What 363 AI-scored case interviews show

Across the 363 case interviews analysed, synthesis sits 7.5 points below the mean of the eight competencies and structure 6.7 points below, while chart reading sits 6.5 points above. That leaves 14 points between the highest and the lowest competency.

Candidates can usually read a figure, react to new information and hold a conversation with the interviewer. What costs them points happens at the two ends of the case: opening with a plan that really breaks the problem down, then closing with a clear recommendation rather than a recap of what was just said. Maths comes right behind.

If you are short on time, this is where it pays off most. The opening and the close can be practised on their own, in a few minutes, without redoing a whole case each time.

To work on those two moments: Build an issue tree, The MECE principle, Close with the Minto pyramid.

Among candidates with at least six case interviews, communication gains 12.7 points between their first three interviews and their last three, and maths 7.3 points. The overall score rises by 3.9 points, and 12 of these 17 candidates finish higher than they started.

The progress does not show straight away: between the first and the third interview, the mean score stays about the same. The first few sessions mostly go into discovering the format, speaking out loud and keeping the pace. The gap opens after that, and it comes first from what becomes automatic: leading the exchange, saying what you are doing, calculating without losing the thread.

Chart reading and business sense barely move, because they start higher. Structure and synthesis do improve, but more slowly than the rest: repeating interviews is not enough to fix them, they need dedicated work.

To speed up what improves fastest: Practise mental maths, Questions asked in interviews.

An interview of 15 to 25 minutes is scored like one of more than 35 minutes: the gap between the three longest brackets is no more than 0.2 point. Only interviews under a quarter of an hour stand apart, 7.1 points below the others.

A case stopped that early has usually not had time to reach the recommendation, which leaves part of the competencies with nothing to score. The average case interview lasts about 25 minutes, and nothing suggests that going longer earns more.

So you do not need to block out an hour for a session to count. Twenty minutes carried through to the recommendation beat a long case dropped halfway.

To pick your next case: The interview case library.

The 60 interviews taken on an advanced case score 2.4 points above those taken on a beginner case, not below. The intermediate level sits between the two.

The type of case weighs a little more without changing the picture: 5.4 points separate market entry, the highest, from market sizing, the lowest. The score depends more on how you run the case than on the case itself.

One caveat: everyone picks their own case, and those who take an advanced one often have more practice behind them. This does not say an advanced case is easy. It says you should not wait until you feel ready to try one, nor stay on a single type of case.

To vary the types of case: Market entry, Profitability, Market sizing, Practise market sizing.

The study covers full case interviews taken out loud on FrameLab from 15 June 2026 to 12 November 2026, finished and scored. Interviews under five minutes, abandoned interviews, team accounts, fit interviews and the short format focused on the opening of a case are left out. That leaves 363 interviews taken by 65 candidates.

Each interview is run by voice by an AI interviewer, then scored from its transcript, from 0 to 100 on eight competencies: structure, maths, chart reading, business sense, creativity, synthesis, communication and poise. Since November 2026 the score comes from a grid of criteria specific to each format (28 criteria for a full case), each scored on four levels, from absent to excellent; the score of a competency is the weighted mean of its criteria. The 314 earlier interviews were scored by the previous version, which picked out the passages of the transcript tied to each competency and gave them a level on the same scale. With either version, synthesis is the lowest-scored competency.

Progress is computed only on interviews scored by the previous version, so that two scales are not compared. Every result is a mean: none rests on fewer than 30 interviews or fewer than 15 candidates, and no individual data is published.

These are the scores of an AI applying a grid, not the outcome of real interviews at a firm. Each candidate picks their case, their interviewer and their length, so the gaps describe what is observed, not a cause. The figures are updated as new interviews come in.