How Banks Run on AI: 12 Real Case Studies (2026) | TLY
Home  /  AI Case Studies  /  How Banks Run on AI
AI Case Studies

How Banks Run on AI

Twelve real chartered banks, from JPMorgan and Goldman to DBS and CommBank, and what actually happened when they put AI on the work. Sourced outcomes, the tools by name, and the limits each one kept in place.

Banks run on AI across three layers. Generative-AI assistants boost employee productivity, from JPMorgan's LLM Suite for more than 200,000 staff to Goldman Sachs's firmwide GS AI Assistant and BBVA's rollout of ChatGPT Enterprise. Customer-facing virtual assistants resolve everyday banking at scale, led by Bank of America's Erica, Wells Fargo's Fargo, and NatWest's Cora+. And machine-learning risk engines detect fraud and scams in real time at banks like Commonwealth Bank and DBS. The common thread is discipline: a human keeps oversight of every regulated decision.

If you run a line of business at a bank, sit on a risk committee, or lead a team inside one, you have heard two years of AI promises. Some are real and already in production. Many are software vendors describing a bank that does not exist yet. This piece looks at twelve named, chartered banks around the world, the specific tools they use, and the outcomes they have reported, with a source for each. These are institutions, not wealth-advisory boutiques or independent RIAs, and the focus stays on how the bank itself runs.

Read them as a working menu, grouped by what the AI does. Pay as much attention to what each bank kept human as to the numbers it published, because that boundary is the whole story.

Layer one: assistants that make bankers faster

The largest and least glamorous use of AI inside banks is the internal assistant. It drafts, summarizes, searches policy, and answers the routine question so a person does not have to hunt for it. This is where the headcount numbers get big.

JPMorgan Chase: an in-house assistant for 200,000 employees

JPMorgan built LLM Suite, its own generative-AI assistant layered on top of leading foundation models, and rolled it out across the firm. The platform reached about 200,000 onboarded users within eight months of its summer 2024 release, supporting staff with idea generation and content drafting. The point is the scale of adoption at a bank of this size, not a one-off pilot. Source: JPMorgan Chase.

Goldman Sachs: a firmwide GS AI Assistant

Goldman Sachs launched the GS AI Assistant across the firm after testing it with more than 10,000 employees, close to a quarter of its worldwide workforce. The tool handles summarizing, drafting, and analysis for efficiency gains, and the rollout signals that Goldman sees a general assistant as core infrastructure rather than an experiment. Source: Fortune.

BBVA: ChatGPT Enterprise across the whole bank

BBVA expanded its partnership with OpenAI to bring ChatGPT Enterprise to all 120,000 employees across 25 countries, one of the largest enterprise deployments of generative AI in banking. During the earlier phase, employees who received licenses reported saving close to three hours a week on routine tasks, with more than 80 percent using the tool daily. BBVA rolled it out with security and privacy controls and the ability to build internal agents connected to its own systems. Source: OpenAI.

Morgan Stanley: meeting notes and an internal expert on tap

Morgan Stanley built two OpenAI-powered tools for its advisors. Debrief, running on GPT-4, turns a recorded client meeting, with consent, into notes and draft follow-ups that flow into the CRM. Its companion, the AI @ Morgan Stanley Assistant, answers questions against the firm's own research and is actively used by more than 98 percent of advisor teams. Included here as a bank building internal productivity tooling, not as wealth advice. Source: CNBC.

Deutsche Bank: DB Lumina turns days of research into minutes

Deutsche Bank built DB Lumina on Google Cloud and Gemini to speed up its research desk. Work that used to take hours or even days, drafting notes and summarizing dense filings, now takes minutes, and hundreds of research analysts adopted it after the tool went live in September 2024. It runs with guardrails built for a regulated environment. Source: Google Cloud.

Citigroup: agentic assistants that chain steps together

Citi is moving from single-answer assistants to agentic ones. Its upgraded Citi Stylus Workspaces can take a multi-step objective, for example identifying the top branded-card businesses in a market, distilling their strategy, and translating the result, from a single prompt. Citi began with a pilot group of about 5,000 workers and a phased rollout, and says its broader Citi AI tools now reach more than 175,000 colleagues across 80 countries and jurisdictions. Some press reports cite a larger employee figure; the bank's own release describes a staged expansion. Source: Citigroup.

The pattern under layer one

Own the layer, not just the license

Notice what the biggest banks did. They did not simply buy seats. JPMorgan, Goldman, and Citi built their own assistant on top of foundation models so they control access, data, and guardrails. That is the enterprise version of the same move a solo professional makes when they design a repeatable workflow instead of chatting ad hoc.

