AI Candidate Screening: Separate the Score From the Decision

AI candidate screening uses software to evaluate application materials or assessment responses against hiring criteria. It can produce a summary, match score, ranking or recommendation—and, in some configurations, advance or reject an applicant.

But not every automated hiring step is AI. A resume parser extracts information; a knockout rule rejects a specified answer; a scoring model evaluates evidence. The useful question is not just “Does this employer use AI?” It is “What information does the tool use, and what happens because of its output?”

Separate four parts of the screening workflow

These functions can sit inside the same applicant tracking system (ATS), but they do different jobs.

Function	What it does	What to check
Resume parsing	Extracts details into candidate-profile fields	Did job titles, dates and qualifications transfer correctly?
Knockout rules	Disqualify applicants based on configured application answers	Is each requirement genuinely necessary, and is the question unambiguous?
AI evaluation	Scores or categorizes evidence against criteria	What criteria, evidence and missing-information rules determine the result?
Workflow automation	Uses rules or scores to advance, hold or reject candidates	Which actions are enabled, and who can reverse them?

For example, Greenhouse describes parsing as scanning a resume and filling candidate-profile fields. An unsuccessful parse leaves the resume attached but requires manual entry. That is not, by itself, a rejection.

Workable’s screening documentation separately describes knockout questions that automatically disqualify applicants who answer “No,” and an AI agent that evaluates candidates against structured matching criteria. Those are different mechanisms, even if both happen before a recruiter reviews an application.

What an AI screening score actually means

A score needs a definition. It might represent how closely a profile matches employer-selected requirements, rather than a measured probability of successful job performance.

As an illustrative product example—not an endorsement—Workable documents a score out of 100 against an Ideal Candidate Profile, with reasoning for criteria marked “met,” “partially met,” “not met” or “unknown.” It also documents separate controls for automatic advancement and disqualification; automatic disqualification is off by default.

That distinction matters: an 82/100 match score is not automatically an 82% chance of succeeding in the job. Predictive meaning requires separate evidence.

Consider a hypothetical applicant with extensive customer-support experience but no named ticketing platform on their resume. A screener could mark platform proficiency as unknown. That is an evidence gap, not proof that the applicant lacks the skill. If an employer treats both identically, the cutoff can discard candidates worth asking one more question.

Written reasoning helps reviewers inspect a score. It does not independently establish that the scoring method is accurate or job-relevant. Nor does disabling automatic rejection remove every risk: if recruiters review only the highest-ranked applicants, low-ranked candidates may never receive meaningful review.

What candidates can change

Focus on making relevant evidence readable and explicit, not on chasing a universal “ATS score.”

• Use a straightforward document. Follow the employer’s file instructions, use clear sections and keep essential information out of text boxes. Greenhouse identifies columns, graphics, tables, headers and footers as possible sources of parsing problems; this is documented behavior for that system, not proof every ATS fails on them.
• Check extracted fields when the application lets you. Correct missing dates, truncated titles or misplaced qualifications before submitting.
• Connect skills to evidence. If you genuinely used the requested software, name it alongside the task you performed. A keyword list alone tells a reviewer less than a concrete example.
• Read application questions carefully. Resume changes cannot undo an answer that triggers a separate knockout rule. If a requirement is ambiguous, ask rather than misrepresent your eligibility.
• Raise accessibility barriers early. Explain which part of the assessment creates the barrier and ask for an appropriate adjustment. Our AI video interview accommodation guide covers practical requests.

These steps improve clarity. They do not guarantee advancement or reveal an employer’s configuration. For a fuller application workflow, see our resume preparation guide.

What HR should verify before allowing rejection

Start with the decision, then work backward to the evidence:

1. Define job-related criteria. Separate essential requirements from preferences. Document why each criterion belongs in screening.
2. Inspect the underlying evidence. Review the original application alongside extracted fields and AI reasoning. Establish how “unknown” differs from “does not meet.”
3. Pilot without automatic rejection. Compare outputs with a structured human assessment. Include a sample of low-scoring candidates—not just the recommended shortlist—to look for qualified applicants the tool misses. Human ratings are a comparison, not unquestionable ground truth.
4. Give reviewers real authority. Make the original evidence available, allow corrections and overrides, and record the reasons. A nominal human sign-off is not a useful check if reviewers cannot investigate.
5. Monitor the deployed workflow. Track advancement, errors, accommodation issues and overrides by role and stage. Assess whether screening disproportionately excludes protected groups, using lawfully collected data and appropriate expertise. Define when to pause automation. Use our post-deployment monitoring guide for a fuller plan.

For U.S. employers, the EEOC’s selection-procedure guidance emphasizes validation for the positions and purposes involved. Vendor documentation can help, but it does not transfer the employer’s responsibility to the vendor.

Legal checks are separate from product settings

Rules reference checked October 14, 2026. These are selected U.S. requirements, not a comprehensive legal checklist. This article is informational, not legal, HR or employment advice.

Under the Americans with Disabilities Act, covered employers must avoid unlawful disability discrimination in screening and provide reasonable accommodations to qualified applicants with disabilities unless doing so would create undue hardship. DOJ guidance specifically warns that hiring technology can screen out qualified applicants with disabilities, including when the technology comes from another company.

In New York City, qualifying automated employment decision tools—not every ATS feature—are subject to Local Law 144. DCWP’s overview lists requirements including a bias audit within the preceding year, publicly available audit information and required notices. Enforcement began July 5, 2023.

DCWP’s FAQ explains the geographic scope: it includes jobs at an NYC office, fully remote jobs associated with an NYC office, and specified employment-agency uses. For covered uses, NYC-resident candidates must receive notice at least 10 business days before use; the FAQ permits notice on the employer’s employment webpage.

An audit is not a certification that a tool is discrimination-free. DCWP states that Local Law 144 does not prescribe specific action based on audit results, while other anti-discrimination laws still apply. Confirm the applicable jurisdiction and obligations with qualified counsel before deployment.