How to practise case interviews without a partner

It is 11 pm. Your laptop is propped on a pillow, and you are explaining a profit drop to an empty room. In a few months, someone across a table will ask the same question, and you will already know how to begin.

(An illustrative scene.)

You cannot practise a case interview without something asking you questions. Any AI chatbot can do that part; it cannot mimic what the interview does to you. The rest is method: full cases on a timer, every step spoken, answers checked the way interviewers check them, real pressure before the day. The method is the same whether you are in an MBA programme or recruiting as an undergrad.

So what does a case partner actually do in the room?

They play the interviewer: push back, release data only when you ask the right question, keep time, score you. Each job has a solo replacement.

• Pushback. Record yourself reading the script below, with pauses, and play it back mid-case. One rule: every conclusion gets one “why?”

• Data only when asked. Keep the exhibits covered. Ask the exact question first, then reveal. If your question would not have unlocked the data in a real interview, note it.

• Time. A timer per stage. Cases in our library run 15 to 45 minutes; the discipline matters more than the case.

• Scoring. A rubric and a recording, covered below, replace a partner’s “yeah, that was fine”.

The prompt script, to record once and reuse:

“Tell me how you would approach this.” · “What is your structure?” · “Why that branch first?” · “What would you look at next?” · “What does this chart tell you?” · “Are you sure about that number?” · “So what?” · “What is your recommendation?”

Which cases should you practise?

Six types cover most of what firms ask, and each asks a different question.

Type	What it asks	Watch out
Profitability	Profit fell and nobody says why. Decompose it, isolate the driver, say what to do.	Ask whether revenue, cost or both moved, then go one layer down on that side only.
Market entry	Should the client enter, and how. Attractiveness, competition, entry mode, the numbers.	Company problem or industry problem? If competitors face it too, the response is different.
Market sizing	Estimate a number nobody has. Pick a path, state assumptions, sanity-check.	The right power of ten matters more than exact arithmetic.
Pricing	Set or change a price. Cost is the floor, competition the reference, value the ceiling.	Price moves volume the other way; do the net maths, not just the price.
Growth strategy	Revenue has to grow. Existing or new products, existing or new customers, what is executable.	Separate organic from inorganic, then test each for payoff.
Unconventional	Brain teasers with no standard structure. Structure still wins; you build it on the spot.	Find the constraint first. In these cases that is often the whole answer.

One habit works across all six: keep the top layer fixed (profit always splits into revenue and cost) and adapt the layers below to the case in front of you. For cases, browse the cases section, or start from the knowledge section’s frameworks, key terms and failure modes.

How do you run a mock case alone?

On a timer, with every step spoken, and with the review getting at least as long as the solving. Skip the review and the rep was half the work.

1. Pick the case by weakness. Your lowest rubric score from last time picks the type.

2. Read the prompt once, then clarify. Write the objective in one sentence.

3. Build your structure. Hypothesis first, then three or four branches tailored to this prompt. A framework that fits every case unchanged is memorised, not built.

4. Analyse. Speak every calculation and state every assumption before you use it.

5. Read the exhibit. Say what the chart tells you in one sentence before you touch the numbers.

6. Recommend. Answer first, then the reasons, then the risks.

7. Review the recording. This is where the practice actually happens.

Cases to use alone, free: McKinsey, BCG and Bain publish practice cases on their careers sites, and many universities publish casebooks (Duke and Notre Dame keep public collections). For a large library filtered by type and difficulty, browse the Northstar case library.

Worked example: the profitability checks on one prompt

Take a chai cafe chain - an illustrative example, not a real company: profit fell 20% this year. First question, before any maths: did revenue, cost, or both move? Go one layer down on the side that moved, not both. Revenue splits into footfall per outlet, share of visitors who buy, average bill; cost splits into fixed (rent, staff) and variable (tea, milk, cups). Then the checks: was every number benchmarked before you called it good or bad? Did you segment before hunting for causes? Does the recommendation carry a number the data supports? Said into a recording, that run takes a few minutes and exposes exactly what a silent read-through hides: the segment you forgot, the assumption you never stated.

