UBS will screen graduates for AI skill. The questions to ask in your firm

One of the world's largest banks is going to ask job candidates to prove they have used AI on real work.

UBS will require graduates and interns joining its 2027 intake to show they can use AI to improve results and efficiency, and to give examples of how they have used it before. The requirement sits alongside the existing ones rather than replacing them. The Financial Times broke the story, reported on 7 September 2026 by The Next Web and heise online.

UBS is not asking for a list of tools or a certificate. The candidate has to point at a real task and say what changed.

That test costs nothing to copy, and it takes five minutes in an interview you were having anyway. A firm your size hires two or three people a year, so changing your interviews will take years to show. The faster use is on the people you already employ, and the same four questions work for both.

Naming tools tells you nothing

The obvious version of this question is "which AI tools have you used?", and it is worthless. Everyone has used ChatGPT. Listing three more names proves only that they have read an article about AI tools.

Asking for evidence of a changed outcome is much harder to fake and much more useful to you. Somebody who has genuinely used AI on real work can describe the task, what they gave the model, what came back, what was wrong with it, and what they did about that. Somebody who has not will stay vague, because there is nothing specific to say.

The same thing decides who gets value from these tools. It is not the youngest or the most technical: your most experienced people are your best AI users, because knowing the work is what lets you judge the output.

Ask these four questions

Ask them in this order. The whole thing takes five minutes.

1. "Describe a real task you handed to an AI tool. What was the task?"

You are listening for a specific piece of work with a beginning and an end. Examples: a competitor analysis, a spreadsheet reconciliation, thirty applications sorted against a spec.

Good answer: names the task, the context and roughly how long it used to take. Weak answer: "Research and writing, all the time." That is a habit, not a task.

2. "What exactly did you give it, and what did you get back?"

Anyone who has done real work with these tools remembers roughly what they typed and what came back.

Good answer: describes the input, the format they asked for, and what the first attempt actually produced. Weak answer: jumps straight to the result with nothing in between. They have not done it.

3. "What did it get wrong, and how did you know?"

The most important of the four. Someone who says it got nothing wrong has either not checked or not noticed, and both are disqualifying for work that reaches a client.

Good answer: names a specific error, and says how they caught it. Invented citations, a number that did not reconcile, a tone that was wrong for the audience. Weak answer: "It was pretty accurate." Follow up once with "what did you have to change?", then move on.

4. "What would you not use it for?"

You are testing whether they have thought about the limits at all. A candidate who cannot name a single one is a risk in a firm holding client data.

Good answer: names a specific limit. Client-identifiable information, final advice, anything going out without a human read. Weak answer: either "nothing" or a vague worry about accuracy with no example behind it.

The same four questions work on your existing team

Ask them in your next round of one-to-ones. You will find out three things in an afternoon: who is genuinely using these tools on real work, where in your firm it is already helping, and which people are quietly avoiding the whole subject.

If the answers are thin across the board, the problem is unlikely to be your people. Most training on this fails because it teaches work nobody does once they are back at their desk, and we have written about why AI training fails and what to do instead.

Exercises, certificates and enthusiasm all mislead

Do not set a written exercise. They will do it with AI, so it tells you nothing.

Do not ask for a certificate. There are hundreds of them, and none of them shows judgement, which is what you are hiring for.

Do not screen out people who are honestly sceptical. Somebody who has tried these tools on real work and formed a critical view is more useful than somebody who is enthusiastic and vague. Ask question three of both and you will hear the difference.

One more caution. Do not let this outrank whether the person can do the job. UBS added the requirement alongside its existing ones rather than in place of them, and in a firm of fifteen the same holds: this is one signal among several.

FAQ

Is this fair on candidates who have not had access to these tools? It is if you ask about the reasoning rather than the tool. Someone who has never used one can still tell you which part of a job they would hand over, what they would check, and what they would never send out unread. Judge that, not which subscription they happened to have.

Is UBS the only one doing this? No. heise online reports that Santander has also looked for advanced AI users, and that UBS has gone further by writing it into the requirements (heise online).

What if we do not use AI much ourselves? Then hiring for it is premature, and the honest first step is finding out where it would help you. Screening candidates for a capability your firm cannot yet use produces a hire with nothing to do with the skill you hired for.

Could a candidate just prepare good answers to these? They could prepare question one, and question four is easy to rehearse, so treat that one as a tie-breaker rather than a test. Questions two and three are hard to fake, because a rehearsed answer runs out when you ask what the first attempt produced and how they spotted the error.

Should we put it in the job advert? Only if you mean it. If nobody in the firm uses these tools, an advert asking for AI skills will not survive the first question a candidate asks back.

---

If you would like help working out what AI skill actually looks like in your roles, book a conversation.