrowers without loosening the risk standards a bank cares about.
In EY and the Institute of International Finance’s February 24, 2026 survey of 101 banks across 31 countries, 33% of chief risk officers said they now use AI to support credit and market risk modelling, putting machine learning in banking inside mainstream risk practice rather than at its edge. Even so, 72% of chief risk officers describe AI adoption inside the risk function as still in an early stage, credit scoring included.
Fraud Detection and Transaction Monitoring
Fraud detection is the clearest use case for machine learning in banking, because fraud patterns shift faster than any analyst can track by hand. Models trained on historical transactions learn the combinations of location, amount, and timing that mark a transaction as suspicious, then score every new transaction against that pattern in real time, before the payment ever settles.
Adoption of these fraud tools is now mainstream rather than experimental. In the same EY and IIF survey, 61% of chief risk officers reported active AI deployment in fraud and financial crime detection, the highest adoption rate of any risk function surveyed that year.
Cybersecurity remains banks’ single largest near-term risk at 86% in that survey, and 41% of chief risk officers already use AI to monitor cyber and operational risk alongside fraud. A deeper breakdown of how these systems work is in artificial intelligence in fraud detection.
None of this replaces human judgment. Analysts still review flagged transactions, tune thresholds, and decide which alerts warrant a call to the customer. These models narrow a flood of transactions down to the handful that need a person’s attention, which is a different job than automating fraud decisions outright.
Customer Churn and Propensity Modeling in Banking
Machine learning in banking also predicts which customers are about to leave. Churn models score every account on transaction frequency, product usage, and complaint history, then rank accounts by how likely they are to close within the next few months. Retention teams use that ranked list to prioritize outreach instead of guessing which relationship manager to call first.
Propensity models work in the other direction, predicting which existing customers are likely to want a specific product next, a mortgage, a savings account, an upgraded card. Banks feed these scores into automated follow-up messages so the right offer reaches the right customer without a manual campaign build each time.
Personalization is still more promise than delivery industry wide. J.D. Power’s 2026 U.S. Financial Health Support and Advice Study, released May 27, 2026 and based on 29,855 banking customers, found just 20% say their provider always personalizes what they receive, though satisfaction scores rise by 238 points among the customers who feel understood. Machine learning in banking, applied to churn and propensity scoring, is how banks close that gap without guesswork.
Risk-Based Pricing for Loans and Deposits
Pricing is another quiet application of machine learning in banking. Instead of one interest rate for every borrower in a risk tier, models adjust pricing at the individual level based on predicted default probability, collateral value, and relationship history. A borrower with strong data behind them can get a better rate than the tier average would otherwise suggest.
The same logic extends to deposit pricing and fee structures, where models balance competitiveness against margin using segments far more granular than a rate sheet allows, so these pricing models sit inside a bank’s risk management function, not just its product team.
These pricing models still need the same guardrails as any other statistical model: back testing, fair lending review, and a named owner accountable for outcomes. Speed without oversight is not a feature.
Anti-Money Laundering Monitoring and Model Risk
Machine learning in banking has powered anti-money laundering monitoring for years, one of RegTech’s oldest use cases. Models score transaction patterns against known laundering typologies, structuring, rapid movement between accounts, transfers sitting just under reporting thresholds, and surface the ones worth a human investigator’s time.
These are exactly the kind of models supervisors mean when they talk about model risk. On April 17, 2026, the Federal Reserve, the OCC, and the FDIC jointly issued SR 26-2, which supersedes and replaces two earlier letters, SR 11-7 and SR 21-8. It applies most directly to banking organizations with over $30 billion in total assets, the tier most exposed to these models at scale.
SR 26-2 draws a clean line. Generative and agentic AI models are explicitly outside its scope because they are novel and rapidly evolving. Its principles instead apply to traditional statistical and quantitative models and non-generative, non-agentic AI models, exactly the scorecards, fraud models, and AML monitors described above. Machine learning in banking is precisely what supervisory model risk guidance already covers.
The EU AI Act and Classical Machine Learning Models
European law treats machine learning in banking unevenly depending on what the model decides. Under Annex III of the EU AI Act, systems used to evaluate creditworthiness or establish a credit score are classified high risk, with one exception carved out for systems used to detect financial fraud.
That exception matters. A credit scorecard sits inside the high-risk category and carries the Act’s full obligations: documentation, human oversight, and testing. A fraud detection model, built the same way on similar data, sits outside it, proof that classical models fall into several regulatory buckets, not one.
The timeline is layered. Prohibited practices and AI literacy obligations applied from February 2, 2025. Governance rules and the general penalty regime followed on August 2, 2025. General applicability, Article 50 transparency duties, and the Article 101 fining power arrive August 2, 2026.
The Annex III obligations for credit scoring, the kind of model described earlier, were pushed to December 2, 2027 by the Digital Omnibus, Regulation (EU) 2026/1744, in force since July 27, 2026. The full detail is in the EU AI Act framework.
Data Quality: The Limiting Factor for These Models
Every machine learning in banking model depends on one resource more than any algorithm choice: clean and available data. Deloitte’s 2026 banking and capital markets outlook, published October 30, 2025, reports that more than 90% of data users inside banks say the data they need is often unavailable or takes too long to retrieve, a self-reported figure from Deloitte’s own survey.
That gap shows up in results. According to Evident’s 2025 analysis, cited in the same Deloitte outlook, only 4 of 50 banks studied reported realized ROI from their AI use cases. Machine learning in banking cannot outperform the data it is trained on, and most banks are still fixing the plumbing before they scale the models.
The Financial Stability Board’s October 10, 2025 report on AI adoption names model risk, data quality, and governance, the exact ingredients these models depend on, as one of four vulnerabilities regulators are watching, alongside third-party dependencies, market correlations, and cyber risk.
The EY and IIF survey backs this up from inside the bank. 80% of chief risk officers name data quality and availability as the primary barrier to wider AI use in risk management, and 79% call data and AI skills essential over the next few years. Data, not model design, is the bottleneck most banks are working through right now.
Explainability and Oversight of Machine Learning Models in Banking
None of this works without someone who can explain why a model made a decision. That expectation applies here as much as to any other automated system. Pedro Machado of the European Central Bank said in a February 24, 2026 speech that if a bank cannot explain why a model behaves the way it does, in terms that matter for decision making, then it does not control that model.
Machado also reported that more than 85% of large banks under European supervision already use AI in some form, spanning IT operations, legal and document review, and front-line applications. The governance gaps he flagged, unclear accountability, thin senior management oversight, and weak challenge from risk, compliance, and audit teams, are the same gaps SR 26-2 is written to close, and the reason these models need governance as much as they need data.
Machine learning in banking earns its keep only when a named owner can answer for a model, when validation teams test it before and after deployment, and when audit can trace a decision back to the data that produced it. That discipline, more than the algorithm, separates a well-governed model from a liability.
Where to Start with Machine Learning in Banking
Pick one machine learning in banking model type from this list, credit scoring, fraud monitoring, churn prediction, or pricing, and audit how it is governed today. Ask who owns it, when it was last validated, and whether SR 26-2 or the EU AI Act changes what is required of it.
A small pilot beats a large plan. Start with the process causing the most manual work right now, and measure the result before expanding to a second model.
If you want help scoping a pilot, explore our AI services. Our team works with banks moving these projects from spreadsheet to production every week. Reach out and we can help you find the fastest path to a working pilot.