How do you check your answer?

Review the recording first, the answer key last. The key tells you the answer; the recording tells you how you perform. The first playback can be uncomfortable to sit through, and that is where the review earns its keep. Run five checks, in the order interviewers apply them: did every number get a benchmark against history, competitors or norms? Did you segment before hunting for causes? Is this a company problem or an industry problem? Is the answer the right order of magnitude? And did the recommendation carry a number the data supports?

Then score the five skills the library tags across cases, 1 to 5 each, from the recording and not from memory:

Skill	What a 5 sounds like	Your score
Structuring the problem	A hypothesis up front, tailored branches, no memorised framework recited
Working the numbers	Assumptions stated before use, arithmetic spoken, a sanity check at the end
Reading the data	Every exhibit gets a “so what” before you touch its numbers
Judgement and recommendation	The recommendation leads with the answer and carries a number
Communicating clearly	Signposted, unhurried, no filler

If a stranger following only your audio could not reconstruct your logic, the case does not count as a pass. (Skill names match the five skills our case library tags across cases.)

What do you do after a bad mock?

At some point one will go badly. Treat it as a diagnosis, not a verdict, and fix the specific weakness rather than the whole machine. Structure collapsed? Drill openings only. Maths broke down? Short daily arithmetic drills, spoken, with the fixed-versus-variable-cost basics cold. Communication rambled? Re-record the same case the next day and compare the two takes.

What mistakes do solo candidates make?

The three killers are memorised frameworks, silent maths and unstated assumptions, and all three are invisible on paper and obvious on tape. A memorised framework fails the swap test: transcribe two minutes of your structure, and if changing the company name changes nothing, it was memorised. Silent maths shows up as dead air on the recording, unstated assumptions as numbers with no because-clause. And keep the queue honest: your next case comes from your rubric’s lowest score, not your mood.

When should you stop practising alone and get a live person?

The moment your rubric scores stop moving. Solo reps build the skill; live sessions prove it under pressure. A live session catches what solo practice structurally cannot: an interviewer who reads your hesitation and leans on it, interruptions you did not script, fit questions between cases. Book a peer from your consulting club, an alum, or a paid coach for a few sessions late in prep, and treat them as exam conditions, not practice.

Can a chatbot do the asking instead?

For the asking, yes; for the interview, only partly. A text chatbot will run a case with you for free at 1 am, useful for early reps and structure drills. Judge it against what the interview demands, though: does it hold data back until you ask the right question, push back more than once, and score how you speak rather than what you typed? A text tool reads typing, not speaking, so do not mistake a smooth chat for a passed interview.

Where do reading, peer practice and AI voice practice fit?

Match the route to the stage.

Stage	Practice	Purpose
Early	Reading + drills	Vocabulary and pattern recognition
Middle	Solo, timed, recorded	The method in this post. Volume lives here.
Late	Live pressure	Human mocks + AI voice reps

Reading builds vocabulary, not performance, so it comes first. Volume belongs in the middle: solo recorded reps are free and available at 11 pm. Late in prep, add pressure - human mocks for realism, AI voice practice for the reps a human cannot give at scale.

The empty room does not stay empty. It starts with one well-run rep tonight.

Where Northstar fits

A partner will not be there at 11 pm the night before your interview, or for a sixth rep on a Sunday, and a chatbot will not reproduce what the interview does to you. That gap is where AI voice agents come in, and it is what we made. Northstar is an AI case interview simulator: a live AI interviewer runs a real case with you by voice, gives data only when you ask, pushes back, and ends with a report scoring each skill. Use a human for pressure and fit late in prep, and Northstar for the reps in between. More at the case library.

Try it

Open Wiped Out, a 30-minute profitability case, and run it with the script and method above. Speak every step, record it, and score yourself with the rubric before you peek at any answer.

For a comparison of voice tools, read AI case interview prep: what the tools can and can’t do.

Questions? support@northstarclub.in