McKinsey has introduced an AI chatbot into the early stages of graduate recruitment. Applicants interact with the system as part of assessment, allowing the firm to gather structured information before interviews and later human judgement.
The use case appears narrow, but its consequences are meaningful. Graduate programmes receive large application volumes within compressed hiring cycles, and an automated conversation can give every candidate the same initial questions.
Consistency and scale are useful only if the process remains fair. Once AI influences who receives scarce recruiter attention, “assistance” can shape an outcome even when software does not make the final decision.
Structured Conversations Can Reduce Screening Work
A chatbot can engage every applicant, organise responses, and surface information about communication or problem-solving. Recruiters receive a more consistent starting point than thousands of differently formatted applications and can spend more time on candidates who progress.
McKinsey says the tool gathers information rather than making hiring decisions independently. That boundary should be operationally explicit. Recruiters need to know which signals are generated, how heavily they influence progression, and when a person must examine the original response rather than a score or summary.
The system may also improve access if it allows candidates to participate at flexible times. Conversely, it can disadvantage people unfamiliar with chatbot interaction or those requiring accommodations. Efficiency depends on providing an equivalent and usable route for different applicants.
Recruiters should examine whether the conversation collects information that actually predicts performance. A structured experience may feel more rigorous while merely converting subjective impressions into data fields. Validation must connect each assessed signal to the role, not to a generic image of the ideal consultant.
Bias Can Enter Through Questions as Well as Training Data
Automated hiring risk is often discussed as a biased model, but the design of the assessment matters equally. A question may reward a communication style associated with particular educational or cultural backgrounds, while language analysis may interpret disability or second-language use as weaker performance.
Organisations should compare progression rates across relevant groups, review disputed cases, test alternative phrasing, and monitor changes over time. A model that appeared balanced during a pilot can drift as applicant behaviour or recruitment priorities change.
Human review does not automatically remove bias. Recruiters may defer to an automated result because it appears objective. Training should explain the system’s limitations and require documented judgement when AI affects a sensitive decision.
Periodic independent review would strengthen that control. Auditors can compare the chatbot’s influence with subsequent interviews and hiring outcomes, investigate unexpected group differences, and test whether changes to the model or assessment altered the candidate pool.
Candidates Need Meaningful Transparency and Recourse
Applicants should know they are interacting with AI, how the exchange fits into assessment, which data is retained, and whether conversations train future systems. Transparency allows them to prepare and decide whether they are comfortable participating.
There must also be a route to request accommodation, report a technical failure, or challenge a result. A chatbot can deliver consistent questions while creating an impersonal barrier if no recruiter is reachable when context matters.
Evaluation should go beyond recruiter hours saved. McKinsey needs to examine candidate experience, quality of later hires, demographic effects, false rejections, and whether interviews genuinely become deeper because screening work changed.
Recruitment is a revealing enterprise AI test because it combines repetitive volume with lasting human consequences. The chatbot may help McKinsey manage scale, but responsibility cannot be distributed into the software stack.
The defensible model is one where AI organises early evidence, recruiters understand and challenge it, and candidates retain visibility and recourse. Hiring remains a human decision not because people are automatically unbiased, but because an accountable organisation must be able to explain why it chose one person over another.
