How HR Professionals Can Use AI to Identify Top Talent with Greater Accuracy

The most uncomfortable finding in HR right now is that AI hasn’t sped up hiring. According to SHRM’s 2025 Talent Trends, AI adoption inside HR jumped from 26% to 43% year over year, but cost-per-hire and time-to-hire have both increased over the past three years. Around 70% of organisations still report recruiting challenges, even with AI in the funnel.

The reason isn’t that AI is bad at screening. It’s that the way most teams use it floods the funnel and then asks AI to clean up the flood. If you want AI to actually improve your shortlist quality (not just shrink your reading time), the work isn’t in writing better screening prompts. It’s in building a structured rubric the AI scores against, every time, in a way you can document and defend.

What the data actually says about AI in talent ID

SHRM’s research shows recruiting is the most common HR use case for AI (64% of AI-using HR teams), well ahead of training and engagement. But “use” doesn’t mean “good use.” The increase in cost-per-hire suggests two things are happening simultaneously: candidates are using AI to mass-apply (flooding the funnel), and recruiters are using AI to screen at speed without tightening their decision criteria.

The fix is upstream of the AI. If your job description, scoring rubric, and “must-have” criteria are vague, AI screening produces fast but unreliable shortlists. If they’re specific, AI screening produces a real lift in quality.

Step one: build the rubric before you touch the AI

Before any candidate review, write a four-criterion rubric. Not ten criteria, four. Each one with a 1 to 5 scale and a clear description of what each score looks like in evidence. The rubric is the artefact that anchors every subsequent AI prompt.

“You are a senior talent partner helping me build a structured screening rubric for the [JOB TITLE] role at [COMPANY], reporting to [HIRING MANAGER ROLE]. Job description and key responsibilities: [PASTE JD]. Identify the four most predictive criteria for success in this specific role, distinguishing them from generic ‘good employee’ traits. For each criterion, write: (1) a one-sentence definition, (2) a 1-5 scale with concrete behavioural anchors at scores 1, 3, and 5, and (3) the type of resume or application evidence that would be required to score above a 3. Output as a clean table. After the table, list two criteria that are commonly used for this kind of role but are NOT predictive of success and should be deliberately excluded.”

What you get back is the spine of your screening process. It’s also the document you need if anyone (a candidate, a regulator, your own legal team) ever asks how decisions were made on this role.

Step two: score against the rubric, don’t ask for “best fit”

Here’s where most AI screening goes wrong. Recruiters paste in 50 resumes and ask “which of these are the best fit?” The model produces a confident ranking with no transparency about how it ranked, and the recruiter trusts it because it sounds professional.

The better prompt scores each candidate against the rubric explicitly:

“You are screening resumes against the attached rubric for the [JOB TITLE] role. For each candidate below, produce a row containing: candidate identifier, score (1 to 5) on each of the four criteria, the specific resume evidence that justifies each score (quote the resume text), an overall recommendation (advance / borderline / decline), and any ‘unable to assess from resume’ criteria that would need to be tested in the screening call. Do NOT use any criteria not in the rubric. Do NOT factor in name, gender, age cues, university prestige, or location unless explicitly required by the role. If a candidate’s resume gives you insufficient evidence to score a criterion, score it as ‘unknown’, do not infer.”

This output is auditable. Every score has a quoted piece of evidence behind it. If a hiring manager pushes back on a decline, you can show the working. More importantly, candidates the resume doesn’t fully describe (career changers, returners, non-traditional paths) get an “unknown” rather than a low score, which is the difference between a screening process that finds hidden talent and one that quietly filters them out.

The failure mode: AI as a black box

Here’s the trap. A platform vendor sells you “AI-powered screening” with no visibility into the criteria, no exportable scoring trail, and a “match score” that the system can’t explain. A candidate is rejected, an EEOC or local equivalent complaint follows, and you cannot show how the decision was made. That’s not a hypothetical, it’s the failure mode that’s pushing the hiring industry toward more transparent rubrics rather than less.

The rule of thumb that holds up: if you can’t reproduce the decision and explain it in plain language, don’t make the decision. Tools that fit this rule for HR teams in 2026 include Greenhouse, Lever, Workable, and BambooHR, which support structured scorecards and exportable decision trails. For sourcing, hireEZ provides AI candidate discovery with citation back to the public profile data it used. For prompt-based screening through a general LLM, ChatGPT Team and Claude both let you save the conversation, which is your audit trail.

A worked example: the “five years experience” trap

A common screening criterion is “5+ years of relevant experience.” Asked to score against this, AI will faithfully apply it, and quietly screen out two kinds of strong candidates: career changers with three years of highly relevant work plus seven years of adjacent work, and returners whose recent gap looks like four years of inactivity but is actually two years of caregiving and two years of part-time consulting.

A better criterion (in your rubric, not in the JD) is “demonstrated experience delivering [SPECIFIC OUTCOME].” Now the model is looking for evidence of the outcome, which a career changer’s three years might produce more compellingly than a default candidate’s five.

Run a quick test. Pick a recent role where your final hire was someone unconventional. Re-screen the applicant pool against the rubric you wrote (above) versus the original JD’s criteria. The number of strong candidates that surface from the rubric pass who didn’t surface from the original pass is usually startling. That gap is the cost of vague criteria, and AI scoring magnifies it rather than fixing it.

What this changes about the recruiter’s role

The HR professionals seeing real gains aren’t the ones running the most AI screens. They’re the ones spending more time on rubric design, more time on the screening call, and less time on resume review. The reading is what AI takes off your plate. The judgment is what stays.

Document everything. Save the rubric per role. Save the prompt that produced the scoring. Save the AI output. Save your override notes when you advanced a candidate the AI scored below threshold (which you should be doing on a meaningful percentage of decisions, that’s the human-in-the-loop check that keeps the process honest). The hire that comes from this workflow is genuinely better, not just faster, and the documentation makes it defensible if anyone ever asks how you got there.

The headline finding from SHRM is that AI hasn’t fixed hiring. It’s worth sitting with. AI doesn’t fix bad criteria, it accelerates them. The teams that beat the trend are the ones doing the unglamorous work of writing good rubrics first, and using AI to apply them consistently second.