Auditing AI Skills-Inference and Internal-Mobility Engines Before Promotion Slates Lock

Why HRIS skills clouds are regulated selection procedures, not neutral directories

When an HRIS infers skills from workplace activity and ranks internal candidates for stretch roles or promotions, federal law treats that algorithm as a test. Vendor labels like "career discovery" do not exempt promotion slates from Title VII scrutiny.

Enterprise HR platforms rarely sell "automated promotion rejection" tools. Instead, they ship modules branded as a "skills cloud," "talent marketplace," or "opportunity graph." Once enabled, the model ingests job histories, self-tagged profiles, Jira tickets, code commits, CRM closes, and calendar metadata. It then assigns probability weights to dozens of inferred adjacent skills and surfaces a ranked match score—say, 88% ready—when a director opens a Level 5 lead requisition.

The legal boundary does not turn on whether a human manager clicks the final approval button. Under the Uniform Guidelines on Employee Selection Procedures (29 CFR Part 1607), specifically 29 CFR § 1607.16Q, a "selection procedure" includes any measure, combination of measures, or scored ranking used as a basis for hiring, promotion, demotion, membership, referral, retention, or licensing and certification—as well as selection for training or transfer if it leads to any of those decisions.

State and municipal regimes draw the exact same perimeter for AI and changing work. New York City Local Law 144 (N.Y.C. Admin. Code §§ 20-870 to 20-874; 6 RCNY § 5-300 et seq.) explicitly defines an "employment decision" to cover screening candidates for employment or employees for promotion within the city. In California, the Civil Rights Council’s Automated-Decision Systems regulations (2 CCR §§ 11008.1 and 11013, effective October 1, 2025) sweep in any computational process that makes or facilitates human decision-making regarding promotion, training, performance evaluation, or compensation, while extending mandatory record retention from two years to four years.

How inferred skills and recency decay penalize leave-takers and part-time cohorts

Activity-scraped skills models confuse recent digital volume with actual capability. Without leave-window normalization, workers returning from FMLA, PWFA, or ADA leave suffer steep algorithmic score drops.

Look closely at how commercial skills-inference engines compute proficiency scores. Most combine two signals: lexical co-occurrence (if an employee holds Skill A and Skill B, the model infers Skill C at a 0.72 confidence weight) and rolling recency decay (tickets closed, commits merged, or client decks shipped over the trailing 90 or 180 days). Both mechanisms introduce systematic demographic skew long before a calibration committee meets.

First, lexical inference rewards self-promotional profile writing and male-skewed title inflation in historical resumes. In internal audits across product and engineering organizations, employees who list 25 unverified buzzwords on internal profiles trigger two to three times as many high-confidence "adjacent skill" inferences as peers who list only certified competencies. If managers filter internal slates at an 80% match threshold, unverified ontology expansion quietly determines who gets interviewed.

Second, rolling 90-day and 180-day activity windows treat protected statutory leave as skill decay. Trace the arithmetic on a concrete promotion cycle:

• The cohort denominator: A 420-person software and product division runs semi-annual senior promotions. Exactly 64 employees took at least six weeks of FMLA, PWFA postpartum recovery, or ADA medical leave—or worked an approved reduced-hour caregiver schedule—within the trailing 180-day scoring window. The remaining 356 employees worked continuous full-time schedules.
• The algorithmic velocity drop: Because the HRIS skills engine weights trailing 180-day ticket and repository activity at 35% of its "role readiness" index without pausing the denominator during approved leave, a 12-week parental leave cuts raw activity volume by roughly 46%.
• The selection-rate collision: Among the 356 continuous-schedule employees, 50 clear the 80-point shortlist cut and earn promotions—a selection rate of 14.04% (50 / 356). Among the 64 leave-taking and part-time employees, predominantly birthing mothers and employees with disabilities, only 4 clear the shortlist cut—a selection rate of 6.25% (4 / 64).
• The four-fifths calculation: Dividing 6.25% by 14.04% yields an impact ratio of 0.445 (44.5%), far below the 0.80 (80%) benchmark in 29 CFR § 1607.4D.

