Banks deploy AI in lending at the steps that used to mean an underwriter reading a file: document intake and verification, income and asset calculation, financial-statement spreading, credit-memo drafting, decisioning on bureau and cash-flow data, the adverse-action reasons behind a decline, and monitoring the loan afterwards. Of the 140 banks profiled on this site, 72 record one of those steps on AI and 43 run it in production. The results are in the banks' own numbers: Citi's new AI models in US Cards raised approval rates by 100 basis points; Synchrony's PRISM decides a credit application in six seconds or less and approved nearly 50,000 customers in 2025 through cash-flow underwriting who would otherwise have been declined; TD Bank's agents cut mortgage pre-adjudication from an average of 15 hours to under three minutes; First Hawaiian's Zest AI model took automated card decisions from 4% to 55% of applications; SouthState has about 90% of credit spreads done by AI; Deutsche Bank cut Mittelstand time-to-cash by up to 95%. Regulators allow it under the rules that govern every credit decision: in the US a lender must give the specific principal reasons for adverse action whatever the model, must test it for discrimination and must validate it as a model; in the EU, credit scoring of natural persons is a high-risk use under the AI Act from December 2, 2027. The tables below are computed from those records, and every number links to its source.
DATA AS OF OCT 13, 2026 · 140 banks · 238 documents · 49 vendors · 4,720 sources
What is AI credit underwriting?
AI credit underwriting means models doing the reading, calculating, scoring and drafting inside a lending decision: document intake and verification, income and asset calculation, financial-statement spreading, credit-memo drafting, decisioning on bureau and cash-flow data, adverse-action reason attribution, and portfolio monitoring afterwards. The lender keeps the decision, the notice and the fair-lending duty; 72 of the 140 banks profiled here record one of those steps on AI, 43 in production.
A credit decision has the same four parts whoever makes it: establish who the borrower is and what they earn and own, assess the ability and willingness to repay, decide and price, and tell the applicant why if the answer is no. Banks have scored consumer credit with statistical models since the 1960s and the OCC wrote its first examination guidance on credit-scoring models in 1997, so ‘AI underwriting’ is a change of degree in three places. The models got more complex: gradient-boosted trees and, at the fintechs, transformers trained on bank-transaction data in place of logistic regression on bureau attributes. The data got wider: cash-flow data from connected accounts, with consent, and documents read by vision models instead of typed in. And since 2025 generative and agentic AI has moved into the parts of commercial and mortgage lending that scoring never reached: reading a 200-page loan package, spreading the financials, drafting the credit memo, writing the underwriter's summary.
The record on this site shows all three at once. Supervised decisioning has run in production at card and consumer lenders for years: Synchrony's PRISM has approved more than 180 million accounts since 2018, Bread Financial has over 200 machine-learning models across credit, collections, marketing and fraud, FNBO has originated personal loans through Upstart since 2019, First Hawaiian went live on Zest AI in consumer lending. Cash-flow underwriting is the data story: Synchrony, RBC and Citizens describe it. The 2025–26 wave is generative and agentic: TD Bank's mortgage pre-adjudication agents, DBS's chain of roughly 70 to 80 agents assembling corporate credit memos, BMO's Aura commercial credit agent, SouthState's in-house small language model spreading financials, First Horizon and RBC spreading statements and starting credit memos with AI. Across the whole use case 84 of the 140 banks here record some credit-underwriting activity; this page keeps the 72 whose record describes AI in the lending process itself, and sets aside the rows about lending to AI companies, insurance underwriting and plain digital origination with no model claim.
Which banks use AI in credit underwriting and lending?
43 banks run AI in a lending step in production on the record, led by Citi, Synchrony, TD Bank, RBC, BMO, KeyBank, SouthState, First Hawaiian, Zions, Deutsche Bank, ING, Lloyds, BNP Paribas, DBS and OCBC; 16 are rolling out, 8 are in pilot and 5 have announced plans. Consumer and card decisioning has run on supervised models for years; the 2025 and 2026 wave is generative and agentic AI in mortgage pre-adjudication and in commercial credit, where the file is long and the saving per loan is measured in hours.
The table lists every bank whose profile records an AI activity in origination, verification, underwriting, decisioning or portfolio monitoring, strongest status first and largest bank first within a status. “In production” means the bank or a named vendor says it is live; the date is the earliest sourced event on the bank's timeline for the activity. Where the record is a vendor's case study rather than the bank's own statement (Zest AI on First Hawaiian, Abrigo on Old National, Upstart on FNBO) the detail says so.
| Bank | What the record shows | Status | Date | Source |
|---|---|---|---|---|
| Truist | Wholesale underwriting and adviser analytics. AI and predictive analytics to speed underwriting, scale lead generation in payments and wealth, and improve adviser productivity. | In production | Apr 17, 2026 | The Motley Fool (Yahoo Finance), Apr 17, 2026 ↗ +1 |
| BMO | ARM automated review triage and portfolio monitoring. Embedded guidance on annual client reviews and automated portfolio monitoring cut cycle time by about 15%; prompt-engineered memo tools for bankers. | In production | Mar 26, 2026 | BMO Financial Group (investor relations), Mar 26, 2026 ↗ |
| M&T | Credit monitoring with RDC.AI. Agentic and predictive intelligence on commercial portfolios; earlier issue detection. | In production | Aug 19, 2026 | M&T Bank, Aug 19, 2026 ↗ |
| KeyBank | AI research support for commercial loan approvals. Analyst research time fell from 60% to 25% according to the Chief Data and Analytics Officer; public-company information is aggregated and summarised with gen-AI tools. | In production | Aug 14, 2025 | Google Cloud, Nov 16, 2025 ↗ +1 |
| Synchrony | PRISM across collections, identity and account management. American Banker reports PRISM is used in lending-related decisions across collections, identity, account management, modelling and governance; more than 20 cross-functional teams of credit and technology staff maintain it, and decisions take six seconds or less. The CEO says it has approved more than 180 million accounts since January 2018. | In production | Mar 31, 2026 | American Banker, Jun 1, 2026 ↗ +1 |
| Synchrony | PRISM credit decisioning with cashflow data. A proprietary model analysing about 9,000 data points returns decisions in seconds; cashflow analysis (with consent) lifted approvals by nearly 50,000 in 2025. The sources do not describe it as AI. | In production | Jun 1, 2026 | American Banker, Jun 1, 2026 ↗ +1 |
| Santander | Machine-learning auto credit risk on FICO Platform. Approve/decline decisions, adverse-action reasons and probability-of-default for pricing in the US auto business, with faster model development and governance. | In production | May 7, 2025 | FICO (Business Wire), May 7, 2025 ↗ |
| Western Alliance | Credit agreement review and lending lifecycle redesign. AI review and verification of credit agreements; the lending lifecycle redesign is credited with a 25% cut in cycle time (internal projection). | In production | May 12, 2026 | Western Alliance Bancorporation (SEC filings), May 12, 2026 ↗ |
| Western Alliance | AI as an underwriting theme in innovation banking. Lending to AI and enterprise-software companies with a stated view on AI disruption. | In production | Jan 2, 2026 | Western Alliance Bancorporation, May 11, 2026 ↗ +1 |
| Zions | AI in credit review, appraisal review and financial-statement spreading. AI widens the population of deals the credit-review function examines; people review what the tools flag. | In production | Jan 20, 2026 | Yahoo Finance, Apr 21, 2026 ↗ +1 |
| Zions | Unstructured-document processing for onboarding and origination. AI reads customer, internal and supplier documents to find what is missing. | In production | Mar 10, 2026 | American Banker, Mar 10, 2026 ↗ |
| First Horizon | AI file-structure organisation for loan-application documents. Saves about two to three minutes per document, hundreds of staff hours within three months of use. | In production | Oct 1, 2026 | American Banker, Oct 1, 2026 ↗ |
| Wintrust | AI credit adjudication at Wintrust Specialty Finance. The equipment lender, a division of Beverly Bank and Trust, uses AI to determine borrowers' creditworthiness and is extending AI to decision-making, analytics and probability-of-default modelling, as reported by Equipment Finance News. | In production | Aug 6, 2026 | Equipment Finance News (first published by sister title FinAi News), Aug 6, 2026 ↗ |
| Old National | AI-assisted commercial loan review (Abrigo Loan Review Assistant). Vendor-reported result: loan-reviewer efficiency up about 31% on average, with review capacity expanded while reviewers stay accountable for conclusions. | In production | Jul 7, 2026 | Abrigo, Jul 7, 2026 ↗ |
| SouthState | Automated credit spreading. About 90% of credit spreads performed by AI by mid-2026, up 57% over the twelve months to March; the in-house small language model is used in credit operations. | In production | Apr 23, 2026 | SouthState Bank Corporation (SEC filings), Jul 23, 2026 ↗ +2 |
