Intelletto · Field Analysis Human Capital Intelligence

Scoring the
Scorers.

Roughly 99% of the Fortune 500 filter applicants through an automated score. The number looks like a measure of candidate quality. It isn't. It measures how much a résumé resembles a job description — and that is a different thing, with a much lower ceiling. Here is how well ATS scoring actually works, rated against what effectiveness really means.

Legacy keyword scoring
D Taleo · iCIMS · Workday · Greenhouse · Lever. A lexical-compliance gate.
Modern AI talent intelligence
C+ Eightfold · Phenom · Beamery · HiredScore. Better recall, same input ceiling.
The category

One name for two different machines

The market says "ATS scoring" as if it were a single technology. It is two — and they fail in different ways.

Generation one · lexical

Keyword overlap

Parse the résumé into fields, extract keywords, compute overlap with the job description. Output a match %. Below the cutoff, no human ever looks.

What it actually measures

Whether the right words appear in the right place. A skill in a dated job bullet outscores the same skill in a flat list.

TALEO · iCIMS · WORKDAY (parse) · GREENHOUSE · LEVER
Generation two · semantic

Embedding similarity

Map profile and role into vector space, score proximity. Recognises that "Software Engineer" ≈ "Computer Programmer" and infers unlisted skills.

What it actually measures

Whether the résumé looks like the job — with real nuance. Still not whether the person can do the job.

EIGHTFOLD · PHENOM · BEAMERY · WORKDAY HIREDSCORE

Mechanics per Jobscan, Resume Optimizer Pro, CVCraft (commercial reverse-engineering) and vendor documentation.

The ceiling · the argument that matters most

Every strong predictor sits out of reach

Personnel-selection science ranks how well each method predicts job performance. Sackett et al. (2022) re-ran the canonical evidence and reordered the field. Watch where the résumé score can actually see.

Structured interviewHighest validity
Work-sample testHigh
Job-knowledge testHigh
Empirically-keyed biodataModerate–high
↓ All a résumé score can actually see ↓
Years of experienceLow
Education / credentialsLow
Self-reported skills / keywordsProxy · unvalidated

The irony is the whole story. The methods that actually predict performance — structured interviews, work samples, job-knowledge tests — cannot be run on a PDF. What's left at the screening stage sits at the bottom of the hierarchy, and is asked to do the heaviest cutting in the funnel.

Bar lengths are relative and illustrative of the post-Sackett ordering, not exact coefficients.

Sackett, Zhang, Berry & Lievens (2022), revising Schmidt & Hunter (1998): validity estimates cut ≈.10–.20; structured interviews now top-ranked.

The scorecard

Effectiveness, dimension by dimension

"How effective is the score?" has seven answers, not one. Neither generation earns the precision it displays.

Dimension Legacy keyword AI talent intel.
Predictive validityDoes the score predict performance? D−C
Construct validityDoes it measure capability? DC+
ReliabilitySame input → same score? D+B−
RecallDoes it find the qualified? FC+
Fairness / disparate impactEquitable across groups? DC−
TransparencyCan you see how it decided? FC
Provenance / defensibilityPer-decision record? FC−
F — failing D — poor C — partial B — adequate
The recall hole · what regulators attack

Built to exclude, not to find

Screening is configured against a long list of disqualifiers, not for a short list of must-haves. The cost lands on qualified people.

88%
of employers admit qualified, high-skilled candidates are screened out simply for not matching the exact job criteria.
HARVARD BUSINESS SCHOOL × ACCENTURE — "HIDDEN WORKERS" (2021), 2,250 EXECS · US/UK/DE
27M
"hidden workers" in the US alone — able and willing, filtered out by logic that scores what they lack.
HIDDEN WORKERS: UNTAPPED TALENT (2021)
~50%
of employers auto-eliminated candidates for an employment gap of six months or more.
HIDDEN WORKERS (2021)
Nurses screened for "computer programming" — they meant patient-data entry
Line-repair crews screened for "customer service"
Retail roles requiring "floor-buffing"
"BS""Bachelor's degree" to the parser
"75% of résumés are never seen by a human."

Don't use this number. It traces to a 2012 sales pitch from Preptel — a company that folded in 2013 — with no published methodology. The 88% figure is the defensible one.

The honest counter-reading: the failure is human-configured criteria, not the software executing them. The category optimises for efficiency at the cost of recall.

The bias signature · now litigated

The same résumé, a different name

University of Washington researchers varied 120 race- and gender-associated names across 550+ real résumés, ranked by three production-grade models against 500+ real jobs.

