Banking has the data, the use cases and the budgets, and it shows in adoption. Whether that adoption pays off is a different question.
Finance is, on paper, the ideal AI industry. It runs on data, it has crisp and measurable problems (is this transaction fraudulent? is this borrower a risk?), and it has the money to invest. So if anyone should be reaping clean returns, it's banks. The numbers below show adoption is broad and still climbing, and yet very few banks can point, in public, to realised ROI across their AI programmes. That contradiction is the most useful thing in the data.
The adoption
- In Feedzai's May 2025 survey of 562 fraud and financial crime professionals, 90% of financial institutions said they use AI against fraud, to speed up investigations and catch new tactics in real time.
- Evident's October 2025 AI Index of 50 of the world's largest banks finds "virtually every key metric of AI adoption" moving up, with every bank in it increasing its AI budget.
Fraud detection is the standout because it's the cleanest possible AI problem: a high-volume, pattern-rich task with a clear right answer and immediate feedback. Fewer false positives isn't a soft benefit; every false positive is a legitimate customer's card declined at the checkout, a support call, and a dent in trust.
The ROI gap
- Evident's index works only from public disclosures. As of October 2025, just 4 of the 50 banks (JPMorgan Chase, BNP Paribas, DBS and Santander) reported realised returns across all their AI use cases, up from 2 a year earlier.
- Many banks don't publish firm-wide AI figures at all, so that count measures disclosure as much as results.
Adoption is broad. Public, firm-wide proof of return is rare. Both can be true when most value sits in a few well-run programmes and most banks don't yet measure the rest.
How both numbers can be true
The reconciliation is concentration and measurement. Much of the value comes from a small number of narrow, well-executed programs (fraud detection chief among them) at the institutions that got the fundamentals right. Spread across hundreds of banks and thousands of broader, fuzzier "AI transformation" initiatives, many of which stall between pilot and scale, the average project disappoints even where specific programs pay. It's the same pattern visible across the whole enterprise in the state of enterprise AI: value pools in the disciplined few, not the enthusiastic many.
The lesson hides in which projects work. Fraud detection succeeds because it's narrow, measurable, and has a tight feedback loop: exactly the conditions that let a team prove value and improve. The projects that miss tend to be broad, vaguely scoped "let's add AI everywhere" efforts with no baseline and no single owner. Scope and measurability, not model quality, separate the two.
Why "realised ROI" is so rare
The phrase doing a lot of work in that statistic is realised. Plenty of banks can show a model that scores well in testing, or a pilot that impressed a committee. Far fewer can trace a line from the AI to a number that actually moved on the P&L (cost down, revenue up, or risk reduced) net of what the program cost to build and run. Fewer still publish it. That's a higher bar, and it's the right one. Evident notes that many banks are not yet ready, or able, to publish aggregate figures for their whole AI portfolio, and that the hard step is moving from proof of concept to scaled deployment. Much of the spend sits in initiatives that haven't shipped fully, haven't been measured against a baseline, or haven't run long enough to net out their own integration and maintenance costs. The disappointment is mostly a measurement-and-execution story, not a technology one.
Regulation amplifies the effect in finance specifically. A bank can't just let a model act: every consequential output needs an audit trail, an accountable human, and an explanation a regulator would accept. That human-in-the-loop requirement is exactly right for the stakes, but it also means AI in banking augments expensive experts rather than replacing them, which caps the headline savings and lengthens the path to realised ROI, a pattern our companion piece on generative AI in banking covers in more depth. The institutions that succeed accept this and design for it, rather than chasing an autonomy that compliance would never sign off on anyway, and shipping AI that clears compliance in financial services is exactly that kind of work.
What this means for your team
- Copy the fraud-detection playbook. Narrow problem, clear right answer, fast feedback, measured baseline. Those conditions are what made it pay.
- Be suspicious of broad "transformation" mandates. Value comes from specific workflows with owners and numbers, not sweeping initiatives.
- Budget for the step from pilot to scale. Evident's index describes the largest banks working through exactly that step; plan and resource for it rather than being surprised.
- Define realised ROI up front. If only 4 of the 50 largest banks publish it across their AI portfolio, the discipline of measuring it properly is itself the competitive edge.
Meanwhile, fraud cuts both ways. Feedzai's May 2025 report says more than half of fraud now involves AI, though it doesn't say how it measured that, and 92% of the institutions in its survey said fraudsters use generative AI. It's an arms race, and standing still means falling behind. If you want help finding the narrow, measurable, fraud-detection-shaped opportunities in your own operation, that's a good first conversation to have with our team.