| CIBC | Faster, more consistent credit decisions. Management said on the Q1 2026 call that AI accelerates and improves the consistency of credit decisions; the investor deck lists 'optimizing and accelerating credit decisions'. No model types or US scope are given. | In production | Feb 26, 2026 | Investing.com, Feb 26, 2026 ↗ +1 |
| CIBC | DocuMind document intelligence for lending. Generative-AI document processing for faster lending decisions and fraud detection: 63,000 documents a month and about 16,000 hours saved each quarter to date. | In production | Jul 31, 2026 | CIBC via CNW (Yahoo Finance), Jul 31, 2026 ↗ +1 |
| Frost | Digital account opening and lending on Blend. Automated review cut manual handling to 3%; lending on Blend since 2020. | In production | Aug 18, 2026 | Business Wire via FT Markets, Aug 18, 2026 ↗ |
| FNB | Credit decisioning and CECL models. The 10-K says FNB relies heavily on models, AI and generative AI in credit decisioning; the investor day lists 16 CECL and stress-testing models; the data-science team runs strategic decisioning systems for credit and risk. | In production | Sep 18, 2025 | F.N.B. Corporation (SEC filings), Feb 24, 2026 ↗ +2 |
| SoFi | Oscilar AI risk decisioning for credit strategy. SoFi selected Oscilar to unify credit underwriting, collections and fraud decisioning; Oscilar reports 50% faster deployment of credit-risk strategies and 30%+ faster decisioning, and Noto says SoFi is testing new credit-model features on enhanced data pools. | In production | Apr 29, 2026 | Oscilar (PR Newswire), Jun 3, 2026 ↗ +1 |
| SoFi | Personal Loan Doc Coach. AI-powered document coach that streamlines the personal-loan application. | In production | Apr 29, 2026 | SEC (SoFi Technologies 8-K), Apr 29, 2026 ↗ |
| Barclays US | Best Egg digital personal-loan origination. Direct-to-consumer platform with $40 billion originated since 2013. | In production | Oct 28, 2025 | American Banker, Oct 28, 2025 ↗ |
| Deutsche Bank | AI-accelerated corporate credit process. Corporate Bank credit re-engineered with AI; Mittelstand time-to-cash cut by up to 95%. | In production | Apr 29, 2026 | Deutsche Bank, Apr 29, 2026 ↗ +1 |
| Banc of California | AI-enabled credit analysis and portfolio monitoring. Listed in the 2026 proxy letter with fraud prevention and operational workflows. | In production | Mar 26, 2026 | SEC EDGAR, Mar 26, 2026 ↗ |
| Texas Capital | Document intelligence for commercial loan packages. AWS says the bank already runs a production capability that extracts obligations, servicing requirements and covenant terms from loan packages of 200 or more pages that analysts read by hand. | In production | Oct 5, 2026 | Amazon Web Services (AWS Industries blog, according to AWS), Oct 5, 2026 ↗ |
| Glacier | Automated commercial loan processing system extended to Guaranty. Chesler told analysts in January 2026 that Guaranty would get Glacier's automated commercial loan processing system after the February conversion; the conversion was completed in the first quarter. Management did not describe the system as AI or machine learning. | In production | Jan 23, 2026 | earnings call transcript (The Motley Fool), Jan 23, 2026 ↗ +1 |
| Eastern Bank | Express Business Loan automated decisioning. Five-minute application and decisions 'often instantly' for small-business loans up to $200,000; built in-house by Eastern Labs. | In production | Sep 1, 2026 | Eastern Bank, Sep 1, 2026 ↗ +1 |
| Axos Bank | AI in commercial credit underwriting and portfolio management. Manual, repetitive underwriting and monitoring tasks automated. | In production | Jan 29, 2026 | Yahoo Finance, Jan 29, 2026 ↗ |
| Axos Bank | Covenant extraction from commercial loan agreements. An agent reads the agreement, presents covenants to an attorney, learns from corrections and uploads them to the commercial originations system for tracking. | In production | Jul 30, 2025 | The Motley Fool (Yahoo Finance), Jul 30, 2025 ↗ |
| City National Florida | Polly product and pricing engine for residential lending. Rules, eligibility and pricing adjusted quickly across portfolio products. | In production | Jul 28, 2026 | HousingWire, Jul 28, 2026 ↗ |
| United Community | SBA lending platform with vendor-described AI underwriting. Newgen's SBA loan solution handles origination, renewals and modifications; Newgen describes AI-powered underwriting capabilities, and United has not said which AI features it uses. | In production | May 14, 2025 | Fintech Futures, May 14, 2025 ↗ |
| First Hawaiian | AI-based automated underwriting in consumer lending (Zest AI). Custom credit model plus process redesign; expanded approvals in previously unscorable/credit-invisible populations; Celent 2025 Model Bank award. Vendor Zest AI reports the credit-card model lifted automated decisions from 4% to 55% of applications. | In production | Jun 5, 2025 | Zest AI (vendor page), Aug 8, 2025 ↗ +1 |
| Seacoast Bank | Generative-AI loan-review documentation (Abrigo Loan Review Assistant). Drafts borrower, facility, collateral, underwriting and cash-flow narratives from approval packages inside DiCOM Loan Review; reviewers keep every credit and risk-rating judgement, check each output and use citations to source documents. Clean line sheets fell from 8 to 10 hours to 4 to 6 hours. | In production | Oct 8, 2026 | Abrigo, Oct 8, 2026 ↗ |
| Banner Bank | New loan and deposit origination system. Part of digital-platform and AI-capability investment; implemented during 2025 and fully live by year-end. | In production | Feb 25, 2026 | StockStory (Yahoo Finance), Apr 23, 2026 ↗ +2 |
| Bread Financial | Machine-learning credit models and possible deep learning. Over 200 machine learning models are embedded in the business; the 10-K says Bread is evaluating the potential use of deep learning and AI tools in credit risk management. | In production | Jan 29, 2026 | Investing.com, Feb 3, 2026 ↗ +1 |
| Hilltop | Digital asset validation inside automated underwriting. PointServ lets borrowers connect financial accounts to the application for instant validation of assets; HousingWire reports automated underwriting approvals up more than 15% after integration. Described as automation, not AI. | In production | Feb 3, 2025 | HousingWire, Feb 3, 2025 ↗ |
| Park National | Automated home valuation. Park's home lending team implemented an automated valuation tool in 2024 that replaced a manual process and speeds closings on home equity loans; the annual report does not call it AI or name the vendor. | In production | Mar 10, 2025 | Park National Corporation (SEC filing, ARS), Mar 10, 2025 ↗ |
| National Bank Holdings | Automated SBA working-capital loans on 2UniFi. The first 2UniFi phase launched in July 2025 with a fully automated SBA working-capital loan offering; NBH calls it automated, not AI, and the CEO said in July 2026 that many applications did not fit the bank's credit risk profile. | In production | Jan 28, 2026 | earnings call transcript (The Motley Fool), Jan 28, 2026 ↗ +2 |
| RBC | Newcomer credit limit increases. Machine learning lets RBC offer roughly 40% more credit limit increases to Canadians with under two years of credit history than conventional methods would allow, about C$300 million of extra credit (October 2026). | In production | Oct 6, 2026 | Royal Bank of Canada (CNW via Financial Post), Oct 6, 2026 ↗ |
| HSBC | Gen-AI credit analysis write-ups. Uses trusted internal and external data to reduce time to complete the credit application process. | In production | Oct 5, 2026 | HSBC Holdings plc, Oct 5, 2026 ↗ |
| Scotiabank | Credit-card pre-approval and mortgage verification. AI-powered credit-card pre-approval decisioning and offer personalisation appear in the Q3 2026 deck; in September 2026 management said AI cut mortgage-underwriting verification time by almost 70%. No model, data or fairness-testing detail is public. | In production | Aug 25, 2026 | Scotiabank (The Bank of Nova Scotia), Aug 25, 2026 ↗ +1 |
| TD Bank Group | Agentic pre-adjudication for mortgages and HELOCs. Autonomous agents classify documents, calculate income, check policy and consent and write an underwriter summary; the Q3 2026 call said the first RESL agentic capability was fully scaled and speeding decisions, with the same approach planned for small business and auto finance. | In production | May 21, 2026 | TD Bank Group, May 21, 2026 ↗ +2 |
| RBC | ATOM-based retail credit adjudication. ATOM supports credit adjudication using transaction histories and non-traditional data, including extending credit to clients with less credit history such as newcomers; in August 2026 RBC cited improved approval rates, faster decisions and increased client acquisition. | In production | Jul 29, 2025 | Royal Bank of Canada (CNW via Yahoo Finance), Jul 29, 2025 ↗ +1 |
| ING | Instant credit decisions and AI-assisted mortgages. 65% instant credit decisions for consumer loans in the Netherlands; machine-learning scorecards for mortgage credit decisions in Australia; online mortgages in Germany with a binding decision in under 30 minutes. | In production | Oct 30, 2025 | ING Groep N.V., Mar 18, 2026 ↗ +3 |
| Santander | Credit-card eligibility at onboarding (Spain). Machine learning and real-time data assess at onboarding whether a new customer may be eligible for a credit card from day one. | In production | Jun 22, 2026 | Banco Santander, Jun 22, 2026 ↗ |
| BBVA | Credit Analysis Pro GPT. Risk analysts' GPT that extracts and compares data from annual reports, ESG disclosures and media; analysts monitor the output for errors. | In production | Nov 20, 2024 | BBVA, Nov 20, 2024 ↗ +1 |
