AI in retail banking: assistants, fraud, credit and personalisation at consumer banks.

What does AI in retail banking mean?

Four things on the record of the consumer banks profiled here: a virtual assistant in the app and the contact centre, fraud and scam models on cards, logins and payments, AI in credit decisions on cards, loans and mortgages, and personalised insights and offers built from the customer's transactions. 95 of the 140 banks here record a customer-facing assistant (76 in production), 107 record AI in fraud (90 in production) and 87 record AI in credit or fair lending.

Retail banking is high-volume and low-value per interaction, which is why it was the first part of banking to automate customer contact and the first to measure AI in billions. A virtual assistant that answers a balance question costs a fraction of a call (KeyBank's figure is about $0.25 against about $9); a fraud model that scores every card swipe in milliseconds is the only way to run a card book at all; and a credit model that decides a $5,000 loan instantly is what makes the loan worth offering. The generative wave changed two of the four functions. Assistants moved from intent-matching menus to language models grounded in the bank's content (Fargo on Google models, Cora's 21 journeys enhanced with generative AI, ING's chatbot going from 46% of chats handled alone in 2023 to 75% in 2025), and personalisation moved from rules to models that write the nudge. Fraud and credit were already machine learning and mostly stayed that way; what changed there is the adversary, who now has the same tools.

Two reading notes for the tables. First, each function table shows one row per bank at its strongest status, up to the limit stated, strongest status and largest bank first; the full lists are on the sibling pages for chatbots, fraud and credit. Second, corporate, wealth and markets records are excluded here and appear on the corporate and investment banking pages; the AI disclosure index ranks the US banks on their whole record.

Which banks run AI assistants for retail customers?

76 banks run a customer-facing assistant in production. Bank of America's Erica has 24 million-plus active users and 3.6 billion cumulative interactions; Wells Fargo's Fargo passed 1 billion interactions in under three years with more than 3 million Spanish-speaking users; Truist Assist has taken 12.5 million-plus requests since 2023, 9 million-plus handled end to end; U.S. Bank's Smart Assistant is used 2.5 million times a month; NatWest's Cora held 12.9 million conversations in 2025, half with no human; ING's chatbot handles 75% of chats without a person, up from 46% in 2023; KeyBank puts the cost of an AI-handled call at about $0.25 against about $9 for a person.

The assistant is the retail AI feature every large bank has and the one the CFPB has written about most. The pattern across the records: the largest banks build (Erica, Fargo, Truist Assist, U.S. Bank's Smart Assistant), banks outside the US often name their assistant and publish its numbers (NatWest's Cora, BBVA's Blue with 150 queries and transactions and the ability to send Bizum payments, DBS's Joy reaching more than 10 million customers, UBS's chatbots serving 1.2 million clients), and regional US banks buy the contact-centre layer. The contact centre is where the second wave lands: Truist generated 4.5 million-plus AI call summaries and saved 36,000-plus teammate hours in a quarter, Ally streamlined more than 6 million customer-care calls in 2025, NatWest saves more than 70,000 hours a year summarising retail calls, and TD's hold time is down 12% since its generative assistant launched.

Which banks use AI for fraud and scam detection?

90 banks run AI fraud or scam detection in production. Bank of America says AI-enabled processes have halved fraud losses since 2018 and its AI has analysed person-to-person payments for fraud 7 million times; Commonwealth Bank screens more than 20 million payments a day, sends more than 40,000 warning alerts a day and cut customer fraud losses by over 20%; Santander automated up to 90% of card-fraud claims in Brazil with an error rate below 1%; DBS flags high-risk transactions in under 10 milliseconds; United Community's fraud losses fell 50% over two years.

Fraud is where retail banks publish the hardest numbers because losses are a line in the accounts. The records split between the transaction models that have run for a decade and the scam layer built since 2023, when authorised push-payment fraud (the customer makes the payment) became the dominant loss: Commonwealth Bank's NameCheck on first-time payments and its 40,000-plus daily warning alerts, DBS's behavioural-science prompts, WaFd's voice-print-verified consumer wires. Deepfakes are the new adversary; FinCEN's 2024 alert is the reference document and AI fraud detection covers the vendor layer (BioCatch, Pindrop, Socure, Feedzai).

Which banks use AI in consumer credit decisions?

87 banks record AI in credit underwriting or fair lending. Citi's new AI models lifted its cards approval rate by 100 basis points; Synchrony approved nearly 50,000 customers in 2025 through cashflow underwriting who would otherwise have been declined, with up to 150,000 expected in 2026; ING decides 65% of Dutch consumer loans instantly; TD's agentic mortgage pre-adjudication went from 15 hours to under three minutes; Scotiabank cut mortgage-underwriting verification time by almost 70%; Regions' ARC analysed 100 million-plus mortgage repricing scenarios with solutions in under five minutes.

