Ninety-two percent of global banks now report active AI deployment in at least one core banking function. On the surface, AI in banking looks like a solved problem, a box every institution has already checked. Look at the returns, and a very different picture shows up. IDC’s research on this same wave of adoption found that the frontrunners, the banks treating AI as a rebuild of how decisions get made rather than a software purchase, are earning a 2.84x return on their investment. The banks trailing behind are earning 0.84x. That number means the project is, on average, losing money.
That gap is the real story of AI in banking right now. It isn’t a story about whether the technology works. It’s a story about the difference between buying a chatbot and rebuilding how decisions get made, and most of the public conversation still treats those two things as the same. Vendors selling either version call both of them AI in banking in the same breath, which is part of why a bank’s board can approve a budget for the technology without anyone agreeing on which of the two they actually bought.
This article breaks down what’s really inside that gap: what the technology actually is underneath the marketing, which banks are pulling ahead and why, where it’s being used inside a bank today, and where the genuine risk still sits.
Three different technologies share one label
Most of the confusion around AI in banking starts here: the phrase covers three genuinely different technologies, each at a different stage of maturity, and most articles talk about all three as if they were one thing.
The oldest and most mature is predictive machine learning, models trained to spot patterns in transaction data and flag the ones that look wrong. This is what powers fraud detection, the single most widely deployed AI use case in financial services. It’s been running quietly inside banks for the better part of a decade, well before anyone started calling it AI in marketing copy.
The newer, more visible technology is generative AI: large language models that draft, summarize, and answer questions in natural language. Bank of America’s Erica is the clearest large-scale example, a virtual assistant that has logged more than 3.2 billion customer interactions. This is what most customers picture when they hear AI in banking, because it’s the layer they actually talk to.
The newest and least settled technology is agentic AI: systems built to plan and carry out multi-step work with minimal human prompting at each step. An agentic system doesn’t just answer a question, it decides what to do next, pulling a document, checking it against a rule, escalating it if something doesn’t match, and moving to the next case without a person clicking through each stage. This is the technology behind a striking 2025 statistic: fifty of the world’s largest banks announced more than 160 agentic AI use cases in that year alone.
Three different technologies, three different maturity levels, three different risk profiles. Most confident-sounding statistics about AI in banking are quietly blending all three together.
Why some banks are profiting and others aren’t
The list of banks running real AI deployments is no longer a short roster of innovation labs. JPMorgan, HSBC, Citi, DBS, and Bank of America are the names that show up most often in case studies, and in early 2026 Goldman Sachs confirmed an enterprise deployment of Anthropic’s Claude, a sign that even the most conservative parts of the industry are now comfortable putting a frontier model inside core workflows.
The more revealing fact is the split inside that group of adopters. McKinsey projects that the pioneers in AI in banking, the institutions moving fastest and most deliberately, stand to gain roughly a 4 percentage point advantage in return on tangible equity over slower-moving peers, a profitability metric boards and investors track closely. That lines up with IDC’s 2.84x versus 0.84x return split mentioned earlier. The technology itself is available to every bank on roughly equal terms. The return on it is not, and the difference comes down to whether a bank treated the rollout as an IT purchase or as a change to how the business actually operates.
Regulation is shaping that split too, in a way that cuts both directions. The Monetary Authority of Singapore introduced FEAT principles, fairness, ethics, accountability, and transparency, specifically for AI in financial services, giving Asia-Pacific institutions a concrete framework instead of a vague instruction to be careful. DBS, operating directly under that framework, is repeatedly cited as one of the region’s fastest movers on agentic deployments, reporting cost reductions of 20 to 40 percent across the functions it has automated. Clear rules, in this case, appear to have sped adoption up rather than slowed it down.
The US is a useful contrast. The Federal Reserve, OCC, and FDIC jointly issued new model risk guidance in April 2026 that explicitly excludes generative and agentic AI from its scope, calling the technology too novel and fast-moving to govern under the old rulebook. That’s an unusual admission for a regulator to make, and it leaves US banks deploying the newest, fastest-growing layer of AI in banking with less guidance than they have for the older, simpler one.
Adoption is outrunning the rules that govern it
Fraud detection became mainstream first and fastest. As of 2026, 90 percent of financial institutions use AI for fraud detection, a number that’s climbed steadily over several years rather than spiking suddenly. It’s the oldest, best-understood layer, and the adoption curve reflects that.
