AI in Finance · Lesson 02

The Application Tour

Five jobs wear the "AI" label in finance. Under the label they are wildly different — different machines, different track records, and very different odds of actually working. This tour gives you the map, and one new tool: the fit test.

The win By the end of this page you can hear any of the five big applications named — fraud, credit, trading, robo-advice, document work — and say which kind of AI it really is, how well AI fits that job, and the one smart question to ask about it.

First, 30 seconds of Lesson 01

From memory, before reading on: what are the two kinds of AI, what does each eat and produce, and what is each one's signature failure? Say it out loud. (Answer's in Lesson 01 and the glossary if you blank — but try first. The reach is the exercise.)

The new tool: the fit test

Not every job suits machine learning equally. Three questions predict whether an AI application is on solid ground, long before you see any vendor's accuracy claim:

  1. Is there lots of labelled history? ML learns from examples. Millions of past transactions marked fraud/not-fraud is a feast; a handful of ambiguous cases is starvation.
  2. Is there a clear right answer? "Was this transaction fraudulent?" has a yes/no ground truth. "Was this a good investment?" depends on horizon, luck, and what you compare it to.
  3. Does the future resemble the past? ML assumes tomorrow rhymes with yesterday. Card-fraud patterns mostly do. Markets — where every player adapts to every other player — often don't.

Three yeses: AI probably genuinely works there. Each no: subtract confidence. This one test explains most of what you'll see on the tour.

The five stops

Stop 1 · Predictive ML

Fraud & anomaly detection fit: excellent

Models score each transaction against learned normal patterns and flag the odd ones in milliseconds. The oldest and best-fitting use in the book: oceans of labelled examples, a clear yes/no target, and fraud patterns that repeat. That's three yeses. The GAO lists it among the most established uses in banking GAO 2025.

Built-in risk: false positives — legitimate customers blocked — and bias in who gets flagged. Smart question: "What's the false-positive rate, and who reviews the flags?"

Stop 2 · Predictive ML

Credit scoring & underwriting fit: good, but high-stakes

Models estimate default risk to decide who gets a loan and at what price. Technically a good fit — plenty of repayment history, a clear target (did they repay?). The trouble isn't accuracy, it's stakes: the training data encodes decades of past lending discrimination, so the model can reproduce bias through innocent-looking proxies like postcode. And when the law demands a reason for a denial, a black-box model struggles to give one GAO 2025.

Built-in risk: algorithmic bias; black-box denials the lender can't explain. Smart question: "Can you explain an individual denial, and has the model been tested for disparate impact?"

Stop 3 · Mostly rules, some ML

Algorithmic trading fit: weakest of the five

Most "algo trading" is execution plumbing — rules that slice big orders and route them fast, no learning involved. The ML part — predicting price moves — faces the worst possible fit-test scores: no stable ground truth, and a future that actively stops resembling the past, because markets are adversarial — every pattern found gets traded away by the people who found it. That's why brilliant backtests keep failing live: overfitting is the house specialty here. The CFTC's concern is less about any one firm's model and more about many firms' models moving together CFTC 2024.

Built-in risk: overfit strategies; herding — similar models amplifying the same swing. Smart question: "How did it perform after the backtest, out of sample, live?"

Stop 4 · Rules wearing an AI label

Robo-advice fit: fine — because it barely uses AI

You answer a risk questionnaire; software maps you to a model portfolio of cheap index funds and rebalances on schedule. Useful and cheap — but under the hood it's mostly rules and standard portfolio math, decades old, with little learning anywhere. This is the tour's prime site of AI-washing: the "AI" in the pitch is often the questionnaire's if-then table. The advice is templated, not bespoke.

Built-in risk: mislabelling — paying an AI premium for a lookup table; advice too generic for your situation. Smart question: "What exactly does the model learn, and what would change if you removed the word AI?"

Stop 5 · Generative AI

Document work fit: genuinely new, needs a checker

Summarising 200-page filings, drafting reports, answering customers in chat, helping staff search policy manuals. This is the post-2022 wave — the FSB names document summarisation, information retrieval, and code generation as the genuinely new use cases FSB 2024. The capability is real; the catch is the signature failure you already know: fluent, confident hallucination. So every serious deployment keeps a human between the output and anything that matters.

Built-in risk: hallucinated facts flowing into reports, filings, or customer answers. Smart question: "Who checks the output before it's used, and what happens when it's wrong?"

The pattern to keep Notice the shape of the tour: AI works best where the job is boring, repetitive, and data-rich (fraud), and worst where the headlines are most breathless (beating the market). The excitement of a use case and its actual fit are close to inversely related. That inversion is one of the most reliable hype detectors you'll ever own.

Check yourself

From memory — a miss now is a memory later.

Now run the fit test in the wild

Pick two of the five applications and find a real vendor page or news story for each (search "AI fraud detection platform", "AI trading", "robo advisor" — anything current). For each, work through:

The drill — attempt before opening
  1. Which stop is this? And which kind of AI — predictive, generative, or rules wearing the label?
  2. Run the fit test. Labelled history? Clear right answer? Future like the past? Score it out of three.
  3. Ask the stop's smart question of the page. Does the vendor answer it anywhere? (Usually: no. Note what's missing.)
Did it work? Bring both write-ups back to your teacher — especially any claim that wouldn't sort cleanly into one stop, or a fit-test question you couldn't score from what was written. Those edges are where the next lesson comes from.

Read this next

Primary source: US GAO — Artificial Intelligence: Use and Oversight in Financial Services (2025). The plainest-English tour of who actually uses what, written by an auditor with nothing to sell. Skim the use-case sections and notice how closely their benefits/risks framing tracks the five stops above.