Layer two: virtual assistants that serve customers

The second layer faces the customer. These are the banking assistants that answer balance questions, move money, and explain a transaction, at a volume no call center could staff. The numbers here are measured in hundreds of millions of interactions.

Bank of America: Erica passes 3 billion interactions

Bank of America's virtual assistant Erica has surpassed 3 billion client interactions since its 2018 launch, now averaging more than 58 million interactions a month and assisting nearly 50 million users. Erica handles everyday questions, surfaces proactive insights, and routes customers to the right place, which keeps routine volume off human channels. Source: PR Newswire, Bank of America.

Wells Fargo: Fargo handled 245 million requests with no PII sent to the model

Wells Fargo's assistant Fargo handled 245.4 million interactions in 2024, up from 21.3 million the year before, helping customers pay bills, move money, and check activity by voice or text. The engineering detail is the one worth copying: voice is transcribed locally, sensitive data is stripped inside Wells Fargo's own systems, and only cleansed text reaches Google's Gemini Flash 2.0 for intent extraction. No personally identifiable information is passed to the language model. Source: VentureBeat.

NatWest: Cora+ lifts satisfaction and deflects the routine

NatWest upgraded its digital assistant to Cora+ with generative AI in 2024 and later partnered with OpenAI to push further. Cora handled 11.2 million retail conversations in 2024, with roughly half needing no human handoff, and the bank reported a 150 percent rise in customer satisfaction scores for Cora over the year as more queries resolved on the first try. Source: NatWest Group.

Capital One: Chat Concierge, agentic AI on the dealer's website

Capital One built Chat Concierge, a multi-agent assistant on top of Meta's Llama that it customized with its own data. It sits on car-dealer websites and helps buyers compare vehicles and book appointments, with the agents planning, validating their own steps to reduce errors, and explaining options. Some dealers report up to a 55 percent increase in customer engagement, and Capital One says it cut latency roughly fivefold since launch. Source: The Financial Brand.

Layer three: the risk engines that never sleep

The third layer is the one customers never see and would miss most if it failed. Machine-learning models score transactions, messages, and behavior in real time to catch fraud and scams. This is the oldest form of AI in banking and still the highest stakes.

Commonwealth Bank: scam losses down 76 percent

Australia's Commonwealth Bank uses AI-driven detection across payments and its NameCheck and CallerCheck tools to intercept scams before money leaves. The bank reports a 76 percent drop in customer scam losses from the peak, comparing the second half of 2025 with the first half of 2023, alongside more than 900 million dollars invested in FY25 to fight fraud, scams, and financial crime. The AI is one part of a wider defense, and the bank presents it that way. Source: Commonwealth Bank.

DBS Bank: about 370 use cases and a billion-dollar target

Singapore's DBS runs roughly 370 AI use cases on more than 1,500 models, spanning fraud and risk detection, client advisory support, and internal productivity. The bank reported about S$750 million in economic value from AI in 2024 and set a public target above S$1 billion for 2025, a figure it later said it reached. DBS treats AI as an operating discipline built over years, not a single project. Source: CNBC.

The twelve banks at a glance

Use case, what the AI does, the bank, and the honest caveat
Use case What AI does Example bank Caveat
Employee assistantDrafts, summarizes, searches policyJPMorgan LLM Suite (200k staff)Adoption at scale, published time savings vary
Firmwide copilotGeneral analysis and draftingGoldman Sachs GS AI AssistantTested on 10k before firmwide launch
Enterprise chatChatGPT Enterprise plus internal agentsBBVA (120k employees)~3 hrs/week saved reported in pilot
Meeting notesTurns client meetings into CRM notesMorgan Stanley Debrief (GPT-4)98% figure is the companion Assistant
Research deskSummarizes filings, drafts notesDeutsche Bank DB Lumina (Gemini)Analysts still review every output
Agentic workspaceChains multi-step tasks from one promptCiti Stylus WorkspacesPhased rollout from a 5k pilot
Customer assistantAnswers and acts on everyday bankingBank of America Erica (3B+)Handles routine, not complex cases
Voice and chatResolves requests, strips PII firstWells Fargo Fargo (245M in 2024)No PII sent to the model by design
Self-serviceDeflects queries, raises satisfactionNatWest Cora+ (+150% CSAT)About half still need a human
Sales assistantMulti-agent help on dealer sitesCapital One Chat Concierge (Llama)55% is engagement, not sales
Scam defenseScores payments and intercepts fraudCommonwealth Bank (-76% losses)AI is one part of a wider defense
Enterprise valueFraud, advisory, and productivity modelsDBS (~370 use cases)Value figures come from the bank

What this means for anyone who works in a bank

Strip away the scale and the pattern is one any professional can copy. Every bank on this list picked a specific, high-volume task, pointed a governed tool at it, and kept a person on the decisions that carry risk. Nobody handed the credit decision, the compliance sign-off, or the regulated advice to a model. The AI took the drafting, the searching, the routine query, and the first-pass fraud score, and a human owned the outcome.

