How to Crack the McKinsey AI Interview with Lilli
McKinsey now uses Lilli in live case interviews. Here's how to prep, perform, and stand out in an AI-assisted consulting round.
McKinsey has changed what it's testing in case interviews, and most candidates are still preparing for the old format.
Lilli, McKinsey's internal AI platform, is now part of live interview rounds at some locations. Roughly 1 in 3 MBB first-rounds include some form of AI-assisted case component, and that number is climbing. If you're doing case prep with only a casebook and a prep partner, you're building the right skills in the wrong context.
This post breaks down what the Lilli format actually looks like, what McKinsey is measuring, and a four-step approach you can practice before you walk into the room.
What Lilli Is (and What It Isn't)
Think of Lilli like a research analyst who has read every McKinsey engagement ever run, but has no judgment about what matters for your specific problem. You can ask it anything. It will give you something back. Whether that something is useful depends entirely on how precisely you direct it.
That framing matters because it resets the skill you're building. You're not learning to use a chatbot. You're learning to act as the senior thinker who knows what questions to ask, can evaluate the answers critically, and can synthesize the output into a recommendation the client can act on. Lilli handles the retrieval. You supply the judgment.
McKinsey uses Lilli in real engagements every day. The interview is a compressed simulation of what Day 1 actually looks like.
The One Skill the Interview Is Measuring
Before you think about tactics, get the frame right.
The Lilli interview is not testing how well you use AI. It's testing the same thing every McKinsey interview tests: whether you can impose structure on an ambiguous problem, form a hypothesis, and build a defensible recommendation. The AI is a new delivery mechanism for that same assessment.
The candidates who struggle are the ones who open the tool first. They type a vague prompt, get a plausible-sounding response, and start building an answer around it. The candidates who succeed treat Lilli like a junior analyst who needs a very specific question to return something useful. They build the structure first. They bring in the tool to fill targeted gaps, not to generate a starting point.
The interview is not testing whether you can use AI. It's testing whether you're the kind of thinker who could direct AI well in a real engagement.
Practice this on a case that puts you in a realistic AI-assisted environment. BoardroomIQ's Case Simulator Tool mirrors this format: live case prompt, AI co-pilot, structured feedback on how well you're directing vs. deferring.
Step 1: Structure Before You Prompt
The single most common mistake in AI-assisted cases is opening with the tool.
Spend the first 60 to 90 seconds doing what you'd do in any case: clarify the question, identify the decision to be made, and sketch your issue tree. What are the two or three critical questions that, if answered, would crack this problem?
Only then do you direct Lilli, toward specific, bounded tasks. The difference between a weak prompt and a strong one is the difference between "tell me about this market" and "what are the typical fixed cost drivers for a B2B SaaS company operating at $200M ARR?" One produces a wall of text you have to sort through. The other returns exactly what you needed to fill one node in your tree.
Strong AI prompts in a case context are surgical. They fill a specific gap. They are not requests for Lilli to solve the case.
Apply what you just learned
Browse all 100 real boardroom decision cases.
Step 2: Audit Every Output Before You Use It
Lilli is powerful, but McKinsey interviewers know it isn't infallible. The interviewer may deliberately let the tool surface a flawed answer to see if you catch it.
After every output, run a fast three-point check. Is it relevant: does this actually answer the sub-question I asked, or did it drift? Is it complete: are there obvious gaps relative to the case context? Is it defensible: could I walk the interviewer through this logic without it falling apart?
If the answer to any of those is no, say so out loud. Narrating your critical evaluation is not a weakness. It's the signal McKinsey is looking for. Candidates who silently incorporate bad AI output are the ones who get passed on.
Step 3: Use AI to Pressure-Test, Not Just Generate
One underused move: once you've built your own hypothesis, ask Lilli to challenge it.
Something like "what are the strongest counterarguments to a market entry recommendation for this company?" or "what risks am I not accounting for in this cost-reduction analysis?" This mirrors how senior consultants use AI in real engagements, not as a first-draft machine, but as a red-team partner.
It also signals intellectual confidence. You're not afraid to stress-test your own thinking. That's a quality McKinsey actively selects for, and it's obvious when a candidate is doing it genuinely vs. performing it.
Step 4: Own the Synthesis and the Close
The final two minutes of an AI-assisted case are identical to any other case close. You need a clear, direct answer to the question posed, two to three supporting reasons in priority order, and one material risk or caveat the client should monitor.
Lilli cannot do this for you. The synthesis is where your judgment shows up. Candidates who read from Lilli's summary output in the closing are immediately recognizable as people who never actually internalized the analysis.
Write your own closing bullets. Speak them in your own words. The AI was a tool. The recommendation is yours.
How to Practice Before Your Interview
The gap between knowing this framework and executing it under live interview pressure is real. You need reps.
Timed structure drills. Set a 90-second timer and build an issue tree for a prompt before touching any AI tool. Repeat with 10 different case types until the structure comes before the instinct to search.
AI prompt precision practice. Take a practice case and write out the exact prompts you'd give Lilli at each stage. Grade yourself: is each prompt surgical, or is it asking the AI to do your thinking?
Red-team your own hypotheses. After building a recommendation on any practice case, write down the three strongest arguments against it. Then ask: if I were using Lilli, how would I phrase the prompt to surface those counterarguments?
The best way to practice the Lilli format is under realistic pressure, with a case that fights back. Try the McKinsey Office Hours case on BoardroomIQ to build the muscle before the real thing.