McKinsey's AI Interviewer Arrives: Do Job Candidates Have to Talk to a Robot?

Recruitment processes are being rewritten by AI. McKinsey has introduced an AI interviewer that simulates a client into its junior business analyst interviews in North America — candidates must spend about ten minutes conversing with a virtual client, responding to questions and pushback, while the actual evaluation is carried out by a human assessor in the room. This isn't a concept demo; it's a real, live pilot.

A Ten-Minute "Client Conversation"

According to reports from The Wall Street Journal, The Verge, and the New York Post, McKinsey piloted an AI interview segment this fall in its recruiting for some junior business analysts in North America. Instead of traditional behavioral interview questions, candidates face an AI playing a client — for example, an "AI mayor" seeking advice on stadium construction. The interview lasts about ten minutes, and potentially hundreds of candidates may participate.

The key point: the AI itself does not score candidates; the final decision rests with the human assessor in the room. McKinsey has also not published any quantitative data on the system's effectiveness.

This point deserves to be examined on its own. The AI is the vehicle for the conversation, and humans are the decision-makers — this "human-machine division of labor" design leaves a logical fallback. But what candidates can perceive is only the tireless conversational partner on the other side.

Simulated Client Interview is an interview format that combines scenario simulation with generative AI. Its prototype is the case interview that the consulting industry has used for years — candidates must analyze a business problem and offer recommendations within a limited time. Traditionally, a human interviewer plays the role of the client or partner, while the AI version replaces this role with a large language model-driven virtual character, allowing the conversation to be generated in real time, follow up at any moment, and not be constrained by the interviewer's schedule.

It's important to distinguish "AI-assisted evaluation" from "AI playing a role": the former refers to using algorithms to analyze candidates' answers and output scores, while the latter (the model in McKinsey's pilot) only lets the AI serve as the interactive interface, with judgment still belonging to humans. This distinction determines where fairness concerns land — the controversy is not about algorithmic scoring, but about the script design itself.

When the Interviewer Is No Longer Human

From the perspective of candidate experience, this is deeply divisive.

Some people might like this format. There's no nervousness of facing a partner, no small talk, no need to read the other person's micro-expressions — you just need to demonstrate your thought process. For introverted candidates or those who aren't good at social performance, this might even feel fair.

Others will be at a loss. Talking to a machine is inherently counterintuitive, with no past experience to reference. The more central question is: real client conversations have texture — clients sigh, suddenly change topics, push back forcefully because their boss is in the room. How much of that texture can a pre-set script simulate? Currently, there is no public data to answer this question.

First the Robot, Then the Boss

An easily overlooked detail is the sequence of the process: candidates are first interviewed by AI, and only after joining do they report to a human manager.

What kind of employer impression does this "machine first, human later" sequence create? Some people will think it doesn't matter — a tool is just a tool. Others will see it as a signal — does this company treat me as a person, or as a data point? First impressions of an employer are often hard to change. If the first "real conversation" a company gives you is with a machine, how will your expectations for this employment relationship be defined? There's no standard answer to this question, but it genuinely exists.

The Black Box Behind the Script

When you talk to someone who has worked at the company, you can gauge the weight of their words, probe for details, and read hesitation in their expressions — the credibility of this information can be calibrated.

But when an AI plays the client, who sets its lines? Candidates can't see the prompts, don't know what behaviors the system is designed to reward, at what points it will apply pressure, or how its personality parameters are tuned. This is completely different from a seasoned partner who has actually dealt with clients, in terms of information transparency.

To be fair, the human assessor on site is a genuine safety valve in this segment. The final decision is made by a person, and the AI's role is to provide the conversational scenario rather than to judge. But candidates' performance is ultimately shaped by a script they cannot see. The transparency this requires is, at present, far from sufficient.