| OCBC | AI copilot for corporate relationship managers. Generates ESG, financial and industry analysis from lengthy unstructured documents during credit assessment, cutting preparation time by up to 30%. | In production | Oct 10, 2025 | Oversea-Chinese Banking Corporation (via HKMA Banking Disclosure Repository), Mar 10, 2026 ↗ |
| Lloyds | GenAI application for commercial real estate lending. The first generative AI application in Business and Commercial Banking, expediting the tenancy-schedule step of commercial real estate lending. | In production | Feb 13, 2026 | Lloyds Banking Group (SEC filing), Feb 13, 2026 ↗ |
| Lloyds | Mortgage income verification algorithm. A model run on the Vertex AI platform cuts the income-verification step of mortgage applications from days to seconds. | In production | Apr 9, 2025 | Google Cloud (PR Newswire), Apr 9, 2025 ↗ |
| BNP Paribas | Document AI for French mortgage files. Intelligent document processing on the in-house Doc Factory software checks purchase agreements, property tax notices, pay slips and tax notices against the advisor’s credit-planning tool and checks tax-notice authenticity; live since November 2024, covering about 80% of real estate files and saving about 30 minutes a file, with an electronic offer for the simplest 20% planned for early 2026. | In production | Mar 5, 2025 | BNP Paribas, Mar 5, 2025 ↗ |
| Société Générale | Nacre overdraft-overrun decisions and Stella credit analysis. Nacre, a gradient-boosting model in SG retail banking, processes over 40,000 transactions a day, with more than 80% handled automatically and under 5% wrongly accepted against 22% in the old system; Stella structures data from annual reports and cuts credit-analysis time by about 50%, from 8-34 hours to 2-20 hours. | In production | Feb 1, 2025 | Societe Generale, Feb 1, 2025 ↗ |
| PNC | Commercial loan processing (Midland). A Big Five target: AI applied to loan-servicing and processing work in the Midland commercial-loan business. | Rolling out | Feb 10, 2026 | Investing.com, Feb 10, 2026 ↗ +1 |
| TD | Agentic mortgage funding and discharge. Two new group capabilities (13 October 2026) coordinate multi-step document verification and draft a case summary or discharge statement for colleague review; Chun said the pre-adjudication agent is transferable to US lines. Group real-estate lending is Canadian. | Rolling out | Sep 9, 2026 | TD Bank Group (Canada Newswire via Yahoo Finance), Oct 13, 2026 ↗ +1 |
| TD | Agentic lending pre-adjudication. Application summary memos for real-estate secured lending underwriters. | Rolling out | May 21, 2026 | TD Bank Group, May 21, 2026 ↗ +1 |
| Huntington | Gen-AI SBA loan origination (Casca). Huntington is a design partner and investor; Casca software automates much of the small-loan process (Live Oak targets approval in a day or two). Huntington said in August 2025 it would deploy it in 2026, and the credit decision stays with its existing underwriting models and loan officers. | Rolling out | Aug 19, 2025 | American Banker, Aug 19, 2025 ↗ +1 |
| BMO | Aura commercial credit agent. Drafts credit-assessment commentary from approved internal systems; underwriters review. | Rolling out | Aug 6, 2026 | BMO AI Hub, Aug 6, 2026 ↗ |
| Citizens | AI lending onboarding platform. Built for student lending, parts reused for home equity and planned for credit cards, so new scores such as cash-flow underwriting deploy faster. | Rolling out | Jul 17, 2025 | American Banker, Dec 1, 2025 ↗ |
| City National | AI commercial credit and auto-adjudication (RBC). McKay told the Scotiabank summit that RBC is 'just starting to deploy' AI for commercial auto-adjudication and credit-submission preparation, built and demonstrated to the board; he places it in RBC commercial banking and does not name City National. | Rolling out | Sep 9, 2026 | Investing.com, Sep 9, 2026 ↗ |
| Zions | AI in credit examination and treasury-management operations. Simmons names both as areas with AI projects expected to have a major impact over three to four years. | Rolling out | Sep 10, 2025 | Investing.com (conference transcript), Sep 10, 2025 ↗ +1 |
| First Horizon | Automated CRE loan-origination documentation platform. Winning entry from a 2025 employee AI-ideas contest; built for commercial real estate loan origination and being extended to the bank's entire commercial line of business. | Rolling out | Oct 1, 2026 | American Banker, Oct 1, 2026 ↗ |
| First Horizon | Financial-statement spreading and credit-memo drafting. AI spreads customers' financial statements and tax returns and starts generating a credit memo; human verification remains in the loop. | Rolling out | Sep 15, 2026 | Investing.com, Sep 15, 2026 ↗ |
| Deutsche Bank | AI instant mortgage decisions. Target: 60% decrease in mortgage handling time. | Rolling out | Nov 17, 2025 | Deutsche Bank, Nov 17, 2025 ↗ |
| FNBO | Upstart AI personal loans (nationwide expansion). Personal loans originated through Upstart's AI lending platform, with 78% approved instantly as of December 2021; the bank said it would extend the platform to affinity partners. Whether it still runs in 2026 is not public. | Rolling out | Dec 8, 2021 | FNBO and Upstart (Business Wire), Dec 8, 2021 ↗ |
| Eastern Bank | New home-equity origination platform. Being installed in 2026 to replace a legacy platform whose capacity constraints made home-equity growth 'episodic'; the stated aims are speed, scalability and consistency, and no AI feature is claimed. | Rolling out | Apr 24, 2025 | The Motley Fool (Yahoo Finance), Apr 24, 2026 ↗ |
| Sallie Mae | Underwriting and credit-administration improvements. Data-led enhancements to the core private student-loan process under the new technology organisation. | Rolling out | Feb 2, 2026 | Sallie Mae, Apr 27, 2026 ↗ +1 |
| Arvest | Digitised small-business lending (Baker Hill UN/FY). Automated workflows and streamlined underwriting across the lending lifecycle. | Rolling out | Feb 9, 2026 | Baker Hill (PR Newswire), Feb 9, 2026 ↗ |
| Rockland Trust | Commercial loan origination redesign (nCino). A 2026 project to redesign the commercial origination process on nCino; no AI component disclosed. | Rolling out | Mar 9, 2026 | Independent Bank Corp. (SEC filings), Mar 9, 2026 ↗ +1 |
| Beacon Bank | AI-enabled document classification and spreading (Baker Hill ASK BKR). Streamlines credit analysis in small-business and SBA 7(a) lending. | Rolling out | Jun 16, 2026 | Baker Hill (PR Newswire), Jun 16, 2026 ↗ |
| First Merchants | AI at mortgage intake. AI used as a 'front side sifter' for incoming mortgage applications, with the aim of growing volume from about $1 billion to $1.4 billion with the same staff. | Rolling out | Jun 5, 2026 | Louisville Business First, Jun 5, 2026 ↗ |
| ConnectOne | Digital agents and business intelligence in the loan platform. On the Q2 2026 call the bank said nCino digital agents and business intelligence cut time on some manual lending processes by over 50%. | Rolling out | Jul 14, 2026 | Investing.com, Jul 23, 2026 ↗ +1 |
| ConnectOne | Custom commercial-lending agents (16 skills). Two agents built on nCino's Agentic Operating System for high-friction lending tasks; one uses Document Intelligence to update individual and business relationships. | Rolling out | Jul 14, 2026 | nCino (GlobeNewswire), Jul 14, 2026 ↗ |
| Park National | nCino commercial loan origination system. Park selected nCino for a new commercial LOS to speed origination and manage risk, with completion targeted for May 2026; the annual report does not describe AI features, and no later filing confirms go-live. | Rolling out | Mar 9, 2026 | Park National Corporation (SEC filing, ARS), Mar 9, 2026 ↗ |
| DBS | Agentic credit-memo chain for corporate lending. A chain of roughly 70 to 80 agents assembles credit memos for large corporate clients from annual reports, news and internal notes; the relationship manager interrogates and finalises the draft, and DBS caps the steps in any chain with checks between stages. | Rolling out | May 19, 2026 | Computer Weekly, Oct 5, 2026 ↗ +2 |
| TD Bank Group | Automated adjudication in auto finance and small business. The 2025 annual report plans AI to raise automatic-adjudication rates in TD Auto Finance and Small Business Banking; by August 2026 TD Auto Finance Canada had automated about a third of manual funding processes and launched digital income verification. | Rolling out | Dec 4, 2025 | TD Bank Group (SEC filing), Dec 4, 2025 ↗ +1 |
| BMO Financial Group | Aura commercial credit agent. Generative-AI assistant that drafts commentary for key sections of commercial credit submissions from approved internal systems; underwriters review and decide. | Rolling out | Aug 6, 2026 | BMO Financial Group (BMO AI Hub), Aug 6, 2026 ↗ |
| RBC | AI agents in commercial and wholesale credit. Clients' financials are ingested and spread using AI (Q2 2026 call); the December 2025 strategic update lists improved document-ingestion accuracy, shorter credit-memo preparation and auto-decision targets for small business, delivered through AI agents launching throughout 2026. | Rolling out | May 28, 2026 | Yahoo Finance, May 28, 2026 ↗ +1 |
| Deutsche Bank | AI-generated material in the credit process, reviewed by credit officers. Credit officers review, validate and challenge AI-generated material, which the Chief Risk Officer says makes the process faster and more consistent; the CEO described AI-assisted credit and compliance checks as more precise and faster with the decision remaining with humans. | Rolling out | Jan 23, 2026 | Deutsche Bank, Jan 23, 2026 ↗ +1 |
| Deutsche Bank | AI-accelerated Corporate Bank credit process and mortgage decisions. AI used to significantly accelerate corporate credit; Mittelstand time from first touchpoint to cash cut by up to 95%; mortgage handling time targeted 60% lower by 2028. | Rolling out | Nov 17, 2025 | Deutsche Bank, Apr 29, 2026 ↗ +2 |