Under Title VII (42 U.S.C. § 2000e-2(k)(1)(A)), once a disparity of that magnitude emerges across sex, pregnancy, or disability cohorts, the burden shifts to the employer. You must prove that the scoring practice is job-related for the target position and consistent with business necessity—and even then, plaintiffs win by showing a less discriminatory alternative (such as pausing decay clocks during leave) that the employer refused to adopt. Similar telemetry traps arise when productivity scores drive layoffs, detailed in our audit guide for AI productivity metrics in RIF selection.

Why vendor whitepapers fail 29 CFR Part 1607 validation standards

A vendor’s global accuracy slide is not a local validation study under 29 CFR § 1607.5. Employers remain directly liable when unvalidated skill weights screen out protected internal groups.

When people operations teams ask an HRIS vendor how its promotion-matching score was validated, account executives usually send a glossy global whitepaper showing "91% skill-taxonomy precision across 40 million profiles." Hand that PDF to an industrial-organizational psychologist or an EEOC systemic investigator and watch how fast it collapses.

Under 29 CFR § 1607.5 and § 1607.14, selection validation is never abstract. It requires empirical proof tied to the actual job family:

1. Content validity (§ 1607.14C): You must link every scored skill directly to observable behaviors and work products required on Day 1 of the target role, backed by a structured job analysis. Under § 1607.14C(1), content validity is inappropriate for traits or inferred abstract capacities that employees are expected to learn after promotion.
2. Criterion-related validity (§ 1607.14B): If the vendor claims an inferred "leadership agility" or "cross-functional velocity" score predicts promotion success, there must be a statistically significant correlation study comparing those pre-promotion scores against unbiased post-promotion job performance measures, including a fairness review under § 1607.14B(8).
3. No cooperative shield (§ 1607.7): Employers cannot outsource Title VII or FEHA liability to software licensors. Both EEOC technical assistance and California’s 2 CCR § 11008 confirm that an employer using a third-party automated-decision system answers directly for any unjustified disparate impact.

The five-step internal-mobility AI audit and portable worksheet

Audit the candidate pool before managers ever see ranked match percentages. Strip unverified inferences, freeze leave windows in velocity formulas, and run four-fifths impact checks across sex, race, age, and leave cohorts.

Run this five-step control gate before opening any semi-annual promotion cycle or internal talent-marketplace slate.

Copy-ready vendor data demand and pre-slate sign-off log

Require People Analytics and your HRIS vendor to sign this verification log before turning on scored internal shortlists. If the vendor refuses feature-weight transparency, keep automatic ranking off.

INTERNAL-MOBILITY & SKILLS-ENGINE VALIDATION LOG — CYCLE [2026-H2]
System / Module: [HRIS Vendor Name & Model Version]
Target Decisions: [Internal Transfers / Stretch Slates / L4-to-L5 Promotions]
Jurisdictions Covered: [US Multi-State, including CA (2 CCR § 11008) & NYC (LL 144)]

1. FEATURE & INFERENCE HYGIENE
[ ] Probabilistic / unverified "inferred skills" excluded from hard eligibility cutoffs.
[ ] Target role skill criteria limited to <= 7 observable competencies tied to written job analysis (29 CFR § 1607.14C).
[ ] Trailing activity / velocity windows exclude FMLA, PWFA, ADA, and USERRA leave days and normalize for < 1.0 FTE schedules.

2. PRE-SLATE DISPARATE-IMPACT TEST (29 CFR § 1607.4D)
Total Eligible Pool (N): [___]   Shortlist Cutoff Score: [___]
- Highest-Selected Reference Group Selection Rate: [___]% ([n]/[N])
- Lowest Sex / Race / Intersectional Cohort Rate:  [___]% ([n]/[N]) -> Impact Ratio: [0.__]
- Recent Leave / Accommodated Cohort Rate:         [___]% ([n]/[N]) -> Impact Ratio: [0.__]
*Rule: Any ratio < 0.80 or statistically significant gap (p < 0.05) freezes automated ranking until biased inputs are removed.*

3. STATUTORY NOTICE & ARCHIVE
[ ] NYC LL 144 independent annual bias audit summary posted and 10-business-day employee notice issued (if NYC roles covered).
[ ] Full input snapshot, score vector, manager overrides, and validation memo archived for 4 years (2 CCR § 11013).

Signed (People Analytics): ___________________   Date: __________
Signed (Employment Counsel): _________________   Date: __________