A fair question
If AI cannot tell you what a price will do, why are banks and insurers hiring so many people to work on it?
It is a good objection and it deserves a proper answer rather than a dismissal. The answer is that they are using it heavily, spending a great deal on it, and getting real value from it. It is simply that almost none of that work is about seeing the future. It is about handling the enormous volume of text, documents, records and messages that financial firms drown in, and about noticing when something in that flow looks wrong.
Put plainly: the genuine uses are administrative. They would make a terrible advert. Nobody is going to sell you a subscription to reconciling documents.
Fraud and anomaly detection
This is the oldest and most successful use, and it predates the current wave of AI by decades.
A card network sees an unimaginable number of transactions. Somewhere in there are the ones that do not fit: a card used in two distant places within an hour, a spending pattern that breaks sharply from a customer's habits, a sequence of small test payments before a large one. Machine learning is genuinely good at this, because it is a stable problem with an enormous number of labelled examples. Fraud that happened last year still looks like fraud.
Note what the system is doing. It is not predicting anything about the world. It is comparing new activity against an established pattern and flagging what deviates. That is exactly the shape of problem from lesson one that AI handles well.
It also gets things wrong constantly, in both directions, which is why a human sits behind it and why your card occasionally gets declined in a supermarket for no reason you can see.
Reading the paperwork
Finance runs on documents, and there are a staggering number of them. Loan files, insurance claims, invoices, contracts, statements, identity paperwork, annual reports that run to hundreds of pages.
Firms use AI to pull structured information out of unstructured documents: extracting the parties and dates from a contract, matching an invoice against a purchase record, summarising a long filing so an analyst can decide whether it is worth reading in full, sorting an inbox of claims by type. This is the least glamorous work imaginable and probably the largest slice of real value in the industry.
Crucially, the output goes to a person who checks it. A summary of a company's annual report is a reading aid, not a verdict, and the analyst still forms their own view.
Compliance, checks and monitoring
Regulated firms have to know who their customers are, watch for money laundering, screen against sanctions lists, and keep records of what staff said to clients. All of that involves sifting through vast quantities of text and transactions for things that warrant a second look.
AI is used to draft the first pass: grouping alerts, matching names across spellings and scripts, flagging conversations for review, drafting the summary that a compliance officer then reads properly. The regulator holds the firm responsible either way, which is precisely why a human signs off. In most jurisdictions there are rules about explaining automated decisions that affect customers, though they vary considerably, so this is another area where your own regulator is the authority rather than anything you read in general.
The most advanced AI in your bank is probably deciding whether a scanned document is page four of a mortgage application or page four of a completely different mortgage application. Revolutionary, in its way. Not remotely filmable.
Customer service and internal search
The chatbot that answers where your statement is, the search tool that lets a call handler find the right policy wording in seconds, the system that drafts a reply for an adviser to edit and send. Useful, widely deployed, and about as thrilling as it sounds.
Checkpoint
Real institutional use is administrative: spotting anomalies, reading documents, running compliance checks and answering routine questions, with a human accountable at the end of every one.
Risk modelling, and what it is honestly for
This is the one that sounds closest to prediction, so it is worth being careful.
Firms model risk: how a portfolio might behave under a range of conditions, how much capital to hold against a set of bad outcomes, how exposures overlap in a stressed market, how likely a borrower is to fall behind. Modelling is a serious discipline and computers are central to it.
But look at what is being produced. These are ranges, distributions and scenarios, not forecasts of a price on a date. The output of a good risk model is closer to "here is the spread of things that could happen and what each would cost us" than to any statement about what will happen. The work is about surviving uncertainty rather than removing it, and modellers are usually the first people in the building to tell you the model is wrong in ways they can partly name.
The history of finance includes some expensive reminders that a model performing beautifully in ordinary conditions can fail exactly when conditions stop being ordinary. That is not a flaw somebody forgot to fix. It follows directly from the fact that models are built on the past.
Real use against the adverts
| The advert says | The reality inside firms |
|---|---|
| It predicts what prices will do | Nothing produced is a forecast of a price on a date |
| It runs by itself while you sleep | A named human is accountable for the output |
| Anyone can buy access today | Systems are internal, audited and never sold as a subscription |
| Results are guaranteed | Outputs are ranges, flagged for review, and often wrong |
| It is the secret the banks do not want you to have | It is mostly document handling and anomaly flagging |
โ Weak prompt
Prompt
How do hedge funds use AI to predict the market? Explain the techniques so I can do the same.
Output
Institutional investors use machine learning on price and alternative data to forecast movements. You could begin by...
The question smuggles in its own answer. It assumes prediction is the thing being done, so the reply obliges with a plausible sounding description of an activity that does not work the way the question implies.
โ Good prompt
Prompt
Explain the main ways regulated financial institutions use machine learning in their day to day operations. For each one, say what the system produces, who reviews the output, and whether it involves forecasting future prices. Do not give me any investment or trading suggestions.
Output
Fraud and anomaly detection produces flagged transactions reviewed by an investigations team, and involves no price forecasting. Document processing produces extracted fields checked by an operations team...
It asks about the actual work and about accountability, which is checkable, rather than asking for a secret that does not exist.
Why this matters for the adverts
Hold the two pictures next to each other.
Inside a regulated firm: narrow tools, enormous quantities of dull text, humans accountable at every step, outputs expressed as flags and ranges, and a regulator asking why a decision was made. In the advert: a single system that knows what happens next, needs no supervision, cannot lose, and is available to you this evening for a monthly fee.
The distance between those two pictures is the point of this lesson. If institutional AI genuinely could see the future, the last thing it would look like is a subscription sold to strangers on social media. What it actually looks like is a compliance officer reading a machine generated summary and disagreeing with half of it.
Checkpoint
Risk models produce ranges and scenarios rather than forecasts, which is a way of surviving uncertainty rather than a way of removing it.
๐ Quiz
Question 1 of 4Why is fraud detection a good fit for machine learning?