The context
Derek Mobley is a Black man over 40 with disclosed diagnoses of depression and anxiety. Over two years he applied to more than eighty positions at companies including Intuit, Nvidia and Citi. Each company used Workday’s applicant tracking system for initial screening. Each rejected him.
Mobley filed suit in August 2023 in the Northern District of California. The claim was not that any individual employer discriminated. The claim was that Workday’s AI-assisted screening tools were systematically filtering him out on the basis of his race, age and disability across every company he applied to. The suit named Workday as the defendant, not the employers.
In February 2024, Judge Rita Lin denied Workday’s motion to dismiss. A federal court held, for the first time, that a third-party AI hiring tool vendor could be liable as an “agent” under Title VII of the Civil Rights Act, the Age Discrimination in Employment Act (ADEA) and the Americans with Disabilities Act (ADA). Workday’s defense was that it is a software vendor, not an employer. The court found that Workday’s tools exercise a function historically belonging to employers (initial applicant screening) and that Workday exercises meaningful control over how that function operates.
The case had not reached trial as of this writing. But the “agent” theory has survived the motion to dismiss threshold. That is the durable result.
The three failures, and what each one actually cost
Every company named in the complaint made the same three mistakes. The cost to prevent any one of them was measured in hours. The cost of not preventing them is now measured in legal exposure.
| Failure | When it happened | Prevention cost | Cost of not preventing |
|---|---|---|---|
| No adverse impact data in vendor RFP | Vendor selection, 2021-2022 | 2-4 hours: add 7 questions to existing RFP | Federal lawsuit, 2023; defense costs estimated $500k-2M+ |
| No demographic rejection monitoring | Go-live through 2023 | 1 analyst-day per quarter | 2+ years of undetected discrimination across 80+ candidates; reputational exposure |
| No indemnification clause for AI screening decisions | Contract signing, 2021-2022 | One legal redline pass, 2-4 hours | No contractual recovery from Workday for defense costs or damages |
The four steps below close all three gaps. Download the printable checklist to fill in with your team: AI Hiring Tool Control Checklist (PDF, 6 pages).
Step 1: Audit your next vendor before you sign
See also: How do you audit an AI system?
The seven questions below belong in every RFP for an AI hiring tool. They take 20 minutes to add. A vendor who refuses to answer questions 1, 2, or 4 is a disqualifying response.
| # | Add this to your RFP |
|---|---|
| 1 | Provide your adverse impact analysis for AI screening for the 12 months ending [current month], showing pass rates by race, sex, age bracket (over/under 40), and disability status against your customers’ applicant pools. |
| 2 | What is your retest cadence after a model update that could affect bias? Provide results from your last three model updates. |
| 3 | List every configuration parameter available to deployment teams that could affect candidate scoring or ranking. |
| 4 | Describe your training data. What percentage was supplied by existing customers? What was the demographic composition of historical hiring decisions in that data? |
| 5 | What is your process for notifying customers of material model updates? Define “material” in writing. |
| 6 | Have any customers raised a bias concern with your ATS AI tools in the past 24 months? How was each one resolved? |
| 7 | Will you commit to a contractual adverse impact testing requirement, with results shared annually? |
If you are in an active contract with an AI hiring tool vendor and have never asked these questions, ask them now in writing. The response, or the absence of a response, is relevant if a claim is ever made.
Step 2: Audit your contract for the indemnification gap
Procurement and legal sign-off is part of your AI governance structure.
Standard enterprise SaaS contracts contain a limitation of liability clause and an indemnification carve-out. Together, these two clauses mean that even if your vendor’s AI tool produced discriminatory results, the vendor likely has no contractual obligation to cover your defense costs or damages. Most HR procurement teams do not catch this because they are checking a contract for SLA and data residency, not for civil rights liability.
| Contract clause | What the standard SaaS contract says | What you need instead | Vendor resistance |
|---|---|---|---|
| Limitation of liability | Capped at fees paid in prior 12 months | Uncapped for discrimination claims arising from vendor AI decisions | High |
| Indemnification scope | Excludes discrimination and civil rights claims | Covers claims arising from the vendor’s AI screening decisions | Very high |
| Model update notification | No obligation | 30 days written notice before any material model update | Medium |
| Bias testing | No obligation | Annual adverse impact analysis delivered to customer | Low-Medium |
| Configuration audit trail | Vendor owns and controls records | Customer right to retrieve full configuration history on request | Low |
| Training data disclosure | Confidential, no disclosure | Annual disclosure of demographic composition of training data | High |
Very high resistance clauses are negotiable on contracts above approximately $500k annual value. Below that threshold, alternative leverage points include: a mutual right to terminate if adverse impact thresholds are breached; a pricing reduction clause tied to bias test results; or a requirement that the vendor provide bias insurance documentation.
