UBS now makes AI skill a hiring test for junior bankers
Two stories worth your Sunday night: a Swiss bank has quietly turned AI fluency into an entry requirement, and OpenAI published the first real numbers on how far its own researchers have handed their work to agents.
UBS will now ask graduate and intern candidates in global banking and markets to prove they can use AI tools before it hires them. The Financial Times reported that the requirement, which applies to candidates joining from 2027, asks applicants to show they can use AI to improve outcomes and efficiency — meaning recruiters are testing it alongside the old filters of grades, internships and finance promise. It is one of the first times a major financial institution has put AI literacy on the entry ticket rather than in an onboarding deck.
The context makes it sharper than a training story. Goldman Sachs and JPMorgan have spent the last year signalling the opposite move — shrinking or freezing junior classes as AI absorbs analyst grunt work — while UBS is keeping the pipeline and raising the bar at the door. The bank's own careers pages already describe an internal AI Fluency Pathway for graduates and say more than 21,000 employees had earned its AI Citizen badge by July 2026, so the hiring requirement is an extension of a program that has been running for a while, not a press-release pivot.
Here is the thing nobody at a bank wants to say out loud: junior banking used dull work as a teaching device. Building comps, formatting decks, grinding through data room after data room gave analysts pattern recognition years before they had judgment. If agents take that work, someone has to explain where judgment comes from now. UBS's answer — hire people who can already direct the tools, then train them — is more honest than pretending the apprenticeship is intact, but it pushes the cost onto students who now need Excel, valuation and calibrated AI skepticism before their first day.
OpenAI says it hit its "automated research intern" goal, and the internal numbers are the first hard look at what recursive self-improvement actually measures. The company's median researcher went from modest agent use at the start of the year to more than $600 a day of inference at API prices by mid-August, with the 90th percentile above $7,000 a day. Across the research org, OpenAI now runs 3.1 agent-workdays for every human workday of labor. It is on track, it says, for a fully automated AI researcher by March 2028.
The caveats matter as much as the numbers. Success rates are rising across task difficulty buckets, but over half of successful tasks with a four-to-eight-hour human horizon still needed at least one human intervention, and high-level planning remains a sliver of agent output — the agents are writing infrastructure code, debugging runs and watching training jobs, not choosing what to work on. OpenAI also disclosed that when it restricted Astra compute for safety reasons in August, researchers simply moved the compute to other model classes, which offset about 85 percent of the decline.
Read alongside chief scientist Jakub Pachocki's essay published the same day — "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed," and he expects voluntary slowdowns to become commonplace — the message is a company publishing its own acceleration data while arguing the brakes need to be social, not just internal. That is a more defensible position than it sounds, and we argued in The Take — OpenAI's disclosure framework will fail, and the company knows it that voluntary reporting without outside verification is where this story has been stuck all week.
What to watch: whether other banks copy the UBS hiring screen or keep shrinking classes instead — and whether OpenAI's promised RSI reporting becomes a quarterly habit or a one-off.
If a bank hands its entry-level grunt work to agents, where is the next generation of analysts supposed to learn judgment? Tell us in the comments.