| BNP Paribas | Conversational income-and-expenditure assessment. BNP Paribas Personal Finance UK is embedding Inicio AI’s virtual agent, which automates affordability assessments, into its customer journey after a 12-month trial in the Creation Finance app. | Rolling out | Feb 18, 2026 | FinTech Futures, Feb 18, 2026 ↗ |
| Barclays | AI in wholesale credit risk systems. The CEO said the bank is rebuilding wholesale credit risk architecture and using AI to aggregate and analyse data and generate reports to support fast, accurate credit decisions. | Rolling out | Feb 10, 2026 | Yahoo Finance (Quartr), Feb 10, 2026 ↗ |
| Citizens | GenAI credit research and monitoring. Automating credit research and portfolio monitoring of private companies; an AI pilot for commercial prospecting, deal research and pitch creation. | Pilot | Apr 16, 2026 | Citizens Financial Group (SEC filings), Apr 16, 2026 ↗ +1 |
| M&T | AI anomaly detection and LLM experiments in credit. Faster credit decisions and earlier warning of changes in a commercial customer's activity; experimental large-language-model work on credit applications. | Pilot | Sep 18, 2026 | MarketBeat (Yahoo Finance), Sep 18, 2026 ↗ |
| Flagstar | StarIQ for underwriting data discovery. Credit underwriting uses StarIQ to find the data most relevant to faster decisions; risk management and marketing are the heaviest users. | Pilot | Jul 30, 2026 | American Banker, Jul 30, 2026 ↗ |
| Valley | Automation and agentic AI in underwriting and loan servicing. Financial-statement spreading and rent-roll population are automated; agentic AI for underwriting workflow is shown in the deck but, per the chief credit officer, still in exploration. | Pilot | Jul 23, 2026 | Valley National Bancorp (SEC filings), Jul 23, 2026 ↗ +1 |
| FNBO | Upstart AI consumer-lending pilot. AI-based lending software tested to expand access to credit, with the bank's normal credit policies (FICO, debt-to-income) embedded. | Pilot | Oct 21, 2019 | American Banker, Oct 21, 2019 ↗ |
| WaFd Bank | nCino Analyst Digital Partner. AI agent for loan-document review, covenant work, credit risk assessment and financial analysis in commercial lending. | Pilot | May 29, 2026 | nCino, May 29, 2026 ↗ +1 |
| Customers Bank | AI-powered loan closing and account opening. Selected commercial loans closed in seven days versus 30 to 60; account-opening workflow cut 68%. | Pilot | Jan 22, 2026 | Customers Bancorp (Business Wire via Morningstar), Jul 23, 2026 ↗ +1 |
| OceanFirst | Machine-learning loan decisioning pilot. The roadmap lists a loan-decisioning pilot (ML) in the 2023–2025 foundation phase and advanced credit underwriting in the 2026–2027 phase. | Pilot | May 1, 2026 | OceanFirst Financial Corp. (SEC filings), May 1, 2026 ↗ |
| Live Oak | Casca AI loan origination for Live Oak Express. AI loan application and origination for SBA 7(a) loans under $350,000; target of seven to ten days from application to funding. | Pilot | Aug 19, 2025 | American Banker, Apr 28, 2026 ↗ +1 |
| Live Oak | Loan-document ingestion and credit memos. Pilots in which AI ingests loan documents and helps prepare credit memos, with human review retained. | Pilot | Oct 23, 2025 | StockStory via Yahoo Finance, Oct 29, 2025 ↗ |
| Hilltop | Enhanced Desktop Underwriter pilot with Fannie Mae. A single source verification asset report validates income, employment and assets in Fannie Mae's automated underwriting, cutting processing time and raising automated approvals (HousingWire, 2024). Described as underwriting automation, not AI. | Pilot | Nov 1, 2024 | HousingWire, Nov 1, 2024 ↗ |
| MUFG | AI credit experts and AI Loan Expert. Specialised agents, developed with Sakana AI, that incorporate in-house tacit knowledge and help draft credit approval documents; the integrated report says a pilot verified gains in efficiency and quality in complex parts of the loan process, using more than 1,400 evaluations and comments from experienced staff. | Pilot | May 19, 2026 | Mitsubishi UFJ Financial Group, Inc., May 19, 2026 ↗ +2 |
| Commonwealth Bank | AI-assisted and agentic business lending. AI for faster loan decisioning through BizExpress (auto-decisioned small-business loans more than doubled FY26 versus FY23), and agentic capability across the lending process including a first controlled end-to-end business loan pilot. | Pilot | Aug 12, 2026 | Commonwealth Bank of Australia, Aug 12, 2026 ↗ +1 |
| Fifth Third | AI upgrade of lending models. SVP Jay Budzik says the bank plans to expand its balance sheet by billions by moving lending-business models to AI; Schramm lists embedding AI in lending and credit processes. | Announced | Jun 1, 2026 | Northwestern McCormick School of Engineering, Oct 12, 2026 ↗ +1 |
| KeyBank | AI to accelerate credit decisioning. One of four AI themes the CEO set out for 2026, alongside client experience, technology productivity and risk and security monitoring. | Announced | Jan 20, 2026 | Yahoo Finance, Apr 16, 2026 ↗ +1 |
| Pinnacle | Credit intelligence. One of three stated AI priorities: using AI to cut the time from client application to closing. No deployed system, vendor or metric is public. | Announced | Apr 23, 2026 | The Motley Fool (earnings call transcript), Apr 23, 2026 ↗ |
| Zions | Zest AI for lending decisions. Zest AI reports that Zions Bank selected its AI lending technology in the second quarter of 2026 to strengthen lending decisions; scope and products not disclosed by the bank. | Announced | Aug 5, 2026 | Zest AI, Aug 5, 2026 ↗ |
| Zions | nCino platform for commercial and small-business origination. nCino tools including Banking Advisor, Commercial Pricing, and Profitability and Analysis to speed end-to-end lending; the bank has not published AI-specific results. | Announced | Apr 15, 2025 | nCino (GlobeNewswire), Apr 15, 2025 ↗ |
| First Horizon | AI across the commercial lifecycle. Onboarding, origination, underwriting and servicing opportunities under the commercial-strategy head. | Announced | Jul 17, 2026 | First Horizon, Jul 17, 2026 ↗ |
| First Horizon | AI credit-analysis tool and comparable-data analytics. A new tool from the credit analyst team was shown at the March 2026 leadership forum; the Chief Credit Officer sees value in faster financial spreading and in pulling comparables and public-filing footnote data. | Announced | Mar 3, 2026 | Investing.com, Mar 3, 2026 ↗ |
| Columbia | AI in the commercial lending process. Work expected over the next couple of quarters to speed analytics, reporting and decision-making in commercial lending, 'not to decide the credit yes or no'. | Announced | Sep 15, 2026 | Investing.com, Sep 15, 2026 ↗ |
| Texas Capital | Credit agents for portfolio reviews. Announced for the quarter after the Q1 2026 call. | Announced | Apr 23, 2026 | The Motley Fool (Yahoo Finance), Apr 23, 2026 ↗ +1 |
| Sallie Mae | Income verification in underwriting (Nova Credit Income Navigator). Selected on 1 July 2026 to verify consumer income as part of private student loan underwriting. Nova Credit describes a cash-flow data and analytics platform; the release does not call the product AI, name a go-live date or give contract terms. | Announced | Jul 1, 2026 | Nova Credit (Business Wire), Jul 1, 2026 ↗ |
| Arvest | Google Document AI for commercial lending (planned). Named in the 2022 Google Cloud agreement to advance commercial lending decisions; no deployment has been reported. | Announced | Jul 25, 2022 | Arvest Bank, Jul 25, 2022 ↗ |
| WesBanco | Credit underwriting automation. Investor presentations from May 2026 name credit underwriting among the functions where the bank is advancing automation and AI; no tool, vendor or result is disclosed. | Announced | May 4, 2026 | WesBanco (SEC filings), May 4, 2026 ↗ +1 |
| First Financial Bank | Upstart Referral Network personal loans. From February 2021 the bank has lent through the Upstart Referral Network, whose AI lending platform sends qualified applicants to a First Financial-branded application; the programme was announced in April 2021. The bank's own filings do not describe it, and no 2026 status appears in an allowed source. | Announced | Apr 14, 2021 | Upstart (Business Wire), Apr 14, 2021 ↗ |
| BNP Paribas | AI agents for corporate credit memos. Planned deployment of agents on Gemini Enterprise for targeted CIB workflows, starting with the preparation of corporate credit memos and extending to sales, trading, research and structuring. | Announced | Sep 24, 2026 | BNP Paribas, Sep 24, 2026 ↗ +1 |
| MUFG | AI and satellite imagery for collateral appraisal. A planned system for inspecting real estate collateral, due from fiscal 2027, that Nikkei Asia says MUFG Bank may share with regional banks. | Announced | Jul 7, 2026 | Nikkei Asia, Jul 7, 2026 ↗ |
Three patterns stand out. First, the big banks' published results are in cards and consumer credit, where the models are old and the volumes make a 100-basis-point approval gain (Citi) or 50,000 extra approvals (Synchrony) worth announcing. Second, mortgage is where agentic AI has moved fastest in 2026, because the file is long and the rules are explicit: TD, ING, Scotiabank, Lloyds, BNP Paribas, First Merchants and Deutsche Bank all describe it, and ING's change lead says the obvious thing out loud: mortgages look like the worst domain to apply AI to, which is why the win is large. Third, commercial lending at regional banks is the widest category by count, because spreading, covenant extraction and memo drafting are hours of analyst work per deal and the vendors (nCino, Abrigo, Baker Hill, Covecta) sell them as agents: Zions, KeyBank, First Horizon, ConnectOne, Axos, WaFd, Valley. 51 rows are in production; the AI disclosure index ranks the US banks on their whole record.