If your contract is already signed and contains none of these clauses, flag it to legal as a known risk and negotiate them into the next renewal.
Step 3: Run the four-fifths rule on your current ATS data
The Uniform Guidelines on Employee Selection Procedures (EEOC, 1978) require any employment selection procedure to be validated against adverse impact. The four-fifths rule is the standard threshold. It applies to AI screening tools. Here is how to run it.
The formula: selection rate for group X ÷ selection rate for highest-selected group
If the result is below 0.80, you have a four-fifths rule breach that requires review.
Step-by-step:
- Pull from your ATS: total applicants and screening pass-throughs for the last 12 months, broken down by race, sex, age (over/under 40), and disability status (if disclosed).
- Calculate selection rate for each group:
passed screen ÷ total applicants in group. - Identify the group with the highest selection rate.
- Divide each other group’s selection rate by the highest. Flag any ratio below 0.80.
- If you find a breach: document it immediately, involve legal, and determine whether the breach existed in prior periods.
What a breach looks like in practice:
| Group | Applicants | Passed screen | Selection rate | Ratio to highest | Compliant? |
|---|---|---|---|---|---|
| White | 1,200 | 360 | 30.0% | 1.00 (reference) | Yes |
| Asian | 300 | 87 | 29.0% | 0.97 | Yes |
| Hispanic | 240 | 62 | 25.8% | 0.86 | Yes |
| Black | 180 | 36 | 20.0% | 0.67 | No (threshold: 0.80) |
| Over 40 | 400 | 96 | 24.0% | 0.80 | Borderline (review) |
| Disability disclosed | 60 | 9 | 15.0% | 0.50 | No (threshold: 0.80) |
The 0.67 and 0.50 ratios in this example are exactly the pattern Mobley alleges. If data like this exists in Workday’s customer dashboards and no one acted on it, that is relevant to the liability question.
Do not run this calculation once. Run it quarterly and log the results. The log becomes your documented monitoring record.
Step 4: Set up the three monitoring checkpoints
Adverse impact monitoring for an AI hiring tool is not a model performance problem. It is a data analysis task. One person, one day per quarter, three numbers to check.
| Checkpoint | What to measure | Threshold | Action if triggered |
|---|---|---|---|
| Screening stage | Rejection rate by demographic group, four-fifths rule | Any ratio below 0.80 | Suspend AI screening, involve legal, run root cause |
| Configuration drift | Have any screening parameters changed since last review? | Any undocumented change | Document the change, require review sign-off retrospectively |
| Vendor model updates | Did the vendor update their model since last quarter? | Any update without documented re-test | Request updated bias analysis from vendor before next screening cycle |
The three checkpoints take under two hours if your ATS has a reporting export. If your ATS does not expose rejection data by demographic group, that absence is itself a finding: you are operating a screening system you cannot audit.
Step 5: Brief your recruiting team on what AI can and cannot decide
This briefing is the practical implementation of an AI acceptable use policy for your recruiting function.
The legal exposure in this case is not only about the vendor. It is about whether your team understood the tool, used it within its intended scope, and had a human override path. If you cannot answer yes to all three, your team is exposed, not just your vendor.
Run this briefing with every recruiter who uses the ATS. It takes thirty minutes. Document that you ran it and when.
| Topic | What to tell your team | What your team should be able to say back |
|---|---|---|
| What the AI screening tool does | ”It ranks or filters candidates based on signals in the application. It is not making the hiring decision. You are." | "I use it as a first sort. The final shortlist is mine.” |
| What it cannot see | ”It cannot see photos, names, or demographic data directly. But it can pick up proxies: graduation year, address, school name. These can correlate with protected characteristics." | "I know the output can be biased even if the input looks neutral.” |
| When to override | ”If a candidate does not pass the AI filter but you have reason to believe they are qualified, you can and should flag them for review. Document why." | "I know the override path and I have used it.” |
| What to report | ”If you see a pattern in rejections that concerns you (a role where all shortlisted candidates look similar), report it to your manager. That is how we detect bias before it becomes a claim." | "I know who to tell if something looks off.” |
| Who owns the outcome | ”If a rejected candidate files a discrimination claim, they are claiming the decision was discriminatory. The AI ranked them out. You used that ranking. Both are relevant. The company is liable." | "The decision is mine. I cannot say the software did it.” |
The question that tells you if your team is ready: Ask each recruiter: “If a candidate complained that our AI screening tool treated them unfairly, what would you do?” If the answer is not “I would escalate to HR and document the candidates’s information and the AI output immediately,” your team is not briefed.
