AI 候選人篩選代理：篩選 50 份履歷而不失一致性

Reading fifty CVs in a sitting produces diminishing quality after the first twenty because attention drifts, criteria drift, and the unconscious shortcuts that introduce bias accelerate. The candidate who applies on day one receives more careful consideration than the equally strong candidate who applies on day ten, when the reviewer has read forty CVs and is pattern-matching rather than evaluating. A thorough manual screening of fifty applications takes three to four hours. Speed produces poor decisions, thoroughness produces bottlenecks, and most hiring processes make peace with both.

Must-haves are eliminators — the skills, experience, or qualifications the person cannot do the job without. Nice-to-haves are differentiators — the attributes that make one qualified candidate stronger than another. A must-have list that is too long produces no qualified candidates. A must-have list that conflates essential requirements with preferences produces a screened-in pool that does not reflect the role. The distinction matters: must have experience managing P&L for a finance director role versus ideally has experience managing P&L for a senior analyst role.

Structured candidate screening uses three verdict types: Advance (meets all must-have criteria and should proceed to interviews), Hold (meets some but not all must-have criteria, or meets all must-haves but lacks differentiators that stronger candidates have), and Decline (does not meet essential criteria). Each verdict includes two or three specific reasons written in terms of the defined criteria. Hold candidates provide a second tier to consider if the Advance pool is smaller than expected, rather than requiring a full restart of the sourcing process.

Employment gaps should be treated as questions to explore in a screening call, not as automatic grounds for rejection. Employment gaps that reflect parenting, caregiving, health situations, or career transitions are legal grounds for concern in several jurisdictions if used as rejection criteria, and they are poor proxies for candidate quality. The same applies to frequent job changes, unexplained title changes, or roles below the candidate's apparent level. Flag for context rather than declining automatically, because patterns that look concerning on a CV often have innocent explanations.

When screening criteria are not defined before reviewing applications, each reviewer builds criteria implicitly as they go, and implicit criteria are unstable. The first strong candidate sets an informal benchmark. A candidate who looks like a previous successful hire gets the benefit of the doubt. When multiple reviewers screen the same pool without shared criteria, different reviewers weight different attributes, and the shortlist reflects aggregated biases as much as role requirements. Defining criteria first means the shortlist can be explained and defended to the hiring manager, to challenged candidates, and internally if the process is audited.