If you read finance-and-AI coverage in the business press, you'll think the story is all about algorithmic trading, robo-advisors, and an imminent takeover by hedge fund robots.
Spend a week inside a large bank, and the real story is much more boring, much more human, and — in my opinion — much more interesting.
The real story is risk officers reading fewer reports, compliance teams handling more volume without hiring, and call-centre staff getting better support from their tools. Glamour-free. Useful.
Where AI is genuinely earning its keep
Fraud detection. This is the unshakeable classic. Every retail bank now runs transaction-level ML models that decide, in milliseconds, whether to approve or question a card payment. The best of these have reduced fraud losses by 20-40% while also — crucially — reducing the false-positive rate that used to decline your card at the hotel checkout.
Anti-money-laundering. Regulators require banks to watch for suspicious patterns in their transactions. Legacy systems flag millions of benign transactions and drown compliance teams. Modern ML systems are much more precise — same risk coverage, a fraction of the false positives. Compliance officers at several of our clients have told me, unprompted, that this has saved their sanity.
Document understanding. Banking runs on documents. Contracts. KYC paperwork. Loan applications. Insurance claims. Language models that can read, classify, and extract structure from these documents have quietly replaced what used to be offshore processing centres — and the quality is often better.
Credit underwriting. Traditional credit scoring works, but it excludes people with thin credit files. Modern ML models can incorporate cashflow data, rental payments, utility bills — giving previously "unscorable" customers access to credit responsibly. Done well, this expands financial inclusion. Done badly, it introduces new biases. The difference is in the team's discipline.
Where it's genuinely not
Algorithmic trading. Quantitative funds have been using ML for decades. For retail investors, "AI-powered trading" products are mostly marketing. The signal-to-noise ratio in markets is brutal, and the organisations that actually have edge are paying their engineers more than you and me combined. Don't let someone sell you "AI stock picks." That's a tell.
Robo-advisors. Surprisingly little AI in most robo-advisors. The allocation logic is usually a few rules and tax optimisation — which is fine, and valuable, but it's not AI. Call it what it is.
Customer-facing chatbots that handle complex advice. Banks are experimenting with this. It's going to take a while to get right. The failure modes include accidentally committing the bank to financial products it didn't offer, which is the sort of thing regulators get cross about.
A story from a mid-sized bank
We worked with a mid-sized European retail bank on their compliance transaction monitoring. Their legacy system was flagging around 4% of all transactions. Their compliance team was reviewing each flag manually. 98% of the flags were false positives. Analysts were miserable. Regulators were grumpy about the backlog.
We didn't replace the system. We added a second-stage model — trained on the analysts' own labels over the previous two years — that re-ranked the original flags by likelihood of genuine suspicion.
The analysts kept working from the top of the queue. Same screens, same process. But now the top of the queue was actually interesting. The false-positive rate dropped to around 40%. The backlog cleared in a quarter. Nobody lost their job. The analysts told us they now had time to investigate the interesting cases properly, rather than rubber-stamping benign ones.
Six months later, the head of compliance told me: "I didn't realise, when we started, that you were mostly going to make my team happier. That's been the best outcome."
What we tell financial services clients
A few things we end up saying in the first meeting:
- Don't put an LLM in the critical decision loop without careful evals. Hallucinated facts in a loan decision are a regulatory problem.
- Model risk management is your friend. The frameworks banks use to govern model risk (SR 11-7 in the US, SS 1/23 in the UK) are actually pretty sensible. Lean in, don't fight them.
- Invest in monitoring. Financial models drift. The world changes. Your monitoring should tell you faster than the regulator does.
- Fairness testing isn't optional. Credit and insurance decisions affect people's lives. Test for disparate impact. Keep testing.
The wider picture
The best financial-services AI teams I know are the ones where risk, compliance, and data science teams eat lunch together. That sounds trivial. It isn't. The technical and the regulatory both have to be in the same room when decisions are made, or you end up with either clever systems you can't deploy, or deployed systems that shouldn't have been.
If you're thinking about AI in a financial services context — happy to swap notes. Book a chat.