What results do banks report from AI in lending?
Citi's new AI models in US Cards raised approval rates by 100 basis points; Synchrony's PRISM decides in six seconds or less and approved nearly 50,000 customers in 2025 through cash-flow underwriting who would otherwise have been declined; First Hawaiian's Zest AI model took automated card decisions from 4% to 55% of applications with a 65% approval rate after a year; TD Bank's agentic pre-adjudication cut mortgage files from an average of 15 hours to under three minutes; Scotiabank reports almost 70% less mortgage-underwriting verification time; RBC offers about 40% more credit-limit increases to newcomers; ING takes 65% of Dutch consumer-loan decisions instantly; Deutsche Bank cut Mittelstand time-to-cash by up to 95%; SouthState has about 90% of credit spreads done by AI; KeyBank cut analyst research time in commercial approvals from 60% to 25% of the process; Customers Bancorp closed pilot commercial loans in seven days instead of 30 to 60.
These are the banks' own figures as reported, with the date and the source. They measure different things (approval rate, decision time, share of decisions automated, hours per file, cycle time, origination volume) and are not comparable with each other; vendor-reported figures and targets are labelled as such. The strongest are the ones with a stated before-state and denominator: TD's 15 hours to under three minutes per mortgage file, First Hawaiian's 4% to 55% of applications, KeyBank's 60% to 25% of the approval process.
| Bank | Metric | Value | Date | Source |
|---|---|---|---|---|
| Citi | Cards approval-rate gain from new AI models | +100 basis points | May 7, 2026 | Citigroup, May 7, 2026 ↗ |
| Synchrony | Time to decide a credit application in PRISM | Six seconds or less | Jun 1, 2026 | American Banker, Jun 1, 2026 ↗ |
| Synchrony | Customers approved in 2025 via cashflow underwriting who would otherwise have been declined | Nearly 50,000 (up to 150,000 expected in 2026) | Dec 31, 2025 | American Banker, Jun 1, 2026 ↗ |
| Synchrony | Accounts approved through PRISM since January 2018 | More than 180 million | Mar 31, 2026 | Fortune (commentary by Synchrony CEO Brian Doubles), Mar 31, 2026 ↗ |
| First Hawaiian | Automated credit-card decisions (vendor-reported, Zest AI) | 55% of applications, from 4%; instant approvals 40%, from 4% | Jun 5, 2025 | Zest AI (vendor case study), Jun 5, 2025 ↗ |
| First Hawaiian | Approval rate after one year of AI underwriting (Zest-hosted Celent summary) | 65% | Aug 8, 2025 | Zest AI (vendor page), Aug 8, 2025 ↗ |
| TD Bank Group | Mortgage pre-adjudication time | From an average of 15 hours to under three minutes | May 21, 2026 | TD Bank Group, May 21, 2026 ↗ |
| TD Bank Group | Mortgage funding and discharge cost per file | Funding about C$187 to about C$47; discharge about C$22 to C$23 to about C$11 (CEO targets) | Sep 9, 2026 | Investing.com, Sep 9, 2026 ↗ |
| Scotiabank | Reduction in mortgage-underwriting verification time from AI (management statement) | Almost 70% | Sep 14, 2026 | Investing.com, Sep 14, 2026 ↗ |
| RBC | Newcomer credit uplift | About 40% more credit limit increases for clients with under two years of Canadian credit history; close to C$300 million extra credit | Oct 6, 2026 | Royal Bank of Canada (CNW via Financial Post), Oct 6, 2026 ↗ |
| ING | Instant credit decisions for consumer loans (Netherlands) | 65% | Mar 18, 2026 | ING Groep N.V., Mar 18, 2026 ↗ |
| ING | Complex mortgage processing time (Netherlands, agentic AI) | From seven days to five | Sep 23, 2026 | Investing.com, Sep 23, 2026 ↗ +1 |
| Deutsche Bank | Mittelstand loan time from first touchpoint to cash | Up to 95% shorter | May 28, 2026 | Deutsche Bank, May 19, 2026 ↗ |
| BNP Paribas | Mortgage files using document AI | About 80%; about 30 minutes saved per file | Mar 5, 2025 | BNP Paribas, Mar 5, 2025 ↗ |
| Société Générale | Stella credit-analysis time per analysis | 2-20 hours, down from 8-34 hours (about 50%) | Feb 1, 2025 | Societe Generale, Feb 1, 2025 ↗ |
| OCBC | Corporate relationship manager credit-prep time | Reduced by up to 30% | Dec 31, 2025 | Oversea-Chinese Banking Corporation (via HKMA Banking Disclosure Repository), Mar 10, 2026 ↗ |
| BMO Financial Group | Commercial underwriting time target | Roughly 50% lower by fiscal 2028, with 70% of commercial processes powered by AI | Aug 6, 2026 | BMO Financial Group (BMO AI Hub), Aug 6, 2026 ↗ |
| KeyBank | Analyst research time in commercial loan approvals | 60% → 25% of the process | Nov 16, 2025 | Google Cloud, Nov 16, 2025 ↗ |
| SouthState | Credit spreads performed by AI | ≈ 90% | Jun 30, 2026 | SouthState Bank Corporation (SEC filings), Jul 23, 2026 ↗ |
| Customers Bank | Commercial loan closing time in pilot | 7 days, from 30 to 60 | Jul 23, 2026 | Customers Bancorp (Business Wire via Morningstar), Jul 23, 2026 ↗ |
| Old National | Loan-reviewer efficiency gain with Abrigo Loan Review Assistant (vendor-reported) | ≈ 31% on average | Jul 7, 2026 | Abrigo, Jul 7, 2026 ↗ |
| Seacoast Bank | Clean loan-review line sheet, time to complete | 8 to 10 hours before, 4 to 6 hours with Abrigo Loan Review Assistant | Oct 8, 2026 | Abrigo, Oct 8, 2026 ↗ |
| ConnectOne | Manual lending process time, nCino agents and BI (Q2 call) | Reduced by over 50% on some processes | Jul 23, 2026 | Investing.com, Jul 23, 2026 ↗ |
| Western Alliance | Lending cycle time after the AI-assisted redesign | −25% (internal projection) | May 12, 2026 | Western Alliance Bancorporation (SEC filings), May 12, 2026 ↗ |
| Atlantic Union | Consumer-lending cycle-time reduction from automation (investor-day estimate) | 20% | Dec 10, 2025 | Atlantic Union Bankshares (SEC filings), Dec 10, 2025 ↗ |
| Zions | Mortgage closing pages needing manual review after AI funding QC | −98% | Oct 3, 2024 | Snapdocs (Business Wire), Oct 3, 2024 ↗ |
| Hilltop | PointServ integration: customer adoption / automated approvals | Up more than 65% / up more than 15% | Feb 3, 2025 | HousingWire, Feb 3, 2025 ↗ |
| Wintrust | Wintrust Specialty Finance originations, first half 2026 (with AI credit adjudication) | More than 4,000 deals, roughly $500 million; team of 34 | Jun 30, 2026 | Equipment Finance News (first published by sister title FinAi News), Aug 6, 2026 ↗ |
| SoFi | Faster deployment of credit-risk strategies after moving to Oscilar | 50% faster; 30%+ faster decisioning | Jun 3, 2026 | Oscilar (PR Newswire), Jun 3, 2026 ↗ |
| FNBO | Upstart pilot results and nationwide approvals | Net Promoter Score 83; 78% of loans approved instantly | Dec 8, 2021 | FNBO and Upstart (Business Wire), Dec 8, 2021 ↗ |
| Bread Financial | Machine learning models embedded in the business | Over 200 | Jan 29, 2026 | Investing.com, Feb 3, 2026 ↗ |
How to read them: an approval-rate or approvals-count gain (Citi, Synchrony, First Hawaiian, RBC) is the number that matters to a regulator as well as to a shareholder, because it is where fair-lending questions start; the banks that publish it also describe the testing behind it. A time-per-file figure (TD, Scotiabank, Deutsche Bank, BNP Paribas, Seacoast) measures the reading and drafting the model took over, not the decision. The targets (BMO's 50% lower commercial underwriting time by fiscal 2028, Western Alliance's 25% cycle-time projection) are programme goals, not results, and are labelled.
What does AI do at each stage of a loan?
Nine jobs, in the order a loan moves: intake and document checks, income and asset verification, financial-statement spreading, credit-memo drafting, decisioning and scoring, pre-adjudication summaries, adverse-action reasons, portfolio monitoring and early warning, and loan review. Each has a bank on the record doing it in production or rolling it out, listed in the table below; in every row the model reads, calculates, scores or drafts and a person owns the decision.
