How AI Résumé Screening Works in a Recruiter's Day

1. Parsing and matching: the applicant tracking system's first pass

Most screening starts inside an applicant tracking system (ATS), the database that receives applications for an employer's job postings. The software first parses each résumé, turning free text into fields such as job titles, dates, education and skills, and then compares those fields with the requirements in the posting. Court filings in Mobley v. Workday, Inc. describe one such feature in Workday's own words: Candidate Skills Match "extract[s] skills in the employer's job posting" and the applicant's materials to "determine the extent to which the applicant's skills match the role," and reports the result to the employer as "strong," "good," "fair," "low," "pending" or "unable to score."¹²

Pfizer's published description of HiredScore, a Workday product it uses, follows the same pattern: each application gets a grade from A (strongest match to the requisition) to D, scored independently of the other applicants in the pool. Pfizer states that the grade "does not supplant human judgment" and that recruiters and hiring managers "always have complete discretion" over whom they select.⁴ In the scorecard above that is why matching counts as "Assists": the software sorts, a person decides.

2. Scoring, tiers and the language-model layer

A second family of tools scores the candidate rather than the résumé. Pfizer's 2023 notice listed HireVue alongside HiredScore; the HireVue assessment in the audit Pfizer posted sorts applicants into top, middle and bottom tiers, and the report states that Pfizer evaluates candidates using the top/middle-versus-bottom threshold.⁶⁷ Newer features write text instead of a number. CareerPlug, an ATS for small businesses and franchises, offers a "position-specific qualification summary" that, when a client clicks it, opens with one of three verdicts ("Meets key requirements," "Meets some key requirements," "Missing key requirements") followed by a narrative summary of the applicant against the job.¹³

Whether large language models (LLMs) sort fairly is still mostly tested in labs. A University of Washington study presented at the 2024 AAAI/ACM Conference on AI, Ethics, and Society ran over 500 public résumés against 500 job descriptions through three text-embedding models and found White-associated names significantly favored in 85.1% of tests.¹⁴ That is a prototype-grade result: simulated screening, not an employer's live system. An August 2026 erratum says the paper's gender-only results were inverted by a coding bug; its race and intersectional results were unaffected and were replicated by a second team.¹⁴

3. Conversational assistants: questions, scheduling and offers

The tasks AI most clearly does on its own are the conversational ones. When Chipotle rolled out Paradox's system across more than 3,500 restaurants in October 2024, it said its assistant, named "Ava Cado," would chat with candidates, answer their questions about Chipotle, collect basic information, schedule interviews for hiring managers and send offers "to candidates who are selected by managers."³ The split is explicit: the software handles the messages, a manager makes the choice.

The same pattern runs on the employee side of HR. IBM says its AskHR agent answers routine questions such as "What is the sickness policy?" while human advisors take complex cases; language models classify each prompt, route it to an HR domain and either trigger a task or write an answer. IBM reports a 94% containment rate for common questions and more than 11.5 million employee interactions in 2024 (company-stated; IBM also sells the platform).²

4. AI-led interviews

Interviewing is where the evidence is newest. In a pre-registered natural field experiment with PSG Global Solutions, a recruitment outsourcing subsidiary of Teleperformance hiring for customer-service jobs in the Philippines, about 70,000 applicants were randomly assigned to a human interviewer or an AI voice agent, which disclosed that it was artificial at the start of each call. Human recruiters then evaluated every interview and made the hiring decisions.⁸

Applicants interviewed by the AI agent received offers in 9.73% of cases against 8.70% for human interviewers, which the researchers report as a 12% higher likelihood; they were also 18% more likely to start the job and to still be employed a month later. Transcript analysis found the AI interviews more structured and consistent, which the authors associate with more hiring-relevant information being collected. The paper discloses that the partner firm supplied data but had no role in the analysis, and that one author took an unpaid role at the firm after data collection ended in June 2025.⁸ Because it was a trial in real hiring rather than a standing deployment, it is graded "Pilot."

