How to Use AI for Case Interview Practice

Use AI for case interview practice as a simulator and critic, not as a source of live answers. Give it an official practice case, make it reveal information only when you ask a relevant question, solve the case aloud, and ask for evidence-based feedback after you finish. This trains the skills firms say they evaluate: problem structure, assumptions, quantitative reasoning, communication, and a supported recommendation.

Do not carry the practice setup into a real interview unless the employer explicitly permits AI assistance. As of July 22, 2026, McKinsey's hiring guidance encourages responsible AI use for preparation but prohibits generating real-time interview answers and says candidates should disable AI note-taking or virtual-assistant software before a virtual interview. The boundary is part of the exercise: prepare with AI, then demonstrate your own reasoning when the rules require it.

What should AI case interview practice train?

A case interview presents an ambiguous business problem and evaluates how you work toward a recommendation. The final number or answer matters, but it is not the only output.

Current guidance from the firms is unusually consistent. BCG's case interview preparation page describes a step-by-step process that includes structuring the approach, asking thoughtful questions, analyzing data, doing quick calculations, identifying important factors, and communicating clearly. Bain's interview guidance says it looks for sensible assumptions, quick math, and the ability to build constructively on others' ideas. Deloitte's case and scenario interview tips add practical judgment, an initial hypothesis, a logical story, recommendations, and next steps.

An effective practice loop therefore measures six observable skills:

Do not ask the model for a generic score such as “8 out of 10.” Ask it to cite moments from the transcript: the unanswered clarification, the assumption that changed, the calculation with mismatched units, or the recommendation that ignored a risk. Feedback without transcript evidence is too easy to accept and too hard to improve from.

Choose a source case before choosing a prompt

Start with a case whose facts and intended learning path come from a primary source. McKinsey publishes sample problem-solving cases on its interviewing page, while BCG, Bain, and Deloitte publish cases or interactive practice tools through the pages linked above. These materials are better anchors than asking a model to invent a plausible consulting problem, because they give you a stable fact set and a credible reference solution or teaching path.

Create two packets:

• Candidate packet: the opening prompt and exhibits the candidate is allowed to see.
• Interviewer packet: all case facts, reveal conditions, calculations, and the reference conclusion.

Put the interviewer packet into the AI context, but instruct the model not to reveal it wholesale. The assistant should answer only the question asked, provide an exhibit only at its intended checkpoint, and say when the case packet does not contain an answer. A practice partner should be able to run the same case from the same packets and produce the same facts.

Use synthetic or public practice material. Do not upload a confidential recruiting case, employer document, client information, or another candidate's interview recollection. Before adding resumes, transcripts, or files to any assistant, review the AI interview assistant privacy checklist.

Configure the AI as a restrained interviewer

The model needs role boundaries more than it needs a clever persona. A compact interviewer instruction can be:

Run the supplied practice case. Start with only the candidate prompt. Reveal a fact or exhibit only when the candidate asks a relevant question or reaches the named checkpoint. Do not suggest a framework, solve arithmetic, complete a sentence, or rescue a weak answer during the case. Ask one concise follow-up at a time. Record the candidate's stated assumptions, calculations, intermediate conclusions, and final recommendation. After the candidate says the case is complete, evaluate only against the supplied rubric and cite transcript evidence for every finding.

Add two failure rules:

1. If the interviewer packet lacks a requested fact, say the information is unavailable instead of inventing it.
2. If the candidate asks for the answer during the case, ask what they would investigate next instead of revealing the solution.

Those rules preserve the useful friction of the exercise. If the assistant silently completes the issue tree, catches every arithmetic error immediately, or supplies the “so what,” the candidate is rehearsing tool use rather than case solving.

If you are evaluating how a desktop assistant captures prompts and returns context during a mock session, use the broader desktop AI interview assistant evaluation criteria. Product fit is separate from case quality.

Run the case in four checkpoints

1. Clarify the decision

Restate the client's decision in one sentence. Ask for the objective, time horizon, geography, customer or business unit, and the metric that defines success when those details are missing. Then state the constraints you will treat as fixed.

The assistant should not reward a long list of questions. It should record whether each question changes the scope or analysis. A useful checkpoint output looks like this:

• Objective: decide whether to enter the market.
• Success measure: positive operating profit within three years.
• Scope: one country and one customer segment.
• Known constraint: existing manufacturing capacity cannot expand in year one.
• Remaining ambiguity: required return on investment is not provided.

This gives the structure a target. Without it, a polished framework can analyze the wrong decision.

2. State a structure and an initial hypothesis

Take a short pause, then explain the branches you want to investigate and why each one matters. A market-entry case might require market attractiveness, ability to win, economics, and execution risk; a profitability case may need a different structure. Do not force a memorized template onto the prompt.

