Why Most 'AI Recruiting' Tools Get Candidate Screening Wrong

A lot of products marketed as "AI recruiting" are, under the hood, keyword matching with a chatbot bolted on the front. That's not a criticism of the idea of AI in hiring — it's a criticism of how thin most implementations are. Having built a multi-tenant hiring platform that does CV parsing, AI-scored candidate evaluation, and real-time AI video interviews, the difference between screening that's actually useful and screening that's decorative comes down to a few specific design choices.

Keyword matching isn't scoring

The most common failure mode: a tool extracts keywords from a job description ("5 years React experience," "AWS," "team lead"), extracts the same keywords from a résumé, and reports an overlap percentage as a "match score." This rewards résumé keyword-stuffing over actual fit, penalizes strong candidates who describe their experience in different language than the job posting, and gives hiring teams a false sense of rigor — a number that looks scientific but measures almost nothing about whether the person can do the job.

Actual AI-assisted scoring needs to reason about equivalence, not just string matching: someone who led a team of engineers has "leadership experience" even if their résumé never uses that phrase; someone who built a REST API in Express has relevant backend experience for a Django role even though the frameworks don't match on paper. This is exactly the kind of judgment call that a language model, given the full context of both the résumé and the role, is actually good at — but only if you build the scoring prompt to reason about substance rather than reducing everything to a keyword-overlap calculation first.

Structured extraction, not just a summary

A résumé isn't just unstructured text to summarize — it's a document with real structure (roles, dates, companies, skills) that a hiring workflow needs to query, filter, and sort on later. If your AI layer just produces a paragraph summary of a candidate, you've thrown away the ability to do anything programmatic with it — filter for candidates with 3+ years in a specific stack, sort by total years of relevant experience, flag employment gaps for a recruiter to ask about.

The pattern that's worked well: use the model to extract structured fields (roles, companies, dates, skills, education) from the freeform résumé text, validate that structure against a schema, and store it as actual queryable data — with the AI-generated summary as one field among many, not the only output. This is also what makes CV-based question personalization possible: if you've extracted that a candidate spent two years doing incident response, an interview can generate a question specifically about that, rather than asking generic questions that could apply to anyone.

Interviews are where the real signal is, and where most tools stop short

Résumé screening, done well, narrows a large pool down to a reasonable shortlist. It cannot tell you how someone communicates, thinks under mild pressure, or handles a follow-up question — and that's exactly the information that traditionally required a human recruiter's time on a screening call. This is the part most "AI recruiting" tools skip entirely, because it's genuinely harder to build: it requires a live, adaptive conversation rather than a one-shot analysis of static text.

A real-time AI interview — configurable by role, difficulty, and question focus, informed by the structured data pulled from that specific candidate's résumé — closes that gap. The interview isn't a fixed script; it's dynamically shaped by what's actually in front of it, the same way a good human interviewer adjusts their questions based on what a candidate says.

What "good" actually requires

Building AI-assisted screening that's worth trusting requires three things most shortcuts skip: structured data extraction instead of keyword matching, scoring logic that reasons about equivalence and context rather than string overlap, and — if you're going deep enough to add interviews — a genuinely adaptive conversation rather than a static form with a chat interface wrapped around it. None of that is exotic technology. It's mostly a matter of not stopping at the easiest thing that demos well.