Agentic AI is on a much steeper, much newer curve. Roughly 70 percent of financial services organizations are deploying or actively exploring it in 2026, up from a figure close to negligible only two or three years ago. Wolters Kluwer projects that 44 percent of finance teams will use agentic AI this year, an increase of more than 600 percent.
Regulation is racing to keep pace, not trailing years behind the way it usually does. The EU AI Act’s high-risk provisions apply from August 2026. Colorado’s AI Act takes effect that same June. FINRA issued a 2026 oversight report specifically addressing how supervisory duties apply when an AI agent, not a person, is the one making the decision. None of this is settled law a bank can simply check against and move on. The rulebook is still being written while the deployments it’s meant to govern are already running in production.
Where AI is actually doing work inside a bank
Fraud detection is still the heaviest concentration of AI in banking by deployment volume, but it’s far from the only place doing serious work now. KYC and anti-money laundering compliance has become the second major front, and the productivity numbers there sound implausible until you see how the work actually changed.
One global bank now runs ten coordinated AI agent squads, each made up of four to five individual agents, to handle KYC reviews that used to run on a periodic schedule. Moving to continuous, event-driven due diligence has produced reported productivity gains of 200 to 2,000 percent in that specific function, not because any single step got faster, but because the review no longer waits for a scheduled batch to start.
Customer service is the most visible front, the one customers directly interact with, running through tools like Erica and similar copilots at other major banks. Credit underwriting is quieter but increasingly automated, with agentic systems pulling income, asset, and employment data from multiple sources and cross-checking it against regulatory requirements before a human underwriter ever opens the file.
A newer front is opening up directly inside digital asset infrastructure. As banks roll out tokenized securities and stablecoin settlement, someone has to monitor that activity for fraud and compliance violations, at a transaction volume no manual review team could realistically keep up with.
AI in banking is increasingly the layer doing that monitoring: screening stablecoin transfers the same way it already screens card transactions, and verifying a counterparty’s credentials before a tokenized asset changes hands. Our breakdown of stablecoin trends and regulation in banking covers how fast that infrastructure has moved from pilot to production across Asia. It’s exactly the kind of environment where this layer of AI stops being optional, simply because the transaction volume tokenization is built to unlock would overwhelm a manual compliance team almost immediately.
What’s actually paying for all of this
The financial case is large enough on its own to explain the urgency. Global banking saved an estimated $120 billion from AI in 2025, a figure projected to reach $500 billion annually by 2030. JPMorgan’s fraud system alone is credited with preventing roughly $1.5 billion in losses every year, screening more than 5,700 behavioral signals per transaction, things like typing cadence and payment instruction language that a rule-based system was never built to notice.
McKinsey puts the theoretical ceiling even higher, estimating that AI could cut certain cost categories across banking by as much as 70 percent. After accounting for the rising cost of running the AI itself, McKinsey’s realistic estimate settles closer to 15 to 20 percent in net savings, still a figure it sizes at $700 to $800 billion industry-wide.
There’s a defensive case running in parallel that gets less attention but matters just as much. Generative AI is making fraud easier to commit at the same time it’s making fraud easier to catch. Deloitte projects that generative AI enabled fraud could reach $40 billion in the US by 2027, up from roughly $12 billion in 2023, and Citi reports that AI is already involved in some form in half of all the fraud attempts it sees today.
Banks aren’t only adopting AI in banking to cut costs. Many are adopting it because the people trying to steal from them already have, and a decade-old rule-based system has no chance against a deepfake voice or a synthetic identity built by the same kind of model the bank is now trying to deploy in its own defense.
How these systems actually work
Strip away the vendor language and the mechanics are fairly consistent across most deployments. A machine learning model scores a transaction or a document against patterns learned from historical data, flagging anything that falls outside the expected range. A natural language layer lets the system read an unstructured document, a pay stub, a contract, a customer message, and pull out the specific facts that matter. An orchestration layer then lets several specialized agents pass work to each other: one agent collects documents, a second checks them against a rule set, a third scores the remaining risk, and only the cases that genuinely need judgment land in front of a person.
That handoff design is what actually produces the productivity numbers cited earlier. The ten-agent KYC squads aren’t replacing the compliance analyst. They’re removing the part of the job that was never really analysis in the first place, the manual collecting and cross-checking that used to eat most of the analyst’s day. What’s left for the human is the part of the job that was always the actual job: judgment calls on the cases that don’t fit a clean pattern.
