Banks are hiring quickly for AI, and one of the groups they are adding is the one that checks it: staff who work on AI governance at the biggest banks grew by a third in a year. The banks that rank highest for AI maturity also added AI product managers, engineers and risk specialists. The news reads like a story about banks being careful. It is better read as a preview of where the work goes in every company that puts AI near money.
Checking grows with the AI, not with the headcount
When an AI gives a wrong answer in a chat window, someone rewrites it and moves on. When it gives a wrong answer about a loan, a payment or a fraud alert, a customer is harmed and the bank is out of pocket. So every workflow the AI enters needs a person who decides what a right answer looks like, looks at a sample of the output, and can stop it.
That need grows with how deep the AI goes, and banks are going deeper. Nearly one in three of the new AI projects banks announced in the first three months of 2026 was an agent that carries out a task, up from 15% in the last quarter of 2025. Most are built into one function, such as handling a service request, instead of being a general assistant for the whole company. An assistant only suggests, so a person owns the result. An agent acts, so someone has to own what it did.
The number of banks that can point to a return on their AI went from 8 to 12 in a year, and the top-ranked banks are the ones that hired to run AI in daily work instead of trying it in pilots.
The same job at a much smaller scale
Take a distributor that lets an AI answer refund requests. The owner decides that the AI may settle any case under $100 by itself, that everything bigger goes to a person, and that every Friday someone reads 20 cases chosen at random out of about 200. That reading takes under an hour. If one case in fifty is wrong and the average refund is $60, the mistakes cost about $240 a week, and the Friday review is how the owner finds the pattern behind them within a few weeks instead of at the end of the year.
Notice who does the reading. It is not an engineer. It is the person who has handled refunds for years and can tell a fair decision from a careless one. A bank's governance team does the same thing with more paperwork: set the limit, sample the output, keep a record, and stop the system when it drifts.
What the teller story suggests
The teller kept a job because the machine made branches cheaper and the human part of the work became the valuable part. The same logic points at the person who knows the work. An AI that does the routine cases makes the unusual cases the part customers and regulators care about, and the people who know those cases are the ones a company cannot scale AI without. Capital One, which ranks second among banks, built its AI governance before it deployed agents, and that order is the one smaller companies can copy.
Before you switch on AI for anything that touches money, decide who signs for its work. If you can name that person and what they will look at each week, you are ready. If you cannot, that gap is the first thing to fix.