AI Job Interviews: McKinsey Tests a Simulated Client
McKinsey is testing a simulated AI client in some entry-level interviews. See what the pilot signals for hiring, candidate fairness, and human oversight.
McKinsey's AI Client Interview Is a Test of Judgment, Not Just Technology
A job interview with an AI client can sound like a test of whether candidates can talk to a machine. But that's probably the wrong lens for McKinsey's reported experiment. The more relevant question is whether a simulated client can reliably show how someone listens, reasons, and responds under pressure - and whether hiring teams can judge that fairly rather than just consistently.
According to The Wall Street Journal, McKinsey plans to run a 10-minute AI-client conversation with some entry-level business-analyst candidates in certain North American offices this fall. The simulation is meant to reflect the kinds of situations consultants run into day to day, and McKinsey says it won't be used to imitate an extreme or "unhinged" client. The firm also says people - not the AI - will assess candidates.
That distinction matters. An AI interview might make an exercise easier to repeat, but repeatability isn't the same as fairness. And a simulation that feels realistic still might not be a valid way to predict job performance. Here's what the reported pilot suggests, what it doesn't clarify yet, and what employers should think about before putting an AI character in front of applicants.
What McKinsey is testing
The reported test is narrow. It's for some undergraduates applying for business-analyst roles in some North American offices. The conversation is expected to last 10 minutes, and The Wall Street Journal reports that "hundreds" of candidates may participate. This is information about a trial - not a company-wide rewrite of McKinsey's recruiting process.
McKinsey says the goal is to resemble a client interaction. It also claims to cover about 80% of scenarios consultants face in everyday work. But that number doesn't tell us whether the simulation has been validated against actual job performance. The reporting doesn't explain how those scenarios were chosen, how the system will work in practice, or how McKinsey plans to judge the results of the pilot.
One detail stands out: the AI isn't meant to score applicants. People will be there to observe and evaluate how candidates handle the conversation. In other words, the model helps generate the interview scenario, while humans stay responsible for what it means.
That makes this different from a fully automated screening system that ranks candidates or rejects them without human review. Still, "no automated score" doesn't remove all the concerns people have about AI in hiring. Even without a numeric grade, an AI can steer the interaction - what candidates get asked, how the simulated client responds, and what behaviors an interviewer finds easiest to notice.
What the report does not tell us
Some crucial details aren't public in the reporting. For example, it doesn't say whether candidates can opt for a non-AI alternative, whether the simulation uses voice or text, what information is retained, or whether every applicant faces the same scenarios.
Those aren't just trivia. They determine whether candidates get comparable conditions, how personal data is handled, and whether applicants who communicate in different ways can reasonably participate. Until those pieces are clear, it's safer to treat this as a reported pilot than to guess how effective - or fair - the tool will be.
That caution applies to any new hiring technology. A company can announce what it plans to test; that doesn't automatically prove it improves hiring outcomes.
Why simulate a client conversation
A client-style exercise is easy to defend on practical grounds. Consulting work involves understanding a problem, asking questions, and responding to another person's needs. A short role-play can also give interviewers something more concrete to observe than a candidate's description of what they would do in a hypothetical scenario.
It may also be easier to deliver the same exercise to lots of applicants. In a conventional interview, candidates can end up with different questions, different interviewers, and different levels of prompting. A structured scenario could reduce some of that variation - especially if the prompts and evaluation criteria are controlled.
Still, the presence of AI doesn't guarantee standardization. If the model responds differently based on what a candidate says - or on tone, pacing, or phrasing - some applicants could get a tougher interaction than others. And even a fixed scenario can still turn uneven if the AI behaves unpredictably or if different interviewers interpret the simulation in different ways.
The comparison shouldn't be "AI versus a perfect human interviewer." No process is perfectly consistent. The real question is whether this new exercise produces better evidence about job-relevant skills than the current approach - without creating unfair barriers and at a cost the organization can actually justify.
A simulation measures a slice of the job
A 10-minute conversation can reveal how someone approaches a discussion. But it can't stand in for the full range of consulting work, and it shouldn't be treated as a complete measurement of ability. The outcome depends on the scenario, how comfortable the candidate is with interview role-play, how clearly the task is explained, and how interviewers read what happens.
That's a general limitation of hiring exercises, not something unique to AI. Case interviews, timed tests, and group activities also sample only parts of what a person can do. The practical issue is whether the sample is relevant - and whether decision-makers understand what it does, and doesn't, demonstrate.
For employers, the goal should be useful observation, not a cool demo of "advanced technology." If the simulation feels realistic but tests qualities unrelated to the role, it's technically interesting and operationally weak.
Human scoring is a safeguard, not a guarantee
Keeping people responsible for evaluation is a sensible design choice. It prevents the final call from being handed entirely to a model, and it leaves room for context - like follow-up questions, clarifying misunderstandings, and incorporating the candidate's response into the bigger picture.