What to check quarterly, assigned by role
Adverse impact monitoring only works if someone owns each task. This is the assignment table.
| Task | Owner | Frequency | Output |
|---|---|---|---|
| Pull ATS rejection data by demographic group | HR analyst or TA ops | Quarterly | Spreadsheet with four-fifths ratios by role and group |
| Run four-fifths rule calculation | HR analyst | Quarterly | Flag any ratio below 0.80 to HR director + legal |
| Review ATS configuration for undocumented changes | HR systems admin | Quarterly | Written confirmation: “No changes since last review” or change log |
| Request bias analysis from vendor | HR director or procurement | Annually (or after any vendor model update notification) | Vendor’s adverse impact report, logged and filed |
| Review recruiter override log | HR director | Quarterly | Count of overrides by recruiter and role; flag any role where overrides are zero |
| Legal sign-off on monitoring results | Employment counsel | Annually | Written sign-off that monitoring program meets EEOC standards |
If any of these tasks has no owner, it is not happening. Assign names, not job titles.
What discovery will force into the open
If this case reaches discovery, Workday will need to produce three categories of documentation that most AI vendors have not been required to make public. The answers will define the liability landscape for the industry.
Training data composition. Workday’s screening models are almost certainly trained in part on historical hiring data contributed by Workday customers. If that data reflects prior discriminatory hiring patterns (and the EEOC’s enforcement history suggests this is common), the model encoded those patterns at training time. The bias may have been imported from the companies’ own historical decisions and then applied back to them.
Model update logs and internal testing records. Discovery will establish whether Workday ran adverse impact testing on its models before deployment, and what the results showed. If internal testing identified disparate impact and the company deployed regardless, that is materially different from the bias being undetected.
Per-customer configuration audit trails. The defendant companies configured Workday’s ATS. What signals they weighted, and whether those configurations changed over time, is the kind of documentation most HR teams do not maintain. The inability to produce it is not neutral: it demonstrates the companies were running a system with material effects on candidates without any mechanism to understand what it was doing.
The concentration problem
Mobley applied to more than eighty companies. Each used Workday. Each rejected him. This is not eighty independent hiring decisions. It is one algorithm applied eighty times.
When three or four ATS platforms cover most of the enterprise hiring market, a single discriminatory pattern in one of them does not produce one adverse outcome. It produces a market-wide barrier. The candidate cannot self-select away from the biased system because they do not know which systems their prospective employers use.
The litigation addresses one plaintiff and one defendant. It does not address the structural question: what remedy is appropriate when the harm is market-wide and the information asymmetry is total? That question requires a regulatory answer, not a legal one. The FTC, EEOC, and CFPB have all indicated interest in algorithmic discrimination. The Workday case will be exhibit one in those proceedings.
Where this leaves HR directors
The court’s theory is simple: you selected the tool, you deployed it, you delegated the screening function to a vendor. You cannot delegate the legal obligation that goes with that function. The “we just used the software” defense does not survive the agency theory the court accepted.
The EU AI Act adds a parallel obligation for companies with European operations. Under Annex III, AI systems used in recruitment and employment are classified as high-risk, requiring conformity assessment, bias monitoring and human oversight before deployment. EU-based HR directors face that requirement now, regardless of how Mobley v. Workday resolves in US courts.
The companies in this case made defensible operational decisions with the information they had at the time. What they did not do was ask who was responsible for what the AI decided. Adding that question to the RFP, the contract review, and the quarterly operations review is the entire correction.
Sources: Complaint, Mobley v. Workday Inc., No. 3:23-cv-04146 (N.D. Cal., filed Aug. 10, 2023). Order denying motion to dismiss, Feb. 26, 2024. EEOC, “Uniform Guidelines on Employee Selection Procedures” (1978). EEOC, “Technical Assistance on Artificial Intelligence and the Americans with Disabilities Act” (2023). SHRM, “Workday AI Lawsuit: A Wake-Up Call for HR” (2024). EU AI Act, Annex III (2024).