McKinsey AI interview 2026: how the Lilli round works, and what it actually tests

The headline going round is that the case interview is dead. It is not. What has changed is narrower and more interesting: McKinsey has added a stage where you solve a problem with an AI assistant in the room, and how you handle the assistant is part of what gets observed.

McKinsey started piloting this in January 2026. Candidates are given access to Lilli, the firm's internal generative AI platform, and asked to use it while analyzing a case and refining a recommendation.4 2 The firm has been unusually open about it: in late July 2026, McKinsey's Director of Talent Assessment, Simone Jacobi, gave an on-the-record interview explaining why the firm introduced AI into the process and what candidates should expect.1 That matters, because most of what circulates about consulting assessments comes from candidate debriefs rather than the firm.

Where the AI round sits in the process

The most common misreading is that this replaces something. It does not. It is a layer on top of a process that otherwise looks the way it did last year.

Two practical consequences. First, if you are an undergraduate applying for a Business Analyst role, the Solve assessment is still the gate that eliminates the most people, and it comes first. Nothing about the AI round changes where you should spend your early preparation. Second, because the AI stage sits in the final round, you will have weeks of notice before it becomes relevant to you.

What it is, and what it is not

What is actually being observed

Strip away the novelty and the exercise is a familiar one wearing new clothes. A consultant with a research analyst available has always been judged on how well they direct that person, how skeptically they read what comes back, and whether they can turn it into something a client can act on. The AI has replaced the analyst. The judgment being assessed is the same judgment.

Reported evaluation focuses on four things: how you scope and structure your request, how critically you assess what comes back, whether you push back when the output is thin, and whether you can synthesize it into a clear recommendation.6 The process is iterative rather than linear, so you cycle between asking, evaluating and refining instead of moving in a straight line.3

The difference a good request makes

Concretely, here is the gap between the two approaches candidates take. Both are real patterns, and the first is the more common one.

The second version does the work the interviewer wants to see: it defines the problem space, commits to a structure, and asks the tool to expose its own uncertainty. It is also the version that produces output you can meaningfully challenge, which is the next thing being observed.

How to prepare without access to Lilli

You cannot practice on Lilli. It is an internal tool. But nothing in the reported evaluation criteria depends on that specific product, which means a public assistant is a perfectly adequate rehearsal environment.6

1. Case first, AI second. If your structuring or your maths is shaky, the AI round will expose it faster, not hide it. Fundamentals are still the prerequisite.
2. Practice on real cases, out loud. Take a case, work it with an assistant open, and narrate your reasoning as you go. The narration is the skill, because in the room you have to explain what you are doing and why.
3. Deliberately challenge the output. On every case, find at least two things in the response that are unsupported, and say so explicitly. Build the reflex now so it is automatic under pressure.
4. Practice the synthesis. End every rep with a 60-second recommendation delivered without looking at the screen. This is the part candidates skip and the part interviewers remember.
5. Time it. The iterative loop eats clock. Candidates who have never done this under time constraint consistently run long.

Solve still comes first

Worth restating, because the AI story is loud and the sequencing is easy to get wrong. The AI exercise is a final-round component. The Solve assessment sits between your application and your first interview, and it removes far more candidates than any interview round does. If you are early in the process, Solve is where your hours should go.

Where this is heading

Bain has signaled that it is building an AI-enabled interview component of its own, and has indicated it is interested not only in what candidates produce with a tool but in how they used AI while preparing.5 That second part is worth sitting with. It implies a view where using AI in your preparation is expected rather than suspect, which is close to the opposite of the assumption most candidates are operating under.

Two honest caveats. The Bain signal is reported rather than firm-confirmed, and details are thin. And formats piloted in US final rounds do not always survive contact with a full cycle, so it is possible this looks different by next spring. Prepare for the version that exists, stay loose about the version that arrives.

The most useful thing to take from this is not a technique. It is that the firms are now openly evaluating a skill they previously assumed you would pick up on the job. Building it deliberately is a real edge, and it costs you nothing beyond changing how you run practice you were already going to do.