Most people who fail a case interview are not bad at business. They are bad at being watched while they think. The interviewer is not marking your final number. They are marking whether a messy, underspecified problem becomes an organised one in your hands, out loud, in real time. That reframe changes your prep: stop memorising frameworks, start rehearsing a process.
Who runs these, and where
The classic case interview belongs to strategy consulting: McKinsey, BCG, Bain, the Big Four strategy arms (Strategy&, Monitor Deloitte, EY-Parthenon) and boutiques like LEK and Kearney. It has spread into BizOps and Chief of Staff roles inside tech companies, corporate strategy at banks and retailers, and into product, analytics and data science, where it appears as a whiteboard interview: size this market, define a metric for this feature, explain last month's 12 percent drop in signups.
Anatomy of a case, minute by minute
A 30 to 40 minute case usually breaks down like this.
| Phase | Time |
|---|---|
| Prompt and playback | 2 min |
| Clarifying questions | 2-3 min |
| Structuring pause you asked for | 60-90 sec |
| Structure and hypothesis | 3-4 min |
| Analysis and maths | 12-15 min |
| Exhibit or chart | 5 min |
| Recommendation | 2 min |
Candidates underweight the playback. Thirty seconds of "so the client is a grocery chain, revenue is flat, and they want to know whether to enter meal kits by 2027?" prevents twenty minutes spent on the wrong problem.
Three clarifying questions worth asking
- What does the client mean by success, and over what horizon? "Growth" could mean revenue, margin, share or exit valuation. A 12-month target and a 5-year target lead to different recommendations.
- What is the scope? Geography, product line, segment, time frame. "Is this the UK business only, or the whole European footprint?"
- What data do we have? It signals you will be evidence-led, and tells you which branches are explorable.
Three is the ceiling. Ask seven and you are visibly stalling. If a question would not change what you do next, skip it.
Build a structure, do not recite one
"Profitability equals revenue minus cost" is not a framework, it is arithmetic, and it is a perfectly good backbone. Revenue splits into price and volume, cost into fixed and variable. What gets marked down is reciting 4Cs, 4Ps or Porter's Five Forces at a problem that did not ask for them. The interviewer has heard it four hundred times, and it reads as pattern-matching.
Concrete prompt: a bakery chain with 40 stores has seen operating margin fall 6 points in two years. Why, and what should they do?
A canned answer lists competitors. A tailored one sounds like this:
- Revenue per store: has basket size fallen, or footfall? Has the mix shifted from high-margin pastry to low-margin bread?
- Input costs: flour, butter and energy all moved hard in that window. What share of COGS are they, and did the chain pass any through?
- Labour: wage inflation plus staffing hours. Bakeries are labour-heavy at 4am.
- Footprint: are the 6 points an average hiding a tail of 8 bad stores?
Then the hypothesis: "My instinct is input-cost pass-through rather than demand, because margins move faster than habits. I want to start with the cost side." That sentence is worth more than the four buckets: it shows you have a view and will risk being wrong.
MECE, without the worship
MECE means your buckets do not overlap and together they cover the problem. Splitting customers into "new" and "returning" is MECE. Splitting them into "young", "urban" and "price-sensitive" is not: one person can be all three, and plenty are none. Do not say the word, just do it.
The maths is where people actually lose
Confident structuring collapses the moment someone multiplies 4,800 by 35 under pressure. Drill market sizing, growth rates, breakeven (fixed cost over contribution per unit), weighted averages, and converting per-day to per-year without slipping a zero.
Work to two significant figures. Round openly. Nobody wants 4,873,412; they want "about 4.9 million". State assumptions before you use them, not after.
A worked market sizing
How many flat white coffees are sold per day in London?
- Population: round to 9 million.
- Assume 40 percent are regular coffee-shop buyers, so 3.6 million people.
- Assume 0.7 cups per person per day, since not everyone buys daily: roughly 2.5 million cups.
- Assume flat whites are 20 percent of the espresso mix: 500,000 per day.
Sanity-check aloud: "500,000 a day is 180 million a year. At 3.50 a cup that is a 630 million market for one drink. High but not absurd. If I am wrong it is probably the 0.7 figure." That last line is what scores. Name the shakiest assumption yourself.
Thinking out loud: lines you can copy
The cardinal sin is silent calculation. Twenty seconds of scribbling is twenty seconds where the interviewer sees nothing and cannot help you. Slip a decimal in silence and you are marked wrong. Slip it out loud and they often correct you.
- "Let me take 30 seconds to structure this."
- "I want to look at three areas, and I will start with the one I think matters most."
- "That gives us roughly 40 million, which suggests volume is not the problem."
- "I want to check one assumption before going further."
- "4,800 times 30 is 144,000, plus 4,800 times 5 is 24,000, so 168,000."
- "That result surprises me. Let me re-check the denominator."
- "Stepping back, here is what we know so far."
Reading the exhibit in four steps
Ten seconds of silence, then narrate:
- Title. "This is revenue by segment, 2021 to 2025."
- Axes and units. Millions or thousands? Percent or absolute? Truncated axis?
- Outlier. The line, bar or cell that does not behave like the others.
- So what. "Segment C is the only one shrinking, and it was 45 percent of revenue in 2021, so it explains the decline."
The recommendation, in 30 seconds
Answer first, always. Build to a conclusion like a mystery novel and you fail the format.
"My recommendation is that the client should [action]. Two reasons. First, [quantified reason]. Second, [strategic reason]. The main risk is [risk], because [why]. My next step would be [specific action with a time frame]."
Worked: "Close the bottom 8 stores and reprice pastry up 8 percent. Those 8 stores explain 4 of the 6 points of margin decline, and pastry demand looks price-inelastic. The main risk is brand damage in the towns we exit, so my next step is a two-month price test in 5 stores."
"It depends" is a non-answer. They are testing whether you will take a position on incomplete data, which is the job.
The behavioural half nobody prepares
At McKinsey it is the Personal Experience Interview, roughly half the assessment, and people spend 40 hours on cases and 40 minutes on this. They probe personal impact (persuading a resistant stakeholder), entrepreneurial drive (something you started with no mandate), and leadership (a team you moved without authority). Have five stories ready: the situation in one sentence, the stakes, what you did, the measurable outcome. Expect drilling: "what did she say when you proposed that?"
A three-week prep plan
Week 1, foundations. Read two case books. Fifteen minutes of mental arithmetic daily, no calculator. Five solo market sizings, spoken aloud and recorded on your phone. Listening back is uncomfortable and it is the fastest feedback you will get.
Week 2, live cases. Find a partner and do 8 to 10 full cases, alternating roles. Peers interrupt you, which is what real interviewers do. After each one, ask: "where did I lose you?"
Week 3, pressure. Six more live cases, at least three with a stranger. Rehearse your five stories. One full mock in your target firm's format, clock running. Twenty-five cases is where most people stop improving.
Failure modes to check yourself against
- Calculating in silence
- Building a structure then abandoning it three minutes later
- No hypothesis, so the analysis wanders
- Arithmetic slips from refusing to round
- Answering a question that was not asked
- Refusing to commit to a recommendation
The whiteboard variant
In product and data roles the case shrinks: 20 to 30 minutes, a whiteboard or shared doc, prompts like "activation dropped from 34 to 29 percent, diagnose it". Less profit arithmetic, more segmentation and metric design. Name a north-star metric and its guardrails, and say how you would test a hypothesis with data you could query.
What stays identical: play back the prompt, structure before you dive, state a hypothesis, narrate the reasoning, commit to one recommendation instead of a menu. Show them the inside of your head while it is working.