AI resume screening that quotes the evidence, checks fairness first and leaves every decision to your recruiters
Every applicant read against the job, and every decision cited and checked for fairness.
See one case, screen by screen ↓AI Resume Screening screens and ranks applicants against a job's requirements with cited evidence for recruiters. Seven agents blind every resume, quote the line that meets each requirement, re-read borderline applicants and check the four-fifths rule before any bulk decline. The recruiter decides every applicant with a reason, and the hiring manager signs the shortlist.
How to choose AI solutions for back-office and enterprise work →
Updated
A corporate recruiter carries about 20 open requisitions, and the average job now draws 244 applications. The requirements were agreed loosely in a 30-minute intake meeting, knockout questions are answered yes by everyone, and the first screen is a few seconds per resume and a keyword search. Below a line, applicants are declined in bulk with one template, and the only record of why is a disposition code.
Nobody sees pass rates by group until the yearly bias audit, because voluntary self-ID sits in another system. New York City's Local Law 144 requires an independent bias audit and notice to candidates before any automated screening, the EU AI Act classes recruitment AI as high-risk, and an age-discrimination collective has been certified against a screening vendor. When a claim arrives, counsel rebuilds the reason for one rejection by hand, weeks later.
Estimated split for one role with a few hundred applicants, screened by hand and with the solution.
Seven agents blind, quote, re-read and check fairness; the recruiter decides and the hiring manager signs.
Wednesday 14 October 2026, at Kestrel & Vane Group in New York. REQ-4471, Senior Payroll Analyst in Group Payroll, closed at 23:59 last night with 286 applicants. Megan Hayes is the recruiter and Daniel Okoro, Head of Group Payroll, is the hiring manager. Here is how they work it with the agents, screen by screen, in the working solution.
Megan's morning opens on six decisions, not a pile of resumes. The agents read every applicant by 01:12 and say what they need from her: a fairness alert on REQ-4471 with age 40 and over at 0.74, traced to M4; 2 disagreements between the readers; hidden instructions stripped from 2 resumes. Beside it, the overnight feed reports what each agent did, with a link to the audit trail.
“Review agent re-read 64 blind: 7 corrected, including C-112 (9 EU entities).”
Before the role closed, the requirements agent turned the job description into five must-haves and three nice-to-haves, each with the evidence it accepts. It flagged three phrases: “5–8 years” (an upper bound reads as an age filter), “Bachelor's degree in accounting required” and “Native English speaker”. Megan and Daniel accepted the rewording and locked requirements v2 on 9 Oct.
The requirements agent never locks requirements; the hiring manager does.
Applications came from the careers site, job boards, referrals, agencies and internal applicants; 5 duplicates were merged. Names, photos, school names and graduation years were hidden before any agent read a resume. Three scanned resumes were read by OCR and sent to second review, and C-047 asked to be read by a person, so no agent reads that resume.
C-158: “White-on-white instruction stripped; screened on visible text.” The intake agent never drops an application.
The screening agent marked 2,280 cells, each with the line from the resume, its section, page and confidence. The review agent then re-read 64 applicants blind, without seeing the first reading. On C-112 the first reading had multi-country payroll not met at 0.58, reading “EU payroll” as one region; the second found “Ran monthly payroll for 9 EU entities across Germany, France, the Netherlands, Spain, Italy and Poland” and corrected it to met at 0.94. C-112 moves from rank 61 to rank 9.
Hover any mark and both readings are there, with which agent read each.
Of the 64 re-read, 55 agree, 7 are corrected and 2 are left for her. On C-087 the screening agent read statutory filings as met at 0.81; the review agent read “Reviewed each quarterly and year-end filing prepared by our payroll provider” as partly met. Megan agrees with one reading, with a reason, or opens the resume.
“A person settles; neither agent can.”
Megan selects the 192 applicants who meet three or fewer must-haves and clicks Decline. Before anything happens, the fairness agent shows its plan and runs it: applicants aged 40 and over would be screened in at 12 of 86 (14.0%) against 22 of 117 (18.8%) under 40, an impact ratio of 0.74, below the four-fifths line. Nothing is declined. She can cancel, or ask Daniel.
