AI customer segmentation in banking: how banks personalise with models.

What is AI customer segmentation in banking?

AI customer segmentation replaces fixed demographic segments with a score per customer, refreshed from transaction data: propensity to take a product, likelihood to leave, the next best action for a banker, the insight to show at log-in, the deposit rate that keeps the money. 58 of the 140 banks profiled here record a personalisation, next-best-action or marketing model, 39 in production.

A traditional segmentation puts customers into groups and markets to the group: mass affluent, young professional, small-business owner. The model-based version scores each customer on what they are likely to do next and acts on the individual. The inputs are the bank's own transaction and log-in data, increasingly joined to external accounts (SoFi's Coach runs on data from 12,000 institutions and 6.5 billion transactions) and to payments data from small-business clients (BMO with Codat). The outputs are an insight at log-in, an offer or a rate, a prompt to a banker, a reminder at the right moment or a lead list with the reasons attached. On the record this is the use case banks describe in the softest language (“personalised experiences”, “proactive insights”) and measure in the hardest numbers (leads, conversion, revenue, cost of deposits).

The CEOs tie it to the whole programme. BMO's Darryl White: “we're personalizing client experiences, augmenting our teams with deeper insights, and automating how we deliver simpler, more proactive support”. HSBC's Georges Elhedery: AI allows “a personalised experience for each customer, delivered in real time and at scale”. OCBC's Tan Teck Long describes “avatars and hyper-personalisation on a real-time basis, 24/7” starting in wealth. The pages on conversational AI and customer service cover the channels those insights are delivered through; this page covers the models that decide what to deliver.

Which banks personalise with models?

39 banks run a personalisation or segmentation model in production on the record, led by Truist, Huntington, BMO, TD, Fifth Third, Regions, FNB, UBS, Scotiabank, Commonwealth Bank and ING; the table lists every sourced record across all use cases, strongest status first.

The table lists every record across the thirteen use cases whose name or description is a personalisation, next-best-action, insights, pricing or marketing model, strongest status first and largest bank first within a status; a bank with several records appears several times. The date is the earliest sourced event on the bank's timeline for that activity. Vendor-reported figures (Personetics at Central Bank, Huntington and Truist) are labelled in the detail.

Two patterns. The insights engine is the entry point for most banks and is bought (Personetics at five banks here, MX at Hancock Whitney), because it needs only the bank's own transaction feed and a widget in the app. The next-best-action engine for bankers is built, because it needs the CRM, the product catalogue and the compliance rules in one place: BMO's AIM shows the rationale to the banker, FNB's Opportunity IQ mines checking-account transactions with machine-learning segmentation, Regions' ClientIQ flags early credit and attrition risk alongside the opportunity. The newest records add a generative layer on top: First Horizon's offer lists come from Salesforce's agentic features, Regions added a copilot to ClientIQ in Q1 2026, and Ally tests content on six agent personas before real customers see it.

What kinds of models do banks use to segment customers?

Eight kinds on the record: next-best-action and next-best-offer engines (BMO, First Horizon, Columbia, FNB), insights engines in the app (Truist, Huntington, Central Bank), predictive foundation models and propensity scores (TD AI Prism, Commonwealth Bank's 2,000-plus real-time models, Scotiabank's payment predictions), deposit and pricing personalisation (Fifth Third, Huntington, FNB), advisor insights (UBS STAAT, Morgan Stanley LeadIQ, Regions ClientIQ), hyper-personalised marketing (ING, Barclays US, CIBC), retention models (OCBC) and synthetic personas with fairness monitoring (Ally, Société Générale).

The table groups the records by what the model does. Each row names the kind, what it decides and the banks on the record, with the bank's own figure where there is one; the bank page carries the source.

The last row is the one compliance teams will look for. Société Générale's Active Shell is a segmentation model in the literal sense: an XGBoost model on more than 250 variables that assesses about 10 million retail clients each month for financial fragility, and it is the only model on the record described with a monthly fairness assessment of its predictions. A model that finds fragile customers can be used to help them or to sell to them, and the fairness test is what makes the difference auditable.

What results do banks report?

BMO: 18,000-plus commercial leads in 12 months and about $85 million of annualised revenue from personalised offers. ING: conversion more than 3% higher across six countries, more than 7 million customers reached. Columbia: about 50% close rate on next-best-offer leads. FNB: leads 6x more likely to convert; cost of deposits down 21 basis points year on year. Truist: 167 million insights and 31 million-plus client interactions in one quarter. Huntington: 14 million insights a month at 4.7 out of 5 satisfaction. OCBC: 20% rise in home-loan retention. Scotiabank: 92% of reminded clients paid.

The banks' own figures as reported, with date and source. Leads, insights delivered, click-through, conversion uplift, revenue and cost of deposits measure different things; an insights count is an activity measure, a conversion uplift or a revenue figure is an outcome. Targets are labelled; vendor-reported figures (Personetics) are labelled.

The figures that connect a model to money are few and worth naming: BMO's about $85 million of annualised revenue from personalised offers in fiscal 2025, ING's more than 3% conversion uplift, FNB's 21-basis-point fall in cost of deposits attributed to machine-learning pricing, SouthState's $2.3 million raised by an agent-run deposit campaign. First Horizon's CFO gives the honest version: offer take rates beat the model, and marketing spend ran a little over budget for two quarters because of it. Associated plans a 25% rise in marketing acquisition spend in 2026, which is the other way the same models show up in a P&L.

What do fair-lending, UDAAP and privacy rules allow?

A segmentation model that touches credit is a credit decision under ECOA and Regulation B: no protected-class inputs or proxies, principal reasons on a decline, testing to prove it; the CFPB's 2026 rule removed disparate-impact liability from ECOA but state civil-rights law still reaches effects. Outside credit, UDAAP forbids offers that are unfair, deceptive or abusive, FCRA limits the use of consumer-report data, and Colorado's, California's and the EU's automated-decision rules add notice, opt-out and review duties. Abroad, the UK's Consumer Duty and Hong Kong's and Singapore's principles require fair outcomes and transparency.

No rule prohibits a bank from modelling its customers. The constraints bite at three points: when the segment decides anything about credit, when an offer or nudge could be unfair, deceptive or abusive, and when the data or the decision falls under privacy and automated-decision law. The first table sets out each constraint with the tracked document or authority; the second lists the documents with what each asks of a bank. The AI credit underwriting page covers the fair-lending mechanics in depth and model risk management the validation layer.

What the banks say about the constraint is thin. Ameris's 10-K lists fair-lending law among the rules applying to its AI and warns of unwanted bias; Renasant describes a standing fair-lending committee; Valley's phrase for its prospecting model, “in a not-creepy way”, is the closest any US bank here comes to describing a UDAAP control in public. The enforcement precedent is the Massachusetts Attorney General's July 2025 settlement with student lender Earnest over AI-driven underwriting.

Which vendors supply personalisation models?

12 vendors in this site's directory describe personalisation, insights or marketing products; Personetics is the most-named engine (five banks). Salesforce, Codat and MX appear in the bank records without a directory page; the largest banks build their own (TD's Layer 6, Commonwealth Bank's engine, BMO's AIM, UBS's STAAT).

The table lists the vendors in this site's AI tools directory whose products cover personalisation, insights, offers or marketing, with the number of banks in the US set that name them on the record. Salesforce (First Horizon, Fulton), Codat (BMO) and MX (Hancock Whitney) appear in the bank records without a directory page. A bank naming a vendor is a disclosure, not a market share.