The detail that separates a well-built deployment from a fragile one is exactly where that human checkpoint sits. A system that only brings in a person after a decision is already made, a payment already sent, a customer already declined, isn’t really human-in-the-loop. It’s human-after-the-fact. The stronger deployments route the escalation before the irreversible step, not after it. That costs a little speed, but it means a person can actually catch a mistake instead of just filing a report on one.
The table below summarizes where each of these technologies currently stands.
| Technology | Maturity | 2026 adoption or impact | Example |
| Predictive ML (fraud, credit risk) | Mature, over a decade in production | 90% of institutions use AI for fraud detection | JPMorgan, ~$1.5B in losses prevented annually |
| Generative AI (copilots, chat) | Established, scaling fast | 3.2 billion+ logged interactions on one platform | Bank of America’s Erica |
| Agentic AI (multi-step workflows) | Early, but accelerating sharply | 70% of firms deploying or exploring it; 600%+ growth in finance team usage | A global bank running 10 coordinated KYC agent squads |
| AI inside digital asset compliance | Emerging, regulatory sandboxes only | Piloted across Singapore and wider Asia-Pacific | Continuous monitoring of stablecoin and tokenized asset transfers |
Four problems banks haven’t solved yet
None of this comes with the risk fully priced in, and it’s worth naming exactly where the open problems sit.
The first is dual-use: the same generative AI strengthening a bank’s defenses is simultaneously strengthening an attacker’s offense, and the two are advancing at roughly the same speed. The second is technical, and harder to explain to a board: large language models are nondeterministic, meaning the same prompt can produce different outputs on different runs, and they hallucinate, generating plausible but incorrect statements with no built-in signal that anything is wrong. That property matters a great deal anywhere an agent is authorizing a payment or signing off on a compliance check, rather than just drafting a memo a human will review anyway.
The third is regulatory, and it’s the most uncomfortable one for a compliance officer to sit with. The newest US model risk guidance, issued in April 2026, explicitly declines to cover generative and agentic AI, the exact systems banks are deploying fastest, and instead tells institutions to apply older risk-management principles in the meantime. Banks are being asked to govern the newest, least understood part of AI in banking using a rulebook written for an older, simpler one, and to do it before the regulator finishes writing rules for the thing actually being deployed.
The fourth sits underneath all three: accountability. When a rule-based system declines a loan, the rule that triggered the decline is documented somewhere, and a person signed off on it. When an agentic system declines the same loan after weighing dozens of inputs through a model that updates its own weights over time, tracing exactly why is harder, and proving it to a regulator or an unhappy customer is harder still. FINRA’s 2026 oversight report flags this directly: the supervision duty doesn’t shrink just because the decision-maker is an agent rather than a person. It gets harder to actually carry out.
None of this is a reason to sit on the sidelines. The institutions earning that 2.84x return aren’t the ones that found a way around these risks. They’re the ones that built governance, audit trails, and human checkpoints into the deployment from the start, instead of bolting them on after something went wrong. AI in banking, at this point, isn’t really optional. The open question for any given institution is whether it’s being deployed with that discipline, or just deployed.
Frequently Asked Questions
1. What is AI in banking?
An umbrella term for three technologies: predictive machine learning (fraud detection, credit risk), generative AI (chatbots like Bank of America’s Erica), and agentic AI (systems that plan and execute multi-step tasks on their own). Each is at a different maturity level.
2. Is AI actually reducing fraud, or just shifting it elsewhere?
Both. JPMorgan’s predictive AI system prevents an estimated $1.5 billion in fraud losses a year, but generative AI is also making fraud easier to commit, through deepfake voices and synthetic identities. Deloitte projects AI-enabled fraud could hit $40 billion in the US by 2027.
3. What’s the difference between agentic AI and a regular chatbot?
A chatbot answers a question. Agentic AI plans and executes a sequence of actions on its own, collecting documents, checking them against rules, escalating exceptions, closing the case, without a person triggering each step.
4. Why do some banks get a higher return on AI than others?
According to IDC, banks treating AI as a redesign of how decisions get made earn a 2.84x return, while banks treating it as a simple software purchase earn 0.84x. The gap usually comes down to senior ownership and data governance, not the technology itself.
5. Can an AI agent be held accountable for a wrong decision?
Not the way a person can. Tracing exactly why an agentic system reached a decision is harder than tracing a documented rule, which is why human checkpoints remain built into stronger deployments.