The table follows a loan from the application to the review after it is booked. Each row names the model class and a bank on the record that runs it; the bank page carries the source. The pattern in every row is the same: the model reads, calculates, scores or drafts, and a person owns the decision that ECOA, the fair-lending laws and the credit policy attach to a person.
| Stage | What the model does | Bank on the record |
|---|---|---|
| Application and document intake | Vision and language models classify what the borrower sent, find what is missing and ask for it; a coach guides the applicant through the file. | SoFi: Personal Loan Doc Coach validates pay stubs and other documents during the application. |
| Income, employment and asset verification | Models read statements, tax returns and connected accounts, calculate income and flag inconsistencies that used to take an underwriter a day. | Lloyds: A mortgage income-verification model on Vertex AI cuts the step from days to seconds. |
| Financial-statement spreading | For commercial credit, AI spreads the borrower's statements and tax returns into the bank's template and populates rent rolls and schedules. | SouthState: About 90% of credit spreads performed by AI by mid-2026, with an in-house small language model. |
| Credit memo and analysis drafting | Generative models draft the narrative sections of a credit submission from approved internal data, news and filings; the underwriter edits and owns it. | DBS: A chain of roughly 70 to 80 agents assembles credit memos for large corporate clients. |
| Decisioning and scoring | A supervised model (gradient boosting, decision trees, transformers at the fintechs) scores the application on bureau, cash-flow and behavioural data and returns an approve, decline or refer decision. | Synchrony: PRISM analyses about 9,000 data points and decides in six seconds or less; more than 180 million accounts approved since 2018. |
| Pre-adjudication and underwriter summary | Agents classify documents, compute income, check policy and consent and write the summary the underwriter reads before deciding. | TD Bank: Agentic mortgage and HELOC pre-adjudication: from an average of 15 hours to under three minutes. |
| Adverse-action reasons | An explainability layer attributes the decision to the principal factors the model actually used, in words the notice can carry and the applicant can act on. | Santander US: Machine-learning auto credit risk on FICO Platform returns approve/decline decisions, adverse-action reasons and probability of default. |
| Portfolio monitoring and early warning | Models watch the booked portfolio for deterioration, covenant breaches and sector shocks and open a review before the annual one. | M&T: Agentic and predictive credit monitoring with RDC.AI on commercial portfolios; earlier issue detection. |
| Loan review | A generative assistant drafts the borrower, facility, collateral and cash-flow narratives of the independent review; the reviewer fills in the judgment. | Seacoast: Abrigo Loan Review Assistant: a clean line sheet in four to six hours instead of eight to ten. |
The stage the banks guard most carefully is decisioning, and the record shows the line they draw. Columbia's chief administrative officer: ‘not to decide the credit yes or no, but just speed up the analytics, the reporting, the decision-making.’ Deutsche Bank: credit officers review, validate and challenge AI-generated material. Valley's chief credit officer: no agentic AI in the core analysis, but proven financial-statement spreading and rent-roll population. The consumer side is different because a scorecard has been the decision for decades; there the question is not whether the model decides but whether the bank can explain it, which is the next section.
What do regulators require of AI in lending?
In the US: ECOA and Regulation B require specific principal reasons for adverse action regardless of the technology, FCRA governs the consumer-report data and scores a model uses, SR 26-2 treats a machine-learning credit model as a model to be validated, and the CFPB's 2026 Regulation B rule removed disparate-impact liability from the federal rule while state attorneys general (Massachusetts v. Earnest, $2.5 million) and state ADMT laws (Colorado from 2027) still reach it. In the EU: EBA/GL/2020/06 governs automated creditworthiness assessment now, the AI Act makes credit scoring of natural persons high-risk from December 2, 2027, and the Consumer Credit Directive gives borrowers a right to human intervention and explanation.
No regulator bans machine-learning credit models; all of them require that the lender can explain the decision, test it for discrimination and validate it. The tracked documents most relevant to an AI lending programme are below, US first and the EU texts after; the full set is on the by-use-case page and the fair-lending documents under fair lending. The OCC versus CFPB on AI in lending comparison sets the prudential and consumer-protection views side by side.
| Document | Authority | Region | Date | Status | What it asks of a bank |
|---|---|---|---|---|---|
| ECOA / Regulation B adverse action (15 U.S.C. 1691(d); 12 CFR 1002.9) | CFPB | US | Oct 28, 1974 | In force | The statute and rule that bind every AI lender in the US: a statement of the specific principal reasons for adverse action within 30 days; 'failed our internal standards' or 'did not reach a qualifying score' is expressly insufficient. The duty is the same for a model and for a human underwriter. |
| FCRA adverse action and credit-score disclosures (15 U.S.C. 1681m, 1681g(f)) | CFPB | US | Oct 26, 1970 | In force | When a consumer report or a score built on one feeds the model, the adverse-action notice must name the reporting agency and disclose the score with up to four key factors that hurt it; alternative data used for eligibility can fall under FCRA's accuracy and dispute rules. |
| Regulation B final rule on disparate impact (April 2026) | CFPB | US | Apr 22, 2026 | In force | Effective July 21, 2026: Regulation B no longer authorises disparate-impact ('effects test') liability under ECOA, and discouragement is narrowed to statements of intent. The adverse-action notice rules in 1002.9 were not amended; state attorneys general and state law still reach disparate impact. |
| CFPB Circular 2022-03 | CFPB | US | May 26, 2022 | Withdrawn | Withdrawn May 12, 2025, but it stated the position the statute still supports: a creditor may not use a model so complex that it cannot give specific and accurate reasons for adverse action. |
| CFPB Circular 2023-03 | CFPB | US | Sep 19, 2023 | Withdrawn | Withdrawn May 12, 2025; it said the sample-form reasons are illustrative and a lender must state the actual principal reason, even an unexpected one such as purchasing behaviour. The underlying rule is unchanged. |
| CFPB withdrawal of 67 guidance documents (May 2025) | CFPB | US | May 12, 2025 | In force | The notice that withdrew both AI adverse-action circulars, with the Bureau's statement that withdrawal does not change the underlying statutes and regulations. |
| CFPB Innovation Spotlight on AI/ML adverse action notices (2020) | CFPB | US | Jul 7, 2020 | Final | The Bureau's first written statement that ECOA and Regulation B can be met with machine-learning models: the sample reasons are illustrative and a creditor may disclose a reason not on the form if it accurately reflects a factor the model used. |
| Joint Statement on Automated Systems (CFPB, DOJ, EEOC, FTC) | CFPB | US | Apr 25, 2023 | Final | Four agencies' statement that existing civil-rights, fair-lending and consumer-protection law applies to automated systems, including models built on alternative data. |
| Massachusetts AG Earnest Operations settlement (July 2025) | Massachusetts AG | US | Jul 10, 2025 | In force | What enforcement of AI underwriting looks like at state level: a $2.5 million settlement over a student-loan refinance model that used a cohort-default-rate variable with disparate impact, an immigration-status knockout and inaccurate adverse-action notices; the lender must now run fair-lending testing and an algorithmic oversight team. |
| SB 26-189 | Colorado AI Act | US | May 14, 2026 | Final | From January 1, 2027 Colorado's ADMT Act requires consumer notice, a plain-language explanation within 30 days of an adverse outcome and human review for covered automated decisions including credit; a lender's ECOA/Reg B adverse-action notice satisfies the notice and disclosure duties. |
| SR 26-2 | Federal Reserve | US | Apr 17, 2026 | In force | A machine-learning credit, scoring or pricing model is a model: inventory, validation and monitoring scaled to materiality. The guidance's own history began with credit scoring (OCC 1997-24); generative credit-memo drafting sits outside it. |
| OCC Bulletin 1997-24 | OCC | US | May 20, 1997 | Superseded | The OCC's first examination guidance on algorithmic credit decisions, rescinded in 2026: it named inconsistent overrides, poor monitoring and fair-lending risk from scoring variables as the concerns examiners still raise about AI models. |
| EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06) | EBA | EU | May 29, 2020 | In force | The EU's standing rulebook for automated creditworthiness assessment: specify where automated models are used, keep model governance and model-risk management proportionate, and make sure staff understand and can challenge model outputs. The EBA treats it as already covering high-risk credit-scoring AI. |
| Regulation (EU) 2024/1689 | EU AI Act | EU | Jul 12, 2024 | In force | Annex III 5(b): AI used to evaluate the creditworthiness of natural persons or establish their credit score is high-risk (fraud-only systems excepted). Deployer duties include human oversight, logging and a fundamental-rights impact assessment. |
| Regulation (EU) 2026/1744 (Digital Omnibus on AI) | EU AI Act | EU | Jul 24, 2026 | In force | Defers the high-risk obligations for stand-alone Annex III systems, including credit scoring, from August 2, 2026 to December 2, 2027; Article 50 transparency duties applied from August 2, 2026. |
| Consumer Credit Directive (EU) 2023/2225 | EU AI Act | EU | Oct 30, 2023 | Final | The recast directive gives a consumer the right to human intervention, a meaningful explanation and the right to contest an automated creditworthiness assessment. |
| Supervision Newsletter: AI use cases for credit scoring and fraud detection (Nov 2025) | ECB | EU | Nov 20, 2025 | Final | What the ECB found in 13 banks' AI credit scoring: decision-tree models dominate, no bank allows self-learning after deployment, and the gaps are quantified benefits, explainability definitions and data standards. |
| ECB Guide to internal models (July 2025, ML section) | ECB | EU | Jul 28, 2025 | In force | For capital models: machine learning must be adequately explainable and its complexity justified by performance; linear and logistic regression are not machine learning for the guide. |
| EBA follow-up report on machine learning for IRB models (EBA/REP/2023/28) | EBA | EU | Aug 4, 2023 | Final | Principle-based recommendations on understanding, documenting, validating and monitoring machine-learning capital models. |
Three things changed in 2025 and 2026 and one did not. The CFPB withdrew its two AI adverse-action circulars in May 2025 and, in its April 2026 Regulation B rule, removed disparate-impact liability from the federal ECOA rule; state enforcement moved the other way, with Massachusetts's $2.5 million Earnest settlement and Colorado's ADMT Act from 2027; and the EU's Digital Omnibus pushed the AI Act's credit-scoring obligations to December 2, 2027. What did not change is the duty in ECOA section 701(d) and Regulation B 1002.9: specific principal reasons for adverse action, within 30 days, whatever the technology. In the model-risk layer, SR 26-2 keeps machine-learning credit and pricing models in scope and leaves generative memo drafting to broader governance; model risk management covers what that means in practice.