State an initial hypothesis as provisional: “I would enter only if the addressable segment is large enough and our capacity can reach a viable unit cost.” Then name the first analysis that could disprove it.

Ask the assistant to flag three issues after the case, not during it:

• a branch that does not connect to the objective;
• overlap or a material missing branch; and
• a hypothesis that never changes despite contrary evidence.

BCG's current preparation guidance explicitly emphasizes structural thinking, classified assumptions, and visible reasoning. The aim is not to recite framework vocabulary. It is to make the next question follow logically from the decision.

3. Analyze exhibits and calculate aloud

For every exhibit, use the same sequence:

1. Read the title, axes, units, time period, and source notes.
2. State the pattern that matters to the client question.
3. Calculate only what the decision requires.
4. Check the units and order of magnitude.
5. Explain how the result changes the hypothesis or next branch.

Keep a calculation ledger in the transcript. For example:

Question: Can year-three contribution cover fixed operating cost?
Formula: customers × annual purchases × contribution per purchase
Inputs: 80,000 × 3 × $18
Result: $4.32 million annual contribution
Check: 240,000 purchases at roughly $20 each should be just under $5 million
Implication: contribution covers the stated $3.6 million fixed cost, before launch cost

After the case, let the assistant recompute the arithmetic and check units against the packet. It should distinguish a math error from a reasoning error. A correct multiplication cannot repair an unsupported input, and a reasonable assumption should not be marked wrong merely because the reference case chose another value.

4. Synthesize before recommending

Pause after each major finding and state an intermediate conclusion: what you learned, why it matters, and what you need next. This prevents the final recommendation from becoming the first time the analysis has a point of view.

Finish with four parts:

1. Recommendation: answer the original decision directly.
2. Reasons: give the two or three findings that carry the conclusion.
3. Risks: name the unresolved condition most likely to change it.
4. Next steps: request the evidence or action that would reduce that uncertainty.

The AI should reject a reason that does not appear in the case record. It should also flag a recommendation that is stronger than the evidence—for example, “enter immediately” when the economics work only if an untested customer-adoption assumption holds.

Review the transcript in three passes

Do not ask for all feedback at once. Separate the review so a fluent delivery does not hide a weak analysis.

Pass one: case logic

Ask the assistant to reconstruct your path from objective to structure, analysis, intermediate conclusions, and recommendation. Every step should connect. Mark branches that were opened but never resolved and facts that appeared in the recommendation without analysis.

Pass two: numbers and assumptions

List every stated number with its source, unit, formula, and use. Separate case facts from candidate assumptions. Recalculate the arithmetic, then identify which assumptions are decision-sensitive. Practice saying what new data would validate those assumptions.

Pass three: communication

Find the first clear statement of the objective, hypothesis, each intermediate conclusion, and the final recommendation. Look for long silent stretches, repeated questions, unexplained jargon, and answers that bury the conclusion. The goal is not maximum speed. BCG warns against rushing and overcomplicating, while Deloitte recommends taking time to compose thoughts and treating the case as a business conversation.

End the review with no more than three changes for the next attempt. Examples include “state units before calculating,” “close each branch with a conclusion,” or “name the key risk in the recommendation.” Then rerun a different case that stresses those behaviors.

Practice failure without letting AI rescue you

Run at least one session with a deliberate constraint:

• the assistant withholds a nonessential fact;
• an exhibit contains irrelevant data;
• your first hypothesis is wrong;
• you make a calculation error and must catch it with a sanity check; or
• the assistant becomes unavailable halfway through.

Continue the case from your own notes. A practice workflow fails if losing the tool also removes your structure, calculations, or ability to speak. Before any permitted workflow involving audio, screenshots, or a meeting application, run the full AI interview assistant preflight with synthetic material.

Keep preparation and the real interview separate

Employer instructions control the real interview. McKinsey's current AI hiring guidance encourages AI for practice and concept explanations, but not for generating real-time answers or completing assessments unless specifically permitted. BCG France provides another explicit example: its candidate honor code excludes AI tools, note-taking plug-ins, and problem-solving software during interviews.

Do not infer permission from technical capability, a remote format, or the absence of proctoring. Ask the recruiter when instructions are unclear. If assistance is prohibited, close the software and use only allowed materials. The broader AI interview ethics guide offers a separate decision framework for disclosure, fairness, and employer rules.

Use AI to expose your reasoning during practice

Good AI case interview practice does not make the case easier. It makes your reasoning inspectable. Use an official case packet, constrain the interviewer, solve aloud through four checkpoints, and demand transcript evidence for every critique. Then rehearse the improvement without depending on the tool.

If you want to run a synthetic remote case with Control, download the desktop app and use the free allowance—five messages and two minutes of voice—to test one short capture-to-critique loop. Keep Control out of the real interview unless the employer explicitly permits it.