Some entry-level interviews at McKinsey gained an unusual counterpart this fall: as the Wall Street Journal reported on October 14, 2026, the firm is piloting a roughly ten-minute conversation with an AI avatar for undergraduate business-analyst candidates at some North American offices, and the avatar plays not the interviewer but the client. Partners or assessors sit in the same office watching how candidates respond and adapt, while the AI itself gives no score and no assessment; the judgment stays human. Hundreds of candidates are expected to go through the exercise this fall, and if it works, the firm expects to extend it to other roles next year.
What those ten minutes actually test
The AI client was trained on different client personalities, for example a professorial type who wants to understand the reasoning behind a recommendation. Assessors are not looking for a standard answer; they watch for in-the-moment response: whether the candidate heard what actually mattered in what the client said, acted on it, or simply pushed through a prepared agenda regardless. In other words, it tests the most everyday consulting skill, keeping a conversation on track and moving it forward while information is incomplete and the other side keeps adding conditions. That differs from automated interviews where an AI runs a case and scores the candidate: here the AI only makes the scenario feel real, and the evaluating is deliberately left to the people in the room.
It deliberately avoids playing a nightmare client
A McKinsey partner involved in designing the pilot says the tool is not meant to mimic an unhinged client or to ask gotcha questions; the aim is to reflect about 80 percent of the scenarios consultants meet in day-to-day work. That restraint matters: the pilot measures whether a candidate's fundamentals are sound, not their stress ceiling, and extreme scenarios would distort the assessment. It should also not be confused with McKinsey's separate pilot in which candidates use its internal AI assistant Lilli on a case; that exercise looks at solving problems with AI, while this one looks at communicating with a client, and conflating the two misjudges the difficulty of either.
What candidates should take from it
For candidates, AI has not rewritten how to prepare: the trainable skills are still listening carefully, confirming understanding, and revising a judgment when new information arrives, and on the day itself the avatar is best treated as a real client rather than a prompt to be gamed. Rehearsing client-style conversations beforehand with a method for turning AI into a mock interviewer is the cheapest way to get used to being probed repeatedly. The broader signal is plain enough: when a consulting firm uses an AI client to screen people, such simulations are moving from training into hiring; yet the fact that scoring stays with humans shows that, at least for choosing people, even a firm selling AI transformation has not handed the decision to AI.
