If I had to hire a junior in marketing or finance tomorrow, I'd weigh their Claude use case heavier than their school. Maybe more.
That's a strong statement coming from someone who has spent fifteen years recruiting out of EDHEC, INSEAD, X, HEC, Sciences Po. The pedigree filter has worked. It still works for some roles. But for a junior, in 2026, it has stopped being the right one.
What changed
A junior's first three years used to be a slow accumulation of operational reps. They'd write a hundred briefs, build a hundred models, pull a thousand data points. The diploma was a proxy for the speed at which they'd absorb those reps and start producing real output.
AI has compressed that curve. A junior who knows how to delegate to Claude, structure their context, iterate on prompts, and verify outputs is producing senior-quality work in week three. A junior without that fluency is still copying templates from old decks in month nine.
The gap between those two profiles is bigger than the gap between an HEC grad and a regional school grad. The AI gap dwarfs the diploma gap.
What a "good use case" actually looks like
"I use ChatGPT" is not a use case. It's table stakes — and probably means they're using the default model on a free plan, missing 80% of what's available.
A real use case has four traits.
It saves measurable time.
"I rebuilt our weekly competitive monitoring report from a 4-hour manual scrape into a 15-minute Claude workflow." That's a use case. There's a baseline, a delta, and the candidate can describe both honestly.
It's specific to a workflow they own.
"I summarise meetings" is generic. "I built a pipeline that takes our raw Notion meeting notes, extracts action items by owner, and pushes them to Asana with deadline inference" is specific. The first one shows familiarity. The second shows ownership.
It involves more than one prompt.
Real use cases are workflows, not single prompts. Look for: chained prompts, system prompts they wrote themselves, reusable templates, agents or skills they configured, an evaluation loop they built to check the output. Bonus points if they ever shipped a custom GPT, a Claude project, or a Cowork plugin.
They've broken it, fixed it, and learned from it.
Anyone serious about AI has a story about a workflow that hallucinated a number into a board deck, or a draft that needed three rewrites because the context was wrong. The candidates who can name the failure mode in detail are the ones to hire. The ones who can't either haven't shipped anything real, or are still over-trusting the model.
How to test for it in 30 minutes
Replace the case study round with this. Send the candidate a real, anonymised problem from your business — three days before the interview. Tell them they can use any AI tool. Tell them you want to see their working, not just the answer.
In the interview, ask three questions :
- Walk me through your first prompt. What context did you provide? What did you not provide, and why?
- Where did the model get it wrong? What did you do about it?
- What would you automate next time? If you had to do this every week, how would you set it up?
You'll learn more about how this person thinks in thirty minutes than from any case interview built on artificial constraints. And you'll filter out, fast, the candidates who are LLM-curious but haven't actually integrated the tool into how they work.
Why this matters for the team you're building
A junior hired on AI fluency in 2026 will, by 2028, be operating at the level a senior operated at in 2023. That's not optimism — it's the same compounding logic — a thread YC and a16z each underline in their 2026 build-lists that made spreadsheets-fluent analysts dominate the 1990s and Python-fluent ones dominate the 2010s.
The mistake is treating AI fluency as a "nice to have" while continuing to optimise for the old signals. That mistake is expensive in two ways. You hire slower people, and you signal to the AI-fluent ones that your company isn't where their capability will compound.
The diploma was a proxy. The AI use case is the actual thing.
One caveat
None of this replaces judgement, written communication, or numerical fluency. A great Claude workflow operated by someone who can't structure a problem still produces nicely formatted nonsense. The bar isn't lowered — the toolset is just different.
But conditional on the basics being there, the AI use case is now the highest-signal data point you can collect on a junior candidate. Use it.