If you are inside a bank, your first move is not a strategy deck. Find the single most repetitive task in your week, whether that is summarizing filings, drafting client follow-ups, or answering the same policy question, and get one approved tool onto it. Then write down what data may enter that tool before anyone pastes anything, because in banking that rule is not optional.

The honest limits

The numbers are self-reported. Almost every figure here comes from the bank or its technology vendor, not an independent audit. They are credible and attributed, but they are the cases these institutions chose to publish. Treat them as what is possible, not what is guaranteed.

Scale hides the exceptions. Erica and Fargo handle enormous volume because they handle the routine. The hard cases, a dispute, a fraud claim, a distressed customer, still go to people, and that is by design.

Regulation is catching up to agentic AI. As banks move from assistants that answer to agents that act, supervisors are watching closely. See our coverage of China's NFRA guidance on safe AI development in banking and insurance and Singapore's MAS work on an agentic-finance runtime. The safe posture today is AI that drafts, scores, and flags, with a human who reviews and signs.

Frequently asked questions

How do banks actually use AI in 2026?

Across these twelve banks, in three layers. First, generative-AI assistants that help employees draft, research, and summarize, such as JPMorgan's LLM Suite, Goldman Sachs's GS AI Assistant, BBVA's ChatGPT Enterprise, and Deutsche Bank's DB Lumina. Second, customer-facing virtual assistants that resolve everyday banking at scale, such as Bank of America's Erica, Wells Fargo's Fargo, and NatWest's Cora+. Third, machine-learning risk engines that detect fraud and scams in real time, such as those at Commonwealth Bank and DBS. In every case a human keeps oversight of regulated decisions.

Will AI replace bank employees?

Not on the evidence here. These banks pointed AI at drafting, research, routine customer queries, and fraud scoring, and kept people on judgment, credit decisions, and compliance sign-off. BBVA reported pilot employees saving about three hours a week, and Wells Fargo's Fargo handles routine requests so staff can focus on complex ones. The pattern is fewer routine interruptions and more time on higher-value work, not fewer bankers.

Which AI tools are big banks using?

A mix of general and specialized systems. JPMorgan built its own LLM Suite on top of leading models; Goldman Sachs runs a firmwide GS AI Assistant; BBVA deployed ChatGPT Enterprise from OpenAI; Morgan Stanley uses OpenAI's GPT-4 for its Debrief and Assistant tools; Wells Fargo's Fargo runs on Google's Gemini Flash 2.0; Deutsche Bank's DB Lumina uses Google Cloud and Gemini; and Capital One's Chat Concierge is built on Meta's Llama. No single vendor wins, and several banks build their own layer on top.

Is it safe for banks to put customer data into AI?

Only with controls decided in advance, which is exactly what these banks built. Wells Fargo designed Fargo so no personally identifiable information reaches the language model; sensitive detail is stripped locally before anything is sent. Deutsche Bank wrapped DB Lumina in guardrails for a regulated environment, and BBVA rolled out ChatGPT Enterprise with security and privacy controls. The rule is the same one that applies to any professional: decide what data may enter a tool before anyone uses it.

How much money is AI saving banks?

The clearest public figure comes from DBS Bank, which reported about S$750 million in economic value from AI in 2024 and set a target above S$1 billion for 2025. Commonwealth Bank credits AI-led fraud and scam defenses, alongside other measures, with a 76 percent drop in customer scam losses. Most banks report time recovered rather than a single dollar number, and the figures come from the banks and their vendors, not an independent audit.

Anthony Guerriero is the founder of The Leveraged Years and a CPA and former Deloitte Senior Manager. He built and scaled a medical logistics company from 6 to 1,800 employees and has advised UHNW clients on cross-border real estate transactions across more than 40 countries. The Leveraged Years teaches senior professionals and operators how to use Claude, made by Anthropic, to do their best work faster without compromising their judgment or professional standards.

Take it further

Get the workflow + SOP.

The full briefing prompts and matching SOP pages that turn these bank patterns into a repeatable desk system live inside the Club. Free with any course, or $49/month direct.

Open the Club →
If this matches your work

Find the right course.

Six diagnostic questions, one course recommendation. We will point you at the program that maps to the work in this briefing, then send your workflow assessment.

Take the selector →