Credit is the function with the most law attached, which shows in how banks describe it: carefully, and often through the result rather than the model. The records that state a mechanism are cashflow underwriting (Synchrony, Citizens' lending platform built for student loans and reused for home equity and cards), agentic document preparation for mortgages (TD, ING's Dutch mortgage assistant that cut complex processing from seven days to five, Scotiabank) and instant decisioning on small consumer loans. Fair-lending testing is disclosed by few banks even though the law requires it; the AI credit underwriting page covers adverse-action and disparate-impact in detail.

Which banks personalise with AI?

25 banks record personalised insights, offers or coaching built on AI. Truist Insights has delivered 1.8 billion insights (an average customer gets 500-plus a year, rated 4.5); Huntington's Heads Up sends 14 million a month across 96 use cases at 4.7 out of 5; ING's hyper-personalised campaigns reached more than 7 million customers with conversion more than 3% higher; Scotiabank's Scotia Intelligence generated 14 million actions since March 2026; SoFi's Coach has held nearly 500,000 conversations with over 90% positive feedback; Lloyds' AI financial assistant serves more than 500,000 Bank of Scotland customers.

Personalisation is the function most often bought rather than built in the US: Personetics is behind Truist Insights, Huntington's Heads Up, U.S. Bank and KeyBank. The generative examples that go beyond insights into coaching are the ones to watch for the regulators' questions about advice: SoFi's Coach, where financial planners evaluate answers and out-of-scope conversations escalate to a person, and Lloyds' AI financial assistant. The segmentation side (which customers get which offer) is on AI customer segmentation in banking.

What results do consumer banks report?

Erica: 3.6 billion interactions, 24 million-plus active users. Fargo: over 1 billion interactions in under three years. Truist: 12.5 million-plus Assist requests, 1.8 billion Insights, 4.5 million-plus AI call summaries. KeyBank: about $0.25 per AI-handled call against about $9 for a person. Bank of America: fraud losses halved since 2018. Commonwealth Bank: customer fraud losses down over 20%. Citi: cards approval rate up 100 basis points. Synchrony: nearly 50,000 extra approvals through cashflow underwriting. TD: mortgage pre-adjudication from 15 hours to under three minutes. ING: 65% instant consumer-loan decisions, 75% of chats with no person.

The banks' own figures as reported, with date and source. The assistant numbers are the most robust on this site because the banks repeat them every quarter and measure them the same way (Erica interactions, Fargo interactions, Truist Assist requests, Cora conversations); the fraud figures are losses or claims, which are audited lines; the credit figures are approval-rate and time deltas the bank chose to publish. Survey percentages about consumer attitudes are left out.

What do the CFPB and other consumer regulators require?

That the technology changes nothing about the duty. The CFPB's chatbot spotlight requires accurate answers and a route to a person; ECOA, Regulation B and FCRA require the specific reasons and key factors behind any credit denial; the 2026 Regulation B final rule resets the disparate-impact standard a credit model is tested against; the 2023 joint statement with the DOJ, EEOC and FTC says complexity is no defence; FinCEN's deepfake alert expects layered identity checks; the EU AI Act makes credit scoring of natural persons high-risk; the FCA's Consumer Duty requires good outcomes whatever the channel; Colorado's ADMT Act adds notice, explanation and human review from 2027.

Retail is the part of banking where AI meets the most law, because the customer is a natural person with statutory rights. In the US those rights are adverse-action notices under ECOA and FCRA, fair-lending testing under Regulation B, UDAAP for anything a chatbot says, and a growing set of state automated-decision laws; in the EU the AI Act adds a high-risk regime for consumer credit scoring and GDPR Article 22 a right not to be subject to a purely automated decision; in the UK the Consumer Duty requires good outcomes through any channel. The documents below are the ones a consumer-bank AI programme is examined against; the tracker holds 238 documents from 41 authorities and the CFPB page lists everything from the Bureau.

The CFPB's 2025 withdrawal of 67 guidance documents, including the 2022 and 2023 circulars on adverse action with complex algorithms, removed interpretation but not the statute: the duty to state specific reasons is in ECOA and Regulation B themselves, and the April 2026 Regulation B final rule is the live text on disparate impact. The 2023 chatbot spotlight remains the clearest statement of what an assistant owes a customer. For the governance a bank wraps around all of this, see AI governance in banking; for the ten risks regulators name and the banks' controls, risks of AI in banking.

Which vendors supply AI for retail banking?

30 published vendors in this site's directory describe assistant, contact-centre, insights, fraud or consumer-credit products; the ones most often named by the banks here are Personetics (5 banks), Glia (4 banks), nCino (4 banks), NICE Actimize (4 banks). The largest banks build the assistant and buy the fraud and data layers.

The table lists the vendors in this site's AI tools directory whose products cover assistants, contact centres, insights, consumer fraud or consumer credit, 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 conversational AI, fraud detection and credit decisioning category guides explain what to check when buying.