White-associated namesfavoured
85%
Female-associated namesfavoured
11%
Black male-associated namesfavoured over white male
~0%

The human-in-the-loop defence is weakening. A 2025 UW follow-up found people mirror biased AI recommendations — the score doesn't just sit there, it propagates into the decisions downstream. Amazon scrapped its own recruiting AI in 2018 for the same failure mode.

Wilson & Caliskan, AAAI/ACM AIES 2024 · UW human-mirroring study, 2025 (N=528).

The reckoning · vendor liability is settled

Mobley v. Workday

The first large-scale test of AI hiring tools in federal court — and it reached the vendor, not just the employer.

1.1B
applications Workday represented were rejected through its tools in the relevant window — the scale at which an error rate compounds.
WORKDAY COURT FILINGS, MOBLEY v. WORKDAY (N.D. CAL.)
JUL 2024
Motion to dismiss denied on an "agency" theory — a vendor whose tool participates in the decision can be liable.
MAY 16 2025
Nationwide ADEA collective conditionally certified. Applicants 40+ screened since Sept 24, 2020.
DEC 2 2025
Notice plan approved; collective explicitly includes HiredScore; 10,000+ employer-clients exposed.
MAR 6 2026
Court rejects Workday's argument that the ADEA doesn't protect applicants. Discovery reaches back to 2020.

Parallel: Harper v. Sirius XM (2025, Title VII); Eightfold faces its own applicant suit; EEOC v. iTutorGroup already settled (AI age discrimination).

The paper shield · why a badge isn't enough

"We ran a bias audit" is a fragile defence

NYC Local Law 144 was the first regime to mandate public bias audits of hiring algorithms. In practice the mechanism barely bites.

Each square = one employer studied. Filled = posted a bias-audit report.

18 / 391

employers had posted an audit report (and just 13 a transparency notice). Cornell calls the gap "null compliance": employers hold enough discretion over scope that a missing audit can't even be read as non-compliance.

1
Issues the enforcing agency found across 32 companies
17
Issues the State Comptroller found in the same companies

Cornell "Null Compliance" (ACM FAccT 2024) · NY State Comptroller audit of DCWP enforcement (Dec 2, 2025). Nearly all posted audits reported impact ratios above 0.8.

Where Intelletto breaks the pattern

Seven things the ATS never computes

A résumé score asks one question: does this document resemble this job? Intelletto runs a different operation — it decomposes, grades, fuses and models the candidate across seven dimensions the applicant tracking system doesn't attempt.

01

Job description

Ingest as-is

Takes the JD as written — often bloated and brittle. The root cause of the recall failure.

Engineered artifact

Treats the JD as a first-class engineered artifact. Role-normative enrichment scores against what the role should require — JD alignment plus a market-expected skills delta.

02

Certifications & skills

Flat keyword

One certification, one keyword. Matched or not matched — no structure beneath it.

Hard / soft decomposition

Decomposes certifications and skills into their hard and soft dimensions, each scored on its own terms against the role.

03

Skill proficiency

Present / absent

Binary. Does the word appear on the page, and ideally in a dated bullet?

Graded with evidence

Grades proficiency by level — drawn from recency, duration and corroboration, not the candidate's self-assertion.

04

Career trajectory

Static snapshot

A flat list of past titles. Direction, velocity and shape of the career are invisible.

Modelled as signal

Models progression, velocity and direction as a predictive signal in its own right.

05

Data fusion

Résumé + form

One self-reported document plus an application form. That is the entire input — and the entire ceiling.

Multi-source · provenance-stamped

Fuses professional, code and social signal from multiple corroborating sources, every record provenance-stamped. This is what lifts the score off the résumé-only ceiling.

06

Culture fit to company

Not attempted

Most vendors explicitly exclude it — scoring only literal, job-relevant fields.

Company-contextual · governed

Models fit to the specific company's context — held inside the audit and compliance layer so it never becomes a proxy for a protected trait.

07

Audit & compliance

Annual badge

A once-a-year bias audit at best — under-enforced, gameable, no per-decision trail.

Per-record provenance

Answerable by decision, built for NYC LL144, the EU AI Act and EEOC from the substrate up.

The one claim to earn by measurement

Data fusion, proficiency grading and trajectory modelling lift Intelletto off the résumé-only input that bounds the ATS category — so the ceiling on the panel above is the ATS's ceiling, not Intelletto's. One discipline remains, and it is a position of strength: criterion validity against real outcomes. Run the study at production scale — score against actual performance and retention, publish the methodology, including the shadow-compare deltas. That turns the category's universal weakness (asserted, retention-keyed validity claims) into the one place Intelletto proves what the rest only state.