Definition
AI in finance is used to detect fraud, score credit risk, automate document checks, and support trading and customer service. Because it affects people's money and access to loans, it's closely regulated.
In simple words
When your card is blocked after a strange purchase, a model flagged it in milliseconds. When a loan is approved in minutes, a model likely scored your risk.
Why it matters
It's one of the largest users of ML, and an excellent lesson on imbalanced data and fairness, because fraud is rare and loan decisions can be biased.
Key terms
- Fraud detection
- Spotting payments that are likely to be dishonest, usually while they are happening.
- Anomaly
- Something that does not fit a person's usual pattern, such as a large payment at 3 a.m. from a new phone.
- Risk score
- A number a model gives to a payment or a person, showing how risky it judges them to be.
- Credit score
- A number summarising how reliably a person has repaid money in the past.
- UPI
- Unified Payments Interface: India's system for instant payments between bank accounts, launched in 2016.
- Social engineering
- Tricking a person into handing over money or secrets, instead of breaking into a computer.
Explained in detail
A check faster than a blink
India's UPI system was launched in 2016 and now carries billions of payments every month. Each one must be approved or stopped within about a second. No team of people could inspect payments at that speed, so banks and payment companies use models. In the instant after you press "pay", a model looks at the details of the payment, compares them with your usual behaviour and with known patterns of fraud, and produces a risk score. A low score lets the payment through. A high score blocks it or asks you to confirm.
| Clue the model looks at | What might raise the risk score |
|---|---|
| Amount | Far larger than you normally send |
| Time | An hour when you are usually asleep |
| Device | A phone never used with this account before |
| Place | A city far from where you were an hour ago |
| Receiver | A new account that many people have reported |
| Speed | Many payments one after another in minutes |
Why this is a hard problem
Fraud is rare. Among many thousands of payments, only a few are dishonest, so a model that flagged nothing would look almost perfectly accurate while being useless. The two kinds of mistake both hurt. A missed fraud costs someone their savings. A false alarm blocks a genuine payment, perhaps a hospital bill at midnight. Banks must choose how cautious to be, knowing that no setting removes both errors. And fraudsters change their methods constantly, so a model trained on last year's tricks must be retrained to catch this year's.
The fraud that AI cannot stop
A fraud model checks whether a payment looks like you. It cannot protect you when you yourself are persuaded to send the money. Most losses today come from social engineering: a caller pretending to be from the bank, a fake customer-care number, a message about a prize, a threat of arrest from someone claiming to be a police officer, or a voice that sounds like a relative in trouble. The payment comes from your phone, at your usual place, approved with your PIN, so to the model it looks genuine. Here your own caution is the only defence.
- You never need to enter your UPI PIN to receive money. A PIN is only for sending.
- A bank will not ask for your OTP, PIN or password by phone or message.
- Pressure to act "right now" is the clearest sign of a scam. Stop and call someone you trust.
- In India, report financial fraud at once by calling 1930 or at cybercrime.gov.in.
Deciding who gets a loan
Banks also use models to estimate whether a borrower will repay. In India a person's credit score runs from 300 to 900 and is built from their history of repaying loans and card bills. Models now add other signals, which can help people with no borrowing history to get a first loan. It also raises questions of fairness. If past lending favoured some groups, a model trained on that history can repeat the pattern. A person who is refused a loan should be able to learn the main reasons and ask for the decision to be looked at again.
How it works
| Use | Technique | Key challenge |
|---|---|---|
| Fraud detection | Anomaly detection, boosting | Fraud is rare, and fraudsters adapt |
| Credit scoring | Logistic regression, boosting | Must be fair and explainable |
| Document checks | OCR, LLMs | Accuracy and privacy |
| Trading | Time series models | Markets change, and overfitting is easy |
Try it yourself
Run the bank's fraud check
Slide to decide when the bank should step in. Watch all three numbers.
Should this payment be flagged?
You are the fraud model. Fraud systems look for what is unusual for this person. A flag means "check this", not "this is fraud".
Weekly groceries at the usual store
A card used in another country ten minutes after a purchase in Pune
Ten large UPI payments to a new account in five minutes at 3 am
₹40 for tea at the shop near home, as on most days
The same rent payment on the same date as every month
Someone who only ever spends small amounts suddenly buys a ₹90,000 phone online
Common mistakes
- Using accuracy for fraud. With 0.1% fraud, a model that flags nothing is 99.9% accurate.
- Believing a backtest that looks perfect. It has very likely overfit.
Check yourself
Summary
- Fraud, credit, documents and trading are the main uses.
- Imbalance and drift make fraud hard.
- Fairness and explanation are required.
Finished reading?