Your Bank Is Now an Algorithm. Here's What It's Actually Doing With Your Money.
Written with AI assistance and reviewed by the NorwegianSpark SA editorial team
Somewhere between 2023 and now, the person who decided things about your money stopped being a person.
Not in a dramatic way. Nobody was marched out of a branch. It happened through a long series of individually reasonable automations — fraud checks, then affordability scoring, then transaction categorisation, then a chat window that answers at 3am — until the honest description of a modern bank is: a large model with a banking licence attached.
All the major neobanks are now rolling out AI assistants that analyse spending patterns and suggest optimisations. Revolut's chatbot already handles basic queries, and N26 has been testing proactive financial coaching (Statrys, Banks.eu, 2026).
That is the marketing layer. This series is about the other four.
Layer 1: fraud and transaction monitoring
This is the oldest and by far the most consequential use of machine learning in banking, and it is the one nobody markets — because its failures are more memorable than its successes.
A model scores every transaction in real time against your behaviour and against patterns learned from millions of accounts. Most of the time it is invisible and genuinely useful: it catches the card-testing charge at 4am that you would not have noticed for a week.
The trade-off is that the same model produces false positives, and a false positive in fraud monitoring means your card declines in a supermarket queue in a country you told them you were visiting. Our deeper explainer is AI fraud detection in banking.
Layer 2: identity, onboarding and AML
Opening an account now runs through document verification, liveness checks, sanctions and politically-exposed-person screening, and a risk score assembled before a human sees anything.
This is where most account rejections happen, and it is why rejections are so rarely explained: the reasoning is a risk model plus a regulatory obligation not to tip off. Practical consequences for anyone opening accounts across borders are in how to open a global bank account and bank accounts for expats.
Layer 3: categorisation and the spending picture
The friendly layer. Transactions get classified, merchants get enriched with logos and names, and you get a pie chart telling you about coffee.
It is useful and it is also the layer with the most quiet errors — a transfer to a friend classified as "entertainment", a business expense filed as groceries. Fine for a rough picture, not fine as a bookkeeping source of truth. If you run a business through it, reconcile against the raw statement.
Layer 4: pricing and limits
Less visible and more important. Models influence your credit limit, whether you are offered an overdraft, what rate you see, and how quickly a large incoming payment is released.
Two accounts at the same institution can be treated differently, and the difference is a score neither customer can see. This is not new — credit scoring has done it for decades — but the inputs are broader now and the decisions are faster.
The polite word is "personalisation". The accurate word is "differential treatment based on a model you cannot inspect".
Layer 5: the assistant you actually see
The chat window. In practice it does three things well — answering account questions, explaining a charge, and starting a dispute — and one thing badly, which is anything requiring judgement about your specific situation.
Which of these are worth a subscription fee is the whole subject of Part 3, where we go through them one by one.
What this means for you, practically
Your rights, briefly and concretely
Two things are worth knowing if you are in the EU or UK.
Automated decisions. Article 22 of the GDPR gives you the right not to be subject to a decision based solely on automated processing where it produces legal or similarly significant effects — with the right to obtain human intervention, express your point of view, and contest the decision (GDPR Art. 22). Banks apply exemptions, but the right to ask for human review is real, and asking for it explicitly changes how a complaint is handled.
The EU AI Act classifies AI systems used to evaluate the creditworthiness of natural persons as high-risk, with obligations around transparency, record-keeping and human oversight (Regulation (EU) 2024/1689).
Neither makes the model go away. Both give you a lever, and the lever works better when you name it.
Is this actually bad?
Mostly, no. Automated fraud monitoring catches things no human review could. Automated onboarding is why you can open an account in an afternoon rather than a fortnight. Categorisation makes budgeting possible for people who would never build a spreadsheet.
The problem is not that banks use models. It is that the models are invisible, the appeals process assumes you know it exists, and the failure mode — losing access to your own money with no explanation — lands hardest on people with the least slack.
That is Part 2.
Where to start if you are picking a provider
Our comparisons: best AI banking apps 2026, best AI global banks 2026 and Revolut vs Wise. For multi-currency holding specifically, Wise remains the reference point on transparent conversion, and Airwallex is the one we point businesses at — reviews in Wise review 2026 and Airwallex vs Wise for business.
Sister-site reading: where cash actually belongs by time horizon in the Yield Ladder, and what your bank costs you at the till in The 3% Tax.
---
Written with AI assistance and reviewed by the NorwegianSpark SA editorial team. NorwegianSpark SA, org. 834 984 172. Some links are affiliate links — see our [disclosure](/disclosure). General information, not financial advice.