“The job may well count them: ask the hiring manager. Nothing is declined until you decide.”
The agent removed one requirement at a time and re-counted. M4, an enterprise payroll system, explains the gap: 11 applicants describe in-house or mainframe payroll engines for 1,000+ employees, which v2 read as “not an enterprise payroll system”. If they count, the ratio is 0.91. It reads self-ID only as totals, from 203 of 286 applicants who shared it.
The fairness agent never changes a requirement itself, and never proposes a change that is not job-related.
Megan clicks “Ask Daniel to clarify M4”. Daniel replies, signs requirements v3 with reason RQ-02, and the 11 are re-read: 52 now meet every must-have, 45 meet 4 of 5 and 189 meet 3 or fewer. The strip turns green at 0.91, above the 0.80 line in every group.
Daniel: “Yes, if it paid 1,000+ people.”
Megan selects by rule the 18 who meet every must-have and two or more nice-to-haves, and advances them with reason A-01. The fairness agent re-checks the selection: the lowest ratio stays 0.91, and with 18 selections it adds Fisher's exact test, no significant difference (p = 0.62). The communication agent drafts 18 phone-screen invitations with slots from her calendar; they go only when she clicks “Approve and send 18”.
The decline notices wait for Priya Raman's 10% spot check. No rejection is automatic.
Daniel sees the 18 as blind IDs with their met requirements and the lines behind them, not names or schools. He marks each Interview or Not now with a reason and signs the shortlist. Names unlock for the people he signed for, and the unlock is logged.
“Tell me who can run multi-country payroll, not who went to which school.”
The same role, seen by the five people who carry it — what their week looked like, and what it looks like now.
Seven agents blind, quote, re-read and check fairness; the recruiter decides and the hiring manager signs.
Turns the job description into testable must-haves and nice-to-haves with the evidence each accepts, flags proxy wording before lock, and measures how the hiring manager reads a few blind profiles.
Pulls applications from every source, merges duplicates, reads scans, hides identity and its proxies before any other agent reads, strips hidden text and holds late applications.
Marks every locked requirement met, partly met or not met, with the quoted line from the resume and a confidence.
Re-reads blind, without the first reading, every applicant at 4 of 5 and every mark under 0.75; corrects with a quote, or disagrees out loud for the recruiter to settle.
Computes selection rates and impact ratios by sex, race/ethnicity, their intersection and age 40 and over at every stage, runs a what-if before any bulk decision and traces a gap to the requirement that drives it.
Drafts invitations, updates, decline notices, the notice before use and explanations on request, from approved templates and the record; routes alternative-process and data requests.
Keeps the yearly bias-audit data set for the independent auditor and the evidence pack for each role.
Reason codes, sign-off. The agents propose; a named person decides.
Ask in plain words, or say what to prepare. Every answer cites the record, and nothing is decided for you.
Why was C-112 corrected?
The screening agent marked M2, multi-country payroll, not met at 0.58: it read “EU payroll” as one region. The review agent re-read the resume blind and quoted “Ran monthly payroll for 9 EU entities across Germany, France, the Netherlands, Spain, Italy and Poland (2,400 employees)”, so it is met at 0.94. C-112 now meets all five must-haves and ranks 9 of 286.
Which applicants meeting every must-have have a payroll certification?
Of the 44 who meet every must-have, 16 list a current payroll certification (N1, a nice-to-have). A certification ranks an applicant higher only among equals; it never screens anyone out.
Who would be affected if I decline everyone missing M2?
147 applicants do not meet M2, multi-country payroll. Declining them would put age 40 and over at 0.72, because the M4 alert is still open. C-087 and C-205 are kept out until their disagreements are settled, and C-047 was read by a person. Nothing is declined.
Draft the bias-audit summary for 2026
From 1 Jan to 13 Oct 2026 the screening tool read 14,920 applicants on 61 roles; 10,722 shared self-ID. By sex the impact ratio is 0.98; by race/ethnicity the lowest is 0.93 (Hispanic or Latino). 9 alerts were raised before bulk decisions and 8 were fixed by a job-related change. It is in Reports for Elena Marchetti to review before the independent auditor sees it.
