- A court just ruled AI hiring vendors are legally "agents" — liable for discrimination like employers themselves.
1.1 billion rejected applications. A nationwide class action. The "we just sell software" defense is dead.
🧵
Mobley v. Workday changed everything. - One applicant. 100+ rejections. Many within minutes, outside business hours.
The court's conclusion: Workday wasn't a neutral tool. It was performing the employer's core hiring function.
The judge drew a sharp line.
A spreadsheet? Simple tool. No liability. - An AI that scores, ranks, and auto-rejects candidates? That's an agent — subject to Title VII, the ADA, and the ADEA.
In May 2025, the court certified a nationwide collective action for everyone over 40 rejected through Workday's platform since 2020. - The scale is staggering. Workday disclosed ~1.1 billion application rejections during that period.
Here's what most people miss: the AI never has to use "age" as a variable.
It learns proxies.
@aol.com email? Legacy tech skills? 15+ years experience? - The algorithm infers age without ever seeing it directly.
This is where LLM wrappers fail catastrophically.
A probabilistic text predictor can't explain WHY it rejected someone. It guesses the next token — it doesn't reason about fairness. - You can't defend "vibes" in federal court.
Our position: recruitment AI needs deterministic logic, not stochastic generation.
The LLM translates. A symbolic engine decides. A knowledge graph grounds every judgment in auditable rules.
Every rejection gets a logic trail. - The EEOC's Four-Fifths Rule is now the standard.
If your AI selects a protected group at less than 80% the rate of the highest group — that's adverse impact.
And the employer is liable even if the vendor built it. - We use adversarial debiasing during training — a second model tries to predict protected characteristics from the output.
If it succeeds, the primary model gets penalized.
Bias isn't audited after the fact. It's architecturally prevented. - The real risk isn't the $1,500/day NYC fines.
It's your company's name on a court-ordered notice sent to millions of rejected applicants — each one a potential plaintiff.
Ignorance is not a defense anymore. - If your AI vendor can't explain exactly why a candidate was rejected — with a traceable logic path — who carries 100% of the legal risk?
#AIRecruitment #EnterpriseAI - We wrote the full analysis — the case law, the technical failure modes, and the architecture that actually holds up in court.
https://veriprajna.com/whitepapers/algorithmic-agent-navigating-liability-deep-ai-recruitment