Where AI Is Actually Working in Finance (and Where the Hype Outruns the Results)

Ask ten people how AI is used in finance and you’ll get ten versions of the same image: a hedge fund supercomputer printing money. That picture is mostly wrong. The real adoption is quieter, more operational, and in a few places more impressive than the hype.

Here’s where it’s actually working.

Fraud detection and anti-money laundering

This is the oldest and most proven use case. Card networks and banks run machine learning models on every transaction, scoring the probability of fraud in milliseconds. The models learn from billions of examples and catch patterns no rules-based system could: an unusual merchant category combined with an unusual time of day combined with a device the customer has never used. The result is fewer false declines and more real fraud caught. If you’ve ever gotten a text asking “did you just make this purchase?”, that was a model.

Credit underwriting

Lenders have used statistical scoring for decades. What’s new is the breadth of data and the speed. Small-business lenders can now underwrite a loan in hours by pulling bank transaction data, accounting software data, and payment history, and feeding it into a model that predicts default probability. The upside is faster, cheaper credit for businesses that banks used to ignore. The risk is fairness: a model trained on biased historical data will produce biased outcomes, and regulators are paying close attention.

Document processing

Finance runs on documents. Loan agreements, invoices, trade confirmations, KYC forms, insurance claims. Reading, extracting, and reconciling them used to require armies of back-office staff. Language models now do a large share of that work, and the accuracy on structured extraction has gotten good enough that humans review exceptions rather than every page.

Research and summarization

Analysts at banks and asset managers now use AI to summarize earnings calls, scan filings for changes, and draft first versions of notes. This is real and widespread. It is also the area where the limitations show most clearly: models hallucinate figures, miss context, and can’t yet form a defensible investment view on their own.

Where the hype outruns reality

Trading alpha. Quant funds have used machine learning for years, and the good ones do generate returns. But the idea that you can point a language model at the market and have it find mispriced securities is fantasy. Markets are adversarial; any easily discoverable pattern gets arbitraged away. The funds that succeed with ML succeed because of data infrastructure, execution, and risk management, not because of a clever model.

Fully autonomous advice. Robo-advisors exist and work fine for basic allocation. But “AI financial advisor that replaces a human” keeps getting announced and keeps not materializing, because advice involves liability, regulation, and trust that current systems can’t carry.

Predicting the macro. Nobody has built a model that reliably forecasts recessions, rates, or inflation. The data is too sparse and the regime changes too often. Anyone selling you one is selling you something else.

The risks that matter

Three things keep risk officers awake. Model opacity: when a model denies a loan, the lender has to explain why, and some models can’t. Correlated behavior: if every firm uses similar models trained on similar data, they’ll make similar mistakes at the same time. And overconfidence: a model that’s right 95% of the time will be wrong on the 5% that matters most if nobody is checking.

The bottom line

AI in finance is real, valuable, and mostly boring. It’s catching fraud, processing paperwork, and speeding up research. It is not a money printer, and the firms treating it as one are the ones most likely to get burned. The interesting question for the next few years isn’t whether AI will transform finance. It’s which firms will build the controls to use it well.