Case interview prep used to mean finding a partner with two free hours and a casebook. As of June 2026, a $20/month AI subscription runs a credible case end-to-end, throws curveballs mid-case, and grades you against a rubric that quotes your own words back to you. It still cannot judge your executive presence over a video link, but everything before that it can drill harder than a tired peer at 11pm.
There is also a new reason to practice this exact skill. McKinsey began piloting an AI-assisted interview round in late 2025, where candidates use the firm’s internal tool, Lilli, to work a case live while interviewers watch how you prompt, challenge, and re-anchor the AI’s output to the client. The pilot is non-evaluative for now (calibration only) and is expanding across U.S. Business Analyst hiring in spring and summer 2026. You will not get Lilli to practice on, but the muscle it tests, structured thinking plus disciplined AI use, is exactly what the workflow below builds.
TL;DR
- Use AI for the feedback bottleneck, not the volume one: 2-3 full cases a day with a quoted-evidence rubric beats 8 ungraded reps.
- Claude Pro ($20/mo) holds long sensitivity chains and full-case context best and does cleaner multi-step math; ChatGPT Plus ($20/mo) wins on volume and Voice Mode for talking through a case out loud. Both run effectively unlimited drills.
- Drip the case, do not dump it: tell the AI to give only the opening prompt, then wait for your clarifications and structure before releasing data one piece at a time.
- See a few real cases first (Case in Point, Victor Cheng), or AI teaches you confident-but-wrong patterns.
- Always run a hard 25-minute clock and a forced 90-second synthesis. The synthesis is where interviewers actually grade you.
The task
Run 5-10 full cases this week and finish each one with a concrete list of what to fix before tomorrow. Volume without feedback is the trap; AI fixes the feedback bottleneck, not the volume one.
When this is the right job for AI
- You already understand profitability, market entry, M&A, and market sizing frames at a basic level.
- You can structure for 90 seconds without freezing — AI is for drilling, not for first contact.
- You will read the feedback and rerun the same case once before moving on.
- You have a target firm tier (MBB, Tier 2, boutique) so the case style can match.
If you have never seen a case, do 2-3 from Case in Point or Victor Cheng first. Otherwise AI gives you a fake-credible case and you learn confident-but-wrong patterns.
Which AI to use (June 2026)
Any of the big three runs a usable case. The differences that matter for prep:
| Tool | Price/mo | Default model | Best for | Watch out for |
|---|---|---|---|---|
| Claude Pro | $20 ($17 annual) | Sonnet 4.6, Opus 4.7 on hard cases | Long sensitivity chains, full-case memory, reliable multi-step math | No native voice; can be over-agreeable unless you force partner pushback |
| ChatGPT Plus | $20 | GPT-5.5 (Instant / Thinking) | Volume drilling, Voice Mode to talk a case out loud | Sometimes leaks the answer early; rein it in with a strict “no verdicts mid-case” line |
| Google AI Pro | $19.99 | Gemini 3.1 Pro | Long casebook PDFs, 1M-token context for whole prep folders | Fewer ready-made case prompts in the wild |
A practical split most candidates land on: Claude for the 2-3 hard graded cases a week, ChatGPT for daily volume and out-loud reps. Free tiers work for a session or two but hit limits fast, and ChatGPT’s US free tier now carries ads. For the underlying model trade-offs, see our ChatGPT vs Claude vs Gemini comparison.
What to feed the AI
- Your target firm tier and the round you are prepping for (R1 fit-heavy, R2 partner case)
- Case type to drill (profitability / market sizing / M&A / capacity / pricing)
- The 1-2 weaknesses last mock surfaced (“rushes structure”, “ignores qualitative”)
- A hard 25-minute clock you actually respect
- Your structure preference (issue tree, hypothesis-driven, MECE branches)
Copy-ready prompt
Run this on a reasoning model (ChatGPT’s Thinking mode or Claude Opus 4.7) so the case logic and math hold up. The one rule that makes or breaks it: the AI must drip information as you ask, not dump the whole case at once. The prompt below enforces that.
You are a senior consulting interviewer running a case. I am prepping for {MBB R2 partner round}.
Case type: {profitability decline}
My known weaknesses: {rushes structure, weak in qualitative drivers, math anxiety on percentage changes}
Preferred structure: hypothesis-driven, 3 MECE branches max.
Run the case as follows:
1. Open with a 2-3 sentence prompt. Stop. Wait for me to clarify and structure.
2. After my structure, push back exactly once on the weakest branch with a partner-style "why did you put X under Y".
3. Mid-case, throw ONE realistic curveball (e.g. a new data point, a CEO preference, a competitor move) and force me to update.