5. Sourcing agents: searching before anyone applies

Before applications arrive, AI agents now search for candidates. LinkedIn says its Hiring Assistant, launched to charter customers in October 2024, became available globally in English by the end of September 2025, and that early adopters were "saving 4+ hours per role, reviewing 62% fewer profiles" (vendor-stated).⁹ The customer names and quotes in that release appear in LinkedIn's own material, so this page grades the case as a vendor claim until an employer confirms it in its own words.

Sourcing also falls outside New York City's audit law. DCWP's guidance says the requirements do not apply when an employer uses a tool to scan a résumé bank, conduct outreach or invite applications; they apply only to assessing people who have applied for a specific position.¹⁵

6. What a New York City bias audit measures

Local Law 144 of 2021 bars employers and employment agencies from using an automated employment decision tool (AEDT) for jobs in the city unless the tool has had a bias audit within the past year, a summary of the results is public, and candidates got notice; DCWP began enforcing it on July 5, 2023.¹⁶ An AEDT is a tool that uses machine learning, statistical modeling, data analytics or AI and "substantially assists or replaces discretionary decision-making." The audit must be done by an independent auditor, and notice must reach candidates 10 business days before use.¹⁵

The core number is the impact ratio. DCWP's rule defines it as the selection rate for a category divided by the selection rate of the most selected category (or, for tools that score, the share scoring above the median divided by the highest such share).¹⁷ CareerPlug's vendor-commissioned audit of its summary feature, prepared by Resolution Economics using the law's framework "solely for methodological purposes," shows the arithmetic: 1,105 of 1,470 female applications got a "meets" verdict (75.2%) and 888 of 1,224 male applications did (72.5%), so the male impact ratio is (888 ÷ 1,224) ÷ (1,105 ÷ 1,470) ≈ 0.965, reported as 0.97. For Asian applicants, 107 of 199 (53.8%) against 912 of 1,178 White applicants (77.4%) gives about 0.69, which the audit attributes to a single position at a single client.¹³ Categories under 2% of the data may be left out, and the same audit excluded 19,613 applications with unknown sex.¹⁵¹³

The law sets no pass mark. DCWP says it "does not require any specific actions based on the results of a bias audit."¹⁵ The rule describes its calculation as consistent with the federal Uniform Guidelines on Employee Selection Procedures,¹⁷ but the often-quoted 0.8 threshold, a rule of thumb from employment-discrimination cases,¹⁸ is not written into Local Law 144. Some posted summaries invoke it anyway: Pfizer's describes impact ratios of at least 80% as consistent with the law's "bias-free" requirements.⁴ How much the audits reveal is debated. A FAccT 2024 study in which 155 student investigators checked 391 employers found 18 posted audit reports and 13 transparency notices, and argued that because employers decide for themselves whether a tool is in scope, an absent audit cannot be read as non-compliance.¹⁸ In December 2025 the New York State Comptroller found DCWP's enforcement ineffective: DCWP received two complaints from July 2023 to June 2025, nine of 12 test calls to 311 never reached it, and where DCWP's review of the same 22 employers found one issue, the auditors found at least 17 instances of potential non-compliance.¹¹ This page explains the law; for what it requires of a specific employer, see DCWP's AEDT page.

7. Illinois, Colorado, the EU and existing discrimination law

Other rules regulate the same tools in different ways:

General anti-discrimination law applies regardless. In Mobley v. Workday, Inc. (N.D. Cal., No. 23-cv-00770), the court on May 16, 2025 granted preliminary certification of a nationwide collective of applicants aged 40 and over on an age-discrimination claim, finding they were similarly situated because of an alleged "unified policy: the use of Workday's AI recommendation system to score, sort, rank, or screen applicants." The order is not a ruling on the merits, and Workday can ask the court to decertify the collective after discovery.¹²