Written by: Karol Sobieraj, Founder & CEO, Digital Colliers
UBS now requires AI skills when hiring junior investment bankers. It's one of the first big banks to make AI literacy a formal hiring requirement. Most mid-market banks still onboard juniors without knowing which systems they'll touch, which AI tools they already use, or what gaps sit between the two.
Why mid-market banks are flying blind
You can't train someone on tools you don't know they're using. The pattern I keep seeing: a junior arrives with GitHub Copilot muscle memory from university. They write Python scripts to pull AML alerts. Nobody on the desk knows they're using an AI assistant. Nobody reviewed the output.
88% of AI proofs-of-concept never reach widescale deployment. For every 33 AI POCs a company launches, only four graduate to production. Mid-market banks are running that failure rate without even tracking which juniors bring AI tooling through the door.
More than 15% of commits from AI coding assistants introduce at least one issue. That number ranges from 17.4% for GitHub Copilot up to 29.1% for Gemini. If you don't know which juniors are writing code with AI help, you can't scope the review load. You can't build check steps into change control. You're carrying silent risk in every script that touches client data or regulatory reporting.
The training-data model: three pieces
The operators shipping this in 2026 structure it in three layers.
Pre-hire fluency check. During interviews, ask candidates which AI tools they've used in the last six months. Document it. Ask them to describe one time the AI output was wrong and how they caught it. You're not testing technical depth. You're testing self-awareness and quality habits.
System-access map. Before Day One, list every internal system the junior will touch in their first 90 days. Credit models, transaction monitoring, Excel pricing sheets, internal Python libs. Map which systems already use AI and which don't. Flag any system where a junior could sneak in AI-generated code or queries without review.
Skills-gap plan. Compare the fluency check to the system map. Find the gaps. If the junior has never used AI to write SQL and your AML team runs queries by hand, you need a training module before they touch production. If they've only used ChatGPT for homework and your desk uses GitHub Copilot in prod, you need to onboard them to the approved tools.
What this catches in production
42% of companies abandoned most of their AI initiatives in 2025, up from 17% in 2024. The average organisation scrapped 46% of AI proofs-of-concept before production. Most of that failure comes from mismatched expectations and invisible tool sprawl. Juniors arrive with tool habits you didn't document. Senior staff assume juniors know the review rules, but nobody wrote them down. Six months later you're carrying technical debt nobody can trace.
If you run the training-data model, the junior's first week includes a tools audit. You walk them through which AI assistants are approved for which tasks. You show them the review checklist for AI-generated code. You map the approval chain for any new tool they want to try. The junior knows the boundaries. The desk knows what's in production. Risk and compliance can scope the exposure.
Map fluency before Day One
UBS moved first because they have the hiring volume and compliance pressure to justify it. Mid-market banks don't need to copy UBS line for line. You need the data underneath: what juniors already know, what systems they'll touch, where the gaps sit.
Build that map before they show up. Train to the gaps in Week One. Run the tools audit before they write their first line of code. The operators who ship this run it as a one-page checklist. Pre-hire fluency, system access, skills gaps. Takes 30 minutes to fill out. It surfaces the risks you're carrying today without any formal process.
If you wait until the junior's already writing Python for the compliance team, you're debugging problems in production.