13 screens from the working solution, on its sample data. Pick one to see it large.
The rules every role is screened against are settings, changed by the TA lead or the HRIS administrator, versioned and logged.
| Setting | Default | Choose from |
|---|---|---|
| No automatic rejection | Always on | Locked on: every decline is a person's decision with a reason code |
| Second review trigger | Every 4 of 5, and any mark under 0.75 | Set the confidence below which a mark is re-read |
| Blind fields | Name, photo, school, graduation year, address, pronouns, age, nationality | Each field on or off |
| Hidden text in resumes | Strip and flag | Strip and flag · Strip only |
| Late applications | Hold; TA lead may admit with a reason | Hold; TA lead may admit with a reason · Hold and decline |
| Decline spot check | TA lead checks 10% | TA lead checks 10% · TA lead checks 20% |
| Fairness alert | Below 0.80; warn below 0.85 | Set both impact-ratio lines |
| Small numbers | Fisher's exact test under 30 selections; categories under 2% counted, not scored | Set the selection count and the share excluded from ratios |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (an applicant-tracking vendor benchmark of 6,000+ employers, 244 applications per job, 2025; SHRM 2025 Recruiting Benchmarking, 8–9 days to screen, via a secondary source; Harvard Business School and Accenture, Hidden Workers: Untapped Talent, 88%). NYC Local Law 144 and the four-fifths rule (29 CFR 1607.4D) are named as context. Kestrel & Vane Group, its roles, applicants and the people named on this page are fictional — characters and sample data in the working solution.
AI Resume Screening screens and ranks applicants against a job's requirements with cited evidence for recruiters. Seven agents blind every resume, quote the line that meets each requirement, re-read borderline applicants and check the four-fifths rule before any bulk decline. The recruiter decides every applicant with a reason, and the hiring manager signs the shortlist.
Every requirement mark carries the line from the resume that meets it, with its section, page, confidence and the agent that read it. When the second reader corrects a mark, both readings stay on the record. Asked why, the communication agent drafts a requirement-level explanation from the same evidence for the recruiter to approve.
No. “No automatic rejection” is always on and cannot be switched off. The recruiter advances, holds or declines each applicant with a reason code, declines wait for the TA lead's spot check, and a decline is blocked while a disagreement or hidden-text flag is open.
Before any bulk decline, the fairness agent computes selection rates and impact ratios by sex, race/ethnicity, their intersection and age 40 and over from voluntary self-ID totals. If a ratio falls below 0.80, it traces the gap to the requirement driving it and proposes a job-related fix for the hiring manager to sign. With small numbers it also runs Fisher's exact test.
Yes. The intake agent hides names, photos, school names, graduation years, address and other proxies before any agent reads a resume, and strips hidden text written to game an AI reader. Self-ID is held in a separate store and the fairness agent sees only totals. Names unlock after the hiring manager signs the shortlist, and the unlock is logged.
The audit agent keeps the year's selection rates and impact ratios by category and intersection, with categories under 2% counted rather than scored, ready to export for your independent auditor. The communication agent drafts the notice before use, and roles missing it are flagged before screening can start.
Yes. The hiring manager locks the requirements and signs the shortlist, the recruiter settles every disagreement between the readers and decides every applicant, and only the TA lead can proceed past a fairness alert, with a written, job-related reason. The agents never send a message without a person's approval.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your job architecture, screening policy, message templates and a sample of past roles — and it goes live once every quality gate has passed. We will walk you through it on one of your own closed roles first.
Every HR answer cited to the policy that applies; sensitive cases always with a person.
1-min demoWorkplace Investigation ManagementEvery finding tied to quoted evidence and tested for gaps before anyone signs.
1-min demoForm I-9 ComplianceAn estimated 86% less fine exposure, every correction signed and never back-dated.
1-min demoWe'll run AI Resume Screening on one of your own closed roles, its job description and its applicants.