But "a human is in the room" isn't a full governance plan. People can still be influenced by an AI-generated interaction, especially when the AI sounds confident or behaves like a convincing stand-in for a client. An interviewer may, without meaning to, treat the model's behavior as a fair proxy for the candidate's skill. Or they might focus on how smoothly someone navigated the interface instead of whether their reasoning held up.
Human review works best when evaluators have a clear rubric, understand what the simulation can and can't measure, and are trained to separate what they can observe from what they assume. And even then, if one interviewer prioritizes directness while another prioritizes rapport, you can still get inconsistent judgments - even if everyone sees the same exercise.
Fairness depends on the whole process
A hiring tool should be assessed across the full candidate experience, not just by checking whether its output includes a score. Useful questions include:
- Does the exercise measure skills that matter for the role?
- Are candidates given clear instructions and comparable opportunities to respond?
- Can applicants request an accommodation or switch to another format?
- Does the simulation hold up across different communication styles?
- Are interviewers using shared criteria and recording evidence, or relying on impressions?
- If the process collects or retains candidate information, what's it used for?
These questions matter because an AI interaction can affect people differently without anyone trying to discriminate. Some candidates may not be familiar with conversational role-play. Others may have speech differences. Some may use an accent the system handles poorly, or may need accommodations. The pilot details about these issues haven't been reported, so it's premature to claim the tool either causes problems or solves them.
The responsible move is to test for those risks before relying on the exercise. That means gathering feedback, looking for patterns in outcomes, and giving candidates a way to flag concerns. It also means being ready to change - or stop - the test if it creates friction without improving hiring decisions.
A practical framework for evaluating AI interviews
If you're considering a similar system, the decision shouldn't start with whether the AI sounds believable. Start with the job and work backward. If you can't explain the purpose of the exercise in plain language, the technology shouldn't be placed between a candidate and a job opportunity.
A practical evaluation can be organized into five steps.
-
Define the skill being assessed
Spell out the capability the exercise is meant to reveal. For a client conversation, that might mean asking clarifying questions or explaining a recommendation. Avoid vague goals like "culture fit" or "communication style." Those terms are hard to measure consistently and can reward similarity to existing employees more than performance in the job.
-
Test whether the scenario reflects the work
Ask people who understand the role to review the scenario and the behaviors you expect to see. A simulation should mirror realistic tasks without forcing candidates to guess hidden rules. If the scenario tests how well someone knows consulting interview tropes rather than how they understand and respond to a client, it may end up selecting for prior exposure - not potential.
-
Keep the assessment criteria visible to evaluators
Interviewers should use a rubric tied to observable evidence. For example: did the candidate clarify what the client wanted? Did they explain the reasoning behind the proposed next step? The rubric shouldn't pretend the interaction is more precise than it is, but it can still make judgments easier to compare and review.
-
Check the candidate experience and outcomes
A pilot should collect information about reliability, accessibility, candidate feedback, and whether the exercise adds evidence beyond what you already get from the rest of the interview process. Employers should also look for whether particular groups are more likely to struggle or drop out - while protecting personal data and avoiding conclusions based on small, misleading samples.
-
Decide what would make you stop
Set clear limits before the pilot starts. If the simulation is inconsistent, creates barriers that can't be addressed, or doesn't improve the quality of assessment, pause it. A test is only valuable if the organization is willing to learn - even from an unfavorable result.
This is where many AI deployments stumble. Teams often measure whether the system can be launched, not whether it improves the decision it was introduced to support. In recruitment, that can leave organizations with a polished new step and no convincing evidence that candidates are being assessed better.
The same discipline applies to vendor claims. Ask what's measured, how the model behaves when candidates respond in unexpected ways, what data gets processed, and what happens when a candidate requests an accommodation. If those answers are unclear, the tool may not be ready for a high-stakes use case.
Conclusion
McKinsey's reported AI-client interview is notable because it places generative AI inside a familiar, high-stakes business process: evaluating job applicants. The real question isn't whether an AI can role-play a client for 10 minutes. It's whether the interaction creates relevant evidence that can be applied consistently and accessibly - while still keeping people accountable for the decision.
The pilot reportedly keeps humans in the loop and appears limited to some entry-level candidates in some North American offices. That's a more restrained approach than handing hiring decisions to an automated score. But what matters next is what hasn't been reported yet: the assessment rubric, accommodations, data practices, and how results from the pilot will be reviewed.
For employers, the standard is straightforward: define the skill, validate the exercise, monitor effects, and be ready to stop. For candidates, the key point is that the tool - based on what's been reported - is a simulated interview scenario, not an AI making the hiring decision. Read The Verge's report on McKinsey's AI client interview for the original coverage, and judge the technology based on the quality of the process around it - not on how human the simulation sounds.