The Hiring Algorithm Is Shared. The Lawsuit May Not Be.

The case has reached the question that determines its scale

Mobley v. Workday began in 2023 with Derek Mobley's allegation that he was rejected, often within hours, from more than 100 positions at companies using Workday's platform. The amended complaint alleges discrimination based on race, age, and disability. Other named plaintiffs later joined. Their current theory is not that a generative model invented a reason to reject them. It is that algorithmic screening can operate as designed and still reproduce unlawful disparities.

On September 21, Reuters reported that the plaintiffs had asked U.S. District Judge Rita Lin to certify four subclasses covering Black applicants, women, applicants over 40, and applicants with disabilities. They point to statistical evidence and common features of the platform. Workday says its tools evaluate job qualifications rather than protected traits, that customers configure and use the product differently, and that the proposed groups cannot satisfy the requirements for class treatment. The court has scheduled a class-certification hearing for March 9, 2027.

That motion does not ask the judge to decide whether Workday discriminated. It asks whether the relevant claims can be adjudicated together. The difference is foundational. A class-certification ruling can determine the practical stakes of litigation, the evidence that matters, and whether alleged injuries scattered across many employers can be treated as the product of one system.

A common interface is not necessarily a common employment practice

Rule 23 begins with numerosity, commonality, typicality, and adequate representation. A damages class also ordinarily must show that common questions predominate over questions affecting only individual members and that class litigation is superior to other methods. The plaintiffs therefore need more than many people encountering a Workday-branded portal. They need a common contention capable of a common answer.

The proposed subclasses span different employers, jobs, qualifications, locations, configurations, and final hiring processes. Those variations support Workday's expected argument that an applicant rejected for one role cannot prove the claim of an applicant rejected for another by pointing only to the shared vendor. If employer A selects its criteria, employer B uses a different model or threshold, and employer C adds human review, the word platform can hide legally important differences.

But variability does not automatically end the inquiry. A shared ranking model, training method, assessment, cutoff logic, or recommendation system could supply a common practice even when customers select different jobs. Rule 23 does not require every fact to be identical. It requires the court to identify what question will drive resolution. The hard work is locating the decision point that is genuinely common rather than assuming that either the logo or the customer contract answers the problem.

Disparate impact law demands a particular practice, not a general suspicion

Title VII's disparate-impact provision requires a complaining party to identify a particular employment practice that causes a disparate impact. The respondent may defend the practice as job related for the position and consistent with business necessity. The statute generally requires each challenged practice to be analyzed separately, unless the elements of the decision process cannot be separated for analysis.

That text makes system architecture legally consequential. If a vendor's scoring function is separable from an employer's job criteria and interview decisions, the parties can fight over the impact and justification of that function. If the output emerges from an inseparable chain of vendor design, customer data, employer settings, and human choices, the statute permits analysis of the decision process as a whole, but the plaintiffs still must connect that process to the alleged disparity.

The first amended complaint alleges that Workday's selection tools disproportionately exclude protected groups and are neither job related nor consistent with business necessity. Those are allegations. Workday disputes them. The court's earlier decision allowing certain claims to proceed decided pleading sufficiency, not liability. The new certification motion raises a different question: whether the alleged practice and causal theory remain common when applied across the proposed subclasses.

Vendor liability does not erase employer responsibility

In 2024, Judge Lin held that the complaint plausibly alleged an agency theory because Workday's customers had delegated traditional hiring functions, including advancing or rejecting candidates, to its tools. She rejected the separate theory that Workday qualified as an employment agency on the allegations then before her and dismissed intentional-discrimination claims. That ruling kept parts of the case alive. It did not hold that every software vendor is an employer or that Workday violated the law.

The EEOC's 2024 amicus brief argued that federal antidiscrimination law can reach an agent performing delegated hiring functions and that an employer cannot avoid statutory duties by outsourcing them. The agency did not take a position on whether the plaintiffs' factual allegations were true. Its institutional point was narrower: the legal analysis should follow the function being performed, not stop at the boundary between customer and vendor.

That is the right frame for procurement as well. An employer cannot treat the contract as a transfer of accountability. A vendor cannot assume that customer customization converts every outcome into somebody else's decision. Responsibility can be distributed. The evidence must show who designed the model, chose the inputs, set the criteria, controlled the threshold, received the recommendation, could override it, monitored outcomes, and retained the records necessary to explain what happened.

The most important product feature may be the record it leaves behind

Applicants often cannot tell whether a person reviewed their materials, an assessment disqualified them, a ranking pushed them below a cutoff, or a recruiter never opened the file. That information asymmetry is not merely frustrating. It shapes whether a person can recognize a legal claim and whether any party can later test causation.

For organizations using automated screening, defensibility depends on versioned records. The useful evidence includes the model and assessment version, the customer's configuration, job criteria, data sources, accessibility accommodations, score or recommendation, threshold, downstream human action, override history, validation work, demographic monitoring, and retention rules. A generic statement that the tool does not use protected traits does not answer whether proxies or historical patterns produce a disparate effect.

The same records will matter on both sides of litigation. Plaintiffs need them to identify a common practice and measure its impact. Defendants need them to show that practices differed, that criteria were job related, that variations matter, or that another stage caused the result. A configurable product without configuration history gives everyone rhetoric and gives the court less evidence.

The court must decide where the system becomes one thing

The class-certification motion arrives at a useful moment. Automated decisions are spreading faster than the vocabulary used to describe them. Platform, AI, algorithm, assessment, ranking, recommendation, and rejection are often treated as synonyms. They are not. Each can refer to a different actor and a different stage in the employment process.

A ruling for either side will not settle the legality of AI hiring. If certification is granted, the plaintiffs will still have to prove impact, causation, and the absence of a sufficient statutory justification while Workday can present its defenses. If certification is denied, individual claims or narrower groupings may remain. The pending motion concerns litigation structure, not a final merits verdict.

Its broader significance is nevertheless clear. The law cannot evaluate automated decisions at the level of branding. It has to locate the practice. Sometimes the shared model will be the common act. Sometimes customer configuration will fracture the claim. The answer should come from evidence about how the system actually made decisions, not from whichever party finds the word customization more convenient.

This article analyzes allegations and pending proceedings in Mobley v. Workday, governing federal law, and public reporting. Workday denies wrongdoing. No court has found that its tools discriminated against the proposed classes, and the class-certification motion remains undecided.