McKinsey still wants a recommendation. It now wants AI literacy too.

Two different McKinsey rooms

QuantumBlack (McKinsey’s AI and analytics platform) still hires people who can talk training data, leakage, and monitoring. That loop is closer to a data science case. Read the QuantumBlack guide if that is the role on the email.

Generalist McKinsey is different. PEI did not go away. The case did not go away. What changed is that “the client wants to use gen AI on claims / service / pricing” is now a normal prompt, and “use GPT” is a fail. They want the same McKinsey habits — a MECE tree, a number, a so-what — applied to a technology you can explain in English.

• Clarify the decision and the metric. AI is a means.
• Size time, leakage, or conversion before naming a model.
• Say whether you would prompt, retrieve (RAG), fine-tune, or give the model tools.
• Put a human on any action that spends money or changes a customer record until evals say otherwise.

• QuantumBlack case interview guide

What “good” sounds like in the McKinsey close

Lead with the answer. Two supports, one of them a number. Name the main risk in the language of the exhibit — hallucination rate, false pays, adoption — not “ethics” as a slogan. Give a Monday step: a 200-claim eval set, a two-week copilot pilot, a kill if leakage moves 50 basis points.

Solve games (Sea Wolf, Red Rock, SFL) are still the online screen for many offices. They are not the AI knowledge test. Do not confuse the two.

CoachNed is independent and not affiliated with McKinsey. Clearclaim Mutual is original practice in a McKinsey-style interviewer voice.

How to prep this week

If you have never cased, do the ordinary method first. Then spend twelve minutes on the AI lesson so the vocabulary is not new in the room. Then sit one AI case out loud. Another article on QuantumBlack will not fix a wobbly close.

• QuantumBlack case interview guide
• McKinsey Solve practice