4. Include one quantitative exhibit I have to read — give me a small table or 4-5 number bullets — and one calculation that uses at least 2 steps.
5. End by asking for a synthesis under 90 seconds.
After my synthesis, give feedback in this format:
- Structure (1-10) + the single specific reason
- Math accuracy + speed (each 1-10, name the slowest moment)
- Hypothesis discipline (did I update on the curveball or paper over it?)
- Communication (signposts, top-down, pause-and-think)
- 3 concrete drills for tomorrow
Do NOT give me the answer until the synthesis. If I ask for the answer, refuse and ask me what I think first.Sample output structure
Prompt: “Our client is a US regional grocery chain. Same-store sales are down 8% year-over-year while the broader market is flat. The CEO wants to know why and what to do.”
Pause. Wait for clarifying questions. Wait for structure.
Pushback: “You put pricing under external — competitive pricing pressure feels external, but our own pricing decisions are internal. Walk me through that.”
Curveball, dropped at minute 15: “By the way, our private-label SKUs are flat year-over-year. Only national brands are down. Does that change your hypothesis?”
Exhibit: a 5-row table of category-level sales declines, with health-and-beauty down 22% while produce is flat.
After synthesis, feedback breaks structure score, math score, hypothesis discipline, communication, and three drills (e.g. “tomorrow: 20 percentage-change calcs in 8 minutes; one full case re-running this one cold; 5 min of synthesis-only drills”).
How to refine
- AI gives the answer too early: add “do not validate or invalidate any of my hypotheses mid-case. Push back as a partner would — with a question, not a verdict.”
- Curveballs feel artificial: ask for curveballs sourced from real consulting situations (a board deadline, a CEO bias, a regulatory letter), not “imagine our competitor lowers prices.”
- Feedback is generic: require a quoted line from your own response next to each criticism. “You said
X. That is hand-wavy because Y.” - Math feels too easy: ask explicitly for 3-step calculations drawn from the five categories that show up in nearly every MBB case — percentages, break-even, compound growth (CAGR / Rule of 72), market-sizing chains, and profitability bridges — with at least one percentage change and one weighted average per case.
- You keep getting the same case type: rotate by giving AI a 5-case plan upfront and asking it to track which you have done.
Common mistakes
- Treating AI like a casebook lookup. The value is in the live interaction, not the prompt text.
- No clock. A case without a clock is a tutoring session, not a mock.
- Skipping the synthesis. Synthesis is where interviewers actually grade you; AI grades it harder than peer partners do.
- Running 8 cases, fixing nothing. One case rerun beats three new ones.
- Letting AI play a friendly interviewer. Ask for partner-style pushback; you need to practice composure under pressure.
FAQ
- Can AI replace a human case partner? For drills and feedback, mostly yes. For executive presence and pacing under camera, no. Book 2-3 human mocks before round 1; a peer or a paid coach catches body language and silence-handling the AI cannot see. For format and math fundamentals to pressure-test the AI’s cases against, IGotAnOffer’s case math guide is a solid free reference.
- How many cases per day is too many? Three full cases (25 min + 15 min feedback each) is the daily ceiling. Beyond that, fatigue makes you learn worse patterns.
- What about behavioral / fit? Use AI separately for fit. Cases and fit need different rubrics; combining them in one prompt waters both down.
- How do I drill math without doing full cases? Ask AI to generate 20 mental-math problems at MBB difficulty across the five recurring categories — percentage changes, break-even, CAGR, market-sizing chains, profitability bridges — with an 8-minute clock and a “no calculator, show your rounding” rule. Target roughly 10 seconds per percentage step and under 30 seconds for break-even with the business read.
- What does a good synthesis sound like? Recommendation first, then 2-3 supporting reasons pulled straight from your analysis, then one risk or next step. Budget 2-3 minutes. Tell the AI to grade specifically whether you led with the answer (top-down) instead of recapping your framework.
- Will this prep me for McKinsey’s Lilli AI interview? Partly. The Lilli round grades how you prompt the tool, whether you challenge its output, and whether you re-anchor it to the client. The copy-ready prompt above trains the same instincts — just add reps where you deliberately ask the AI to be wrong, then catch and correct it.
- Should I tell AI my target firm? Yes. MBB cases are more rigorous on hypothesis discipline; Tier 2 cases lean on market sizing and creative qualitative.
Related
- AI behavioral interview prep
- AI mock interview
- AI tell-me-about-yourself practice
- AI salary negotiation script
- AI company / role research
- AI LinkedIn Profile Rewrite That Reads Like You Wrote It
- AI Portfolio Project Narrative: STAR With a Spine
Tags: #AI writing #Case interview #consulting #job-search-practice