How does a lender explain a machine-learning credit decision?
By attributing the decision to the principal factors the model actually used and translating them into reasons the applicant can act on. Regulation B's sample reasons are illustrative, so unexpected factors may be disclosed if accurate, and a creditor may not hide behind model complexity: the CFPB said so in Circulars 2022-03 and 2023-03, withdrew both in May 2025, and left the statutory duty intact. Vendors return reasons with the decision (Santander US on FICO Platform, First Hawaiian on Zest AI, Affirm's explainable decisions for thin-file applicants), and Optum Bank's chief credit officer puts the industry view in one line: a model built to make an underwriting decision 'shouldn't be a black box'.
Three things have to be true at once. The reasons have to be the model's actual principal factors, not the closest line on the sample form; the Regulation B commentary has said since 2020 that the sample reasons are illustrative and that an accurate unlisted reason is allowed. The reasons have to be specific enough to act on; ‘insufficient credit history’ passes, ‘score below cut-off’ does not. And the attribution has to be stable and reviewable, which is where the model-risk and fair-lending functions meet: a feature-attribution method that gives a different top reason each time the model is retrained will fail validation, and a top reason that is a proxy for a prohibited basis will fail the fair-lending test before it fails a notice.
The vendors have made this a product feature because the banks ask for it. Santander US's machine-learning auto credit risk on FICO Platform returns approve/decline decisions, adverse-action reasons and probability of default together; Zest AI says First Hawaiian received SR 11-7 model-risk documentation and a model compliant with FCRA, CFPB and ECOA rules; Affirm describes explainable decisions for thin-file applicants from its transformer model. The EU adds a borrower-side right: the recast Consumer Credit Directive and GDPR Article 22 give the applicant human intervention and a meaningful explanation, and the ECB's 2025 workshops found that none of the 13 banks reviewed allows its credit model to self-learn after deployment, for exactly this reason.
Which vendors and fintech lenders supply AI credit decisioning?
6 vendors in this site's credit-decisioning category and 3 lending-platform and document vendors that banks here name for underwriting work; the most named are nCino (4 banks), FICO (2 banks), Zest AI (2 banks), Abrigo (1 bank). 10 fintechs on this site run their own underwriting models, from Upstart and Affirm to Nubank and Monzo.
The first table lists the vendors in this site's credit decisioning category and the lending-platform and document vendors that banks here name for underwriting work, with the number of banks in the US set that name them on the record. A bank naming a vendor is a disclosure, not a market share; the category guide explains what to check when buying, including how the vendor supports adverse-action reasons and fair-lending testing.
| Vendor | What it does | Banks naming it | In production |
|---|---|---|---|
| nCino | Cloud lending and onboarding platform for banks with gen-AI (Banking Advisor), role-based AI agents (Digital Partners) and credit analytics. Named by M&T, First Horizon, WaFd Bank, ConnectOne. | 4 | 1 |
| FICO | FICO Platform decisioning and analytics software for origination, account management and fraud (Falcon), plus the FICO Score. Named by TD, Santander. | 2 | 2 |
| Zest AI | AI credit underwriting for banks and credit unions, plus fraud detection (Zest Protect) and gen-AI lending intelligence (LuLu). Named by Zions, First Hawaiian. | 2 | 1 |
| Abrigo | Lending, loan review, CECL and AML software for US community and regional banks, with generative and agentic AI assistants added since 2025. Named by Old National. | 1 | 1 |
| Provenir | Cloud decisioning platform for credit risk, fraud and identity, with a data marketplace, case management and AI models. Named by SoFi. | 1 | 1 |
| Covecta | Agentic AI for banks and lenders: pre-built agents that read documents, draft credit memos and run loan and deposit operations checks. Named by ServisFirst Bank. | 1 | 0 |
| Ocrolus | Classifies, extracts and checks lending documents (bank statements, pay stubs, tax forms) and returns cash-flow and income analytics. Named by SoFi. | 1 | 0 |
| Scienaptic | AI credit-decisioning models and platform for consumer lending, mainly US credit unions, with fraud checks and an agentic AI layer (iCUE). | 0 | 0 |
| Taktile | Low-code decision engine and AI-agent platform for credit underwriting, onboarding, fraud and AML decisions at lenders and insurers. | 0 | 0 |
The build-versus-buy line is clearer in lending than anywhere else in banking. Card and consumer lenders build their own scorecards and have for decades (Synchrony's PRISM, Bread's 200 models, Citi's cards models); regional banks buy decisioning (First Hawaiian and Zions on Zest AI, Santander US and TD on FICO Platform, SoFi on Oscilar and Provenir) and buy the commercial-lending agents inside their origination platform (ConnectOne, WaFd and First Horizon on nCino, Old National and Seacoast on Abrigo, Beacon and Arvest on Baker Hill, ServisFirst on Covecta); and small-business lending has its own fintech layer (Huntington and Live Oak on Casca, FNBO and First Financial through Upstart). The fintech lenders on this site that run their own models are in the second table; they are competitors, partners and, through Upstart and Plaid, suppliers to the banks above.
| Fintech | Category | Lending use cases on its page |
|---|---|---|
| Nubank | Neobank | nuFormer underwriting; AI Private Banker |
| Chime | Neobank | Transformer event-sequence risk models |
| Monzo | Neobank | ML credit decisioning |
| Affirm | BNPL & lending | Transformer-based real-time underwriting; Real-time bank-account signals; ML fraud detection |
| Klarna | BNPL & lending | Real-time underwriting |
| Happen Bank | BNPL & lending | Machine-learning credit decisioning; Automated loan issuance; Point-of-sale home improvement lending; AI use policies |
| Upstart | BNPL & lending | AI personal, auto and HELOC underwriting; Alternative-data credit access; Verification and fraud-detection models; Automated identity verification; AI models supplied to banks and credit unions; AI model deployment inside a national bank |
| Block | Payments | Machine-learning underwriting |
| Adyen | Payments | Behaviour-based underwriting |
| Plaid | Business finance & infrastructure | Transaction foundation model; LendScore cash-flow credit score; LendScore 2 and LendScore Arc; Decagon AI agents with Plaid-verified accounts |
Where is the fair-lending line?
Where it always was: a model may not discriminate on a prohibited basis, its variables must survive disparate-treatment review and, outside the federal ECOA rule since July 2026, disparate-impact testing by state enforcers and under the EU AI Act. 8 banks here record a fair-lending activity tied to AI, from Renasant's standing fair-lending committee and Ameris's 10-K warning to Santander's open-source discrimination-testing code and SoFi's fairness-constrained training patent.
The fair-lending record on this site is short because banks rarely volunteer it, and that makes the rows below worth reading: a standing fair-lending committee that reports to the board (Renasant), a 10-K that lists equal-opportunity and fair-lending law among the laws applying to AI (Ameris), a consent order requiring a fair-lending programme and machine-learning fairness testing (City National), open-source discrimination-testing code with counterfactual comparators (Santander), a patent on fairness-constrained model training (SoFi) and a 10-K warning about algorithmic bias and digital redlining (Dime).
