McKinsey to Test Some Applicants With an AI Client Avatar
McKinsey will add an AI avatar that acts as a client to interviews for some entry-level consulting candidates this fall, according to The Wall Street Journal. The scope is limited. Only part of the entry-level pipeline will use the system at first.
The avatar plays a mayor, according to New York Post. Candidates speak with that synthetic character as the client in the simulation. The format is conversation, not a written exercise.
McKinsey's guidance for virtual candidates tells them to disable AI software they have installed, according to McKinsey. That rule covers tools on the candidate side of the session. The firm's avatar runs on the assessment side.
The broader context here is familiar to anyone who builds tests for conversational software, often called evaluation harnesses. A fixed written case checks structured thinking under set conditions. A live dialogue checks different skills: drawing out missing information, recovering after a misunderstanding, and managing turn taking. Those behaviors are harder to script and harder to grade consistently. They sit closer to client work, where facts arrive incomplete and views shift during discussion. The design tradeoffs are familiar. Every candidate needs the same experience. Open talk still needs clear grading rules, called rubrics. Response speed, a steady persona, and safety limits, called guardrails, must hold constant so the test measures candidate behavior rather than familiarity with synthetic talkers.
In my view, the asymmetry deserves attention. Candidates are asked to switch off their own assistants and then are judged through talk with the employer's system. That is coherent for measurement. Outside help adds noise, while a shared simulation gives everyone the same task. Yet it changes the filter. Comfort with synthetic counterparts becomes part of what is selected for, including patience with stiff pacing, clear follow-up questions, and composure when the other side mishears or loops. Some strong communicators in human settings will adapt at once. Others will need adjustment. How well performance here predicts job effectiveness will depend on how much work dialogue itself comes to involve similar systems.
Looking at what this means for hiring, the interview starts to work as a sample of a different workflow. Preparation shifts from rehearsing frameworks alone toward rehearsing interaction: listening for cues, structuring unclear facts in real time, and steering a talk toward a decision. For hiring teams, a repeatable test allows cleaner comparison across a large pool than different human interviewers can deliver at scale. The risk is overfitting, with candidates learning avatar quirks rather than client skills. Handled carefully, the opportunity is selection that rewards listening and clear structure over polish, and early exposure to work where professional talks increasingly include software on the other end.