Prompt is the instructional text that a user or developer inputs to a large language model, and the model's output content, tone, and boundaries are essentially determined by it. In simulated interview scenarios, prompts typically include character personas (e.g., "You are a budget-sensitive mayor"), conversational goals, pressure point settings, and information that must not be disclosed. Candidates cannot see this content, and therefore cannot judge which pushbacks are deliberately designed stress tests and which are simply random outputs of the model.

Personality parameters or system settings refer to mechanisms that control consistency in model behavior through temperature values, role constraints, and similar methods. With different parameters, the same AI might appear as a gentle listener or as an aggressive interrogator, and candidates have no right to know about this. This is precisely the technical root of the "black box" controversy: it's not that AI is scoring, but that the AI's mode of performance itself is filtering candidates.

If You're Going to Use It, Meet at Least Three Conditions

Author Ashkan Rajaee (founder of TDZ Pro) laid out three bottom-line requirements in the original article. As a practitioner who has led teams in software services and consulting, growing a business from six million dollars to sixty million dollars in scale, his starting point is the cost and fairness of the recruitment process.

1. Disclose in advance: Clearly tell candidates how the simulated interview works and who ultimately makes the decision.
2. Publish data: Can this simulation actually predict on-the-job performance? Currently there are zero public metrics, and results are needed to back it up.
3. Offer a human alternative: Provide another path for candidates unwilling to talk to AI.

None of these three are anti-technology; they simply raise the burden of proof — because the person sitting on the other side of the screen is betting their career. The author also acknowledges in the original article that he speculates one motivation for large companies to introduce AI interviews is to reduce recruiting costs, but this is only his inference; McKinsey has not stated this officially, and its stated goal is to observe how candidates think and respond.

The Boundary Between Fact and Opinion Across Multiple Sources

This article synthesizes reports from three media outlets: The Wall Street Journal, The Verge, and the New York Post. The information that can be cross-verified at the factual level includes: the pilot took place this fall, in North America, targeting some junior business analyst candidates, lasting about ten minutes, with AI not scoring, human assessors making the final decision, and currently no public effectiveness data.

As for "does this mean big companies are using AI to cut recruiting costs," "will candidates' first impressions be damaged as a result," and "is the script black box fair" — these are the original author's judgments and extrapolations, belonging to opinion rather than fact. In discussing such topics, maintaining this dividing line is meaningful.

Would you be willing to let a robot interview you?

Predictive Validity is a core metric in psychometrics and personnel assessment, measuring the extent to which a selection tool can predict a candidate's actual job performance after joining. Even if an interview segment has a smooth experience and sophisticated design, if it has no statistical correlation with subsequent performance, it is ineffective as a screening tool — and may even harm recruitment quality by introducing irrelevant variables.

Assessing predictive validity typically requires long-term tracking data: recording candidates' interview performance, then conducting a correlation analysis with their performance evaluations one to two years after joining. McKinsey's pilot has just launched, and both the sample size and time span are insufficient to support such research. Therefore, "does this AI interview actually work" cannot be answered at this stage — and this is precisely the substantive meaning of the author's demand to "publish data": not to demand an immediate conclusion, but to require that the verification mechanism be incorporated into the process design.

One-line Summary

McKinsey pilots AI client interviews, retaining human decision-making power, but the script black box and missing effectiveness data spark fairness controversy.

Paragraph Summary

This fall, McKinsey piloted AI interviews in some junior business analyst recruiting in North America: candidates must engage in about ten minutes of real-time conversation with an AI playing a client, but the AI itself does not score — the final judgment is made by a human assessor on site, and no effectiveness data has been officially released. This "human-machine division of labor" design leaves room logically, yet allows candidates' performance to be shaped by an invisible set of prompts and personality parameters. The article notes that candidate experiences are polarized and outlines three bottom-line requirements: advance disclosure of the mechanism, publication of predictive validity data, and a human interview alternative. The author also emphasizes that the pilot facts should be strictly distinguished from inferences such as "big companies using AI to cut costs."