| Bank | What the record shows | Status | Date | Source |
|---|---|---|---|---|
| Ameris | AI exposure under fair-lending and equal-opportunity law. The 10-K lists equal opportunity and fair lending laws among those applying to AI and warns of unwanted bias; no AI credit-decisioning model is disclosed. | In production | Feb 26, 2026 | Ameris Bancorp (SEC filings), Feb 26, 2026 ↗ |
| Renasant | Standing fair lending compliance committee. The proxy describes a standing committee that includes the chief risk officer and chief compliance officer and reports to the management ERM committee. The 10-K's AI risk factor cites fair access to financial services and fair treatment of consumers as AI concerns; neither document describes an AI lending model. | In production | Mar 2, 2026 | Renasant Corporation (SEC filings), Mar 18, 2026 ↗ +1 |
| Société Générale | Active Shell financial-fragility early warning. An XGBoost scoring model using over 250 variables and 70 data pipelines assesses about 10 million retail clients each month for risk of financial fragility, with a monthly fairness assessment of predictions; 200 million scores have been computed and 270,000 clients contacted in two years. | In production | Feb 1, 2025 | Societe Generale, Feb 1, 2025 ↗ |
| City National | Fair-lending programme and ML fairness testing. The consent order requires a fair-lending risk-management programme; Marc Fisher oversees fair lending and CRA. RBC holds a granted patent for automated fairness testing of ML models (US 12,651,042 B2, June 2026); the bank has not said City National uses it. | Announced | Jan 31, 2024 | OCC, Jan 31, 2024 ↗ +2 |
| SoFi | Fairness-constrained ML training (patent). US 11,928,730 B1 describes training machine-learning models with a fairness improvement for financial decisions such as credit; SoFi has not said whether it runs in production. | Announced | Mar 12, 2024 | Google Patents (USPTO), Mar 12, 2024 ↗ +1 |
| United Bank | Mortgage-lending record on file with the Federal Reserve. In approving the Piedmont merger the Fed recorded an adverse comment on United Bank's home loans to Black borrowers in West Virginia and Virginia and a satisfactory CRA rating from October 2022. No automated underwriting or AI is disclosed; any later adoption starts from this record. | Announced | Nov 29, 2024 | Banking Dive, Dec 2, 2024 ↗ +1 |
| Dime | Algorithmic-bias and 'digital redlining' risk. The 10-K flags fair-lending exposure from AI use as an evolving regulatory risk. | Announced | Feb 20, 2026 | Dime Community Bancshares (SEC filing), Feb 20, 2026 ↗ |
| Santander | Discrimination testing with counterfactual comparators (mutatis-mutandis). Open-source research code for situation testing of algorithmic discrimination, published with a paper on revisiting the comparator in discrimination testing; the group does not say it is used on credit decisions. | Announced | Jun 25, 2026 | Banco Santander, Jun 25, 2026 ↗ |
What the people running these programmes say about them is below. The line that recurs is the one Optum Bank's chief credit officer draws: when machine-learning models are built to make an underwriting decision, the most important thing is that it should not be a black box. The adjacent pages cover the same stack from the other sides: AI for KYC for the identity checks at the front of the application, AI fraud detection for application fraud, and model risk management for the validation the credit model must pass.
| Bank | Who | Quote | Date | Source |
|---|---|---|---|---|
| Citi | Arvind Purushotham, Global Head of Venture Investing, Citi Ventures | “In the old world, lending and underwriting used to be driven by rules... But in the new world, I think it's going to be a combination of some rules being applied, but then some artificial intelligence being brought in.” | Sep 28, 2026 | American Banker, Sep 28, 2026 ↗ |
| City National | Dave McKay, President and CEO, Royal Bank of Canada | “We're just starting to deploy artificial intelligence into commercial auto adjudication and credit adjudication.” | Sep 9, 2026 | Investing.com, Sep 9, 2026 ↗ |
| WesBanco | WesBanco, Q2 2026 investor presentation | “Advancing automation and AI initiatives to improve efficiency, reduce costs, and enhance scalability across loan operations, call center, and credit underwriting” | May 4, 2026 | WesBanco (SEC filings), May 4, 2026 ↗ |
| Valley | Ira Robbins, Chairman and CEO | “We have already utilized AI to improve access to our internal knowledge base, to rethink legacy back-office processes, including card service requests, certain elements of underwriting, and risk monitoring, and to accelerate data analytics and software development.” | Apr 23, 2026 | Earnings call transcript (The Motley Fool), Apr 23, 2026 ↗ |
| KeyBank | Chris Gorman, Chairman and CEO | “As it pertains to AI, we are focused on a few thematic use cases that will enhance client experiences, accelerate credit decisioning, increase technology productivity, and strengthen risk and security monitoring.” | Apr 16, 2026 | Yahoo Finance, Apr 16, 2026 ↗ |
| BMO | Darryl White, CEO, BMO Financial Group | “Already, agentic AI is being embedded in client services and credit decisioning, and it's expanding rapidly across the functions.” | Mar 26, 2026 | BMO Financial Group (investor relations), Mar 26, 2026 ↗ |
| FNBO | Clark D. Lauritzen, Chairman and President | “Our team has made great strides improving our credit underwriting capabilities, technology, and processes, and we saw gains from these efforts in 2025.” | Mar 18, 2026 | First National of Nebraska / FNBO, Mar 18, 2026 ↗ |
| First Horizon | Bryan Jordan, Chairman, President and CEO | “We're looking at how do we automate credit underwriting and how we use AI in places like that. We're looking at agentic and how we use that in customer service operations.” | Feb 11, 2026 | Investing.com, Feb 11, 2026 ↗ |
What is AI credit underwriting?
AI credit underwriting is the use of machine-learning models and, since 2025, generative and agentic AI in the decision to lend: reading and verifying the application file, spreading financial statements, drafting the credit memo, scoring the applicant on bureau, cash-flow and behavioural data, returning the decision and its reasons, and monitoring the loan afterwards. The lender still owns the decision, the adverse-action notice and the fair-lending obligation; the models do the reading, calculating, scoring and drafting.
Which banks use AI in lending?
On the public record, Citi (new AI models in US Cards raised approval rates by 100 basis points), Synchrony (PRISM decides in six seconds or less), TD Bank (agentic mortgage pre-adjudication, 15 hours to under three minutes), First Hawaiian (Zest AI, 55% of card applications decided automatically from 4%), SouthState (about 90% of credit spreads by AI), KeyBank, BMO, RBC, ING, Deutsche Bank, Lloyds, BNP Paribas, DBS, OCBC, Santander and some 60 others describe AI in origination, underwriting, decisioning or monitoring; the table on this page lists every bank with a sourced record.
Is AI allowed in credit decisions?
Yes, under the same law as any other credit decision. In the US, ECOA and Regulation B require a statement of the specific principal reasons for adverse action regardless of the technology, FCRA governs consumer-report data and scores, and a machine-learning credit model is a model under SR 26-2. In the EU, EBA/GL/2020/06 already governs automated creditworthiness assessment and the AI Act makes credit scoring of natural persons a high-risk use, with obligations now due December 2, 2027. No regulator bans the models; all of them require that the lender can explain the decision and test it for discrimination.
How does a lender give adverse-action reasons from a machine-learning model?
By attributing each decision to the principal factors the model actually used (feature-attribution methods such as Shapley values or counterfactual analysis) and mapping them to reasons the applicant can act on. Regulation B's sample reasons are illustrative, so a reason not on the form is allowed if it accurately reflects a factor the model used; 'failed our internal standards' or 'did not reach a qualifying score' is not allowed. The CFPB's 2022 and 2023 circulars said a creditor may not use a model so complex that it cannot give specific reasons; they were withdrawn in May 2025, but the statutory duty they interpreted is unchanged, and Santander US and First Hawaiian describe vendor models that return adverse-action reasons with the decision.
What is cash-flow underwriting?
Scoring an applicant on the inflows and outflows of their bank accounts, with consent, instead of or alongside a bureau score. Synchrony says nearly 50,000 customers were approved in 2025 through cash-flow underwriting who would otherwise have been declined, with up to 150,000 expected in 2026; RBC's ATOM uses transaction histories and non-traditional data to extend credit to clients with thin files and offers about 40% more credit-limit increases to newcomers; Plaid's LendScore and Upstart's models sell the same capability to lenders.
Does AI in underwriting create fair-lending risk?
It can, and the enforcement record shows how. Massachusetts settled with Earnest for $2.5 million in 2025 over a student-loan model whose cohort-default-rate variable produced disparate impact for Black and Hispanic applicants and over inaccurate adverse-action notices; the OCC's 1997 credit-scoring guidance named the same risk. The CFPB's 2026 Regulation B rule removed disparate-impact liability from the federal ECOA rule, but state attorneys general, state ADMT laws such as Colorado's and the EU AI Act still reach it, and banks such as Renasant, Ameris and Santander describe fair-lending committees, 10-K warnings and discrimination-testing code in response.
Does the EU AI Act apply to bank credit scoring?
Yes. Annex III, point 5(b) lists AI used to evaluate the creditworthiness of natural persons or establish their credit score as high-risk, with an exception for systems used solely to detect financial fraud. A bank that buys or configures such a model is a deployer and must ensure human oversight, use the system per instructions, keep logs and complete a fundamental-rights impact assessment. The 2026 Digital Omnibus deferred the stand-alone Annex III obligations from August 2, 2026 to December 2, 2027; the EBA treats its 2020 loan-origination guidelines as already covering most of what the Act asks.
Which vendors sell AI credit decisioning?
The banks on this site name FICO (TD, Santander), Zest AI (Zions, First Hawaiian), Provenir and Oscilar (SoFi), nCino and its agents (M&T, First Horizon, WaFd, ConnectOne), Abrigo (Old National, Seacoast), Blend (Huntington, BMO, Frost), Baker Hill, Casca (Huntington, Live Oak), Covecta and Upstart (FNBO, First Financial); Scienaptic and Taktile sell decisioning without a bank in this set naming them. Among fintech lenders, Affirm, Upstart, LendingClub, Klarna, Nubank and Monzo run their own models.
When a bank changes how it decides a loan, or a regulator moves on AI in lending, you'll read it the next morning.
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am ET · free
plus every tracker, bank and agent page update, the morning after · leave any morning