Disparate Impact

How disparate impact works in practice

Disparate impact is about outcomes, not motives. Nobody has to intend anything. A minimum typing speed, a requirement for a high school diploma, a credit check, a no-gap employment rule or an AI resume screener can all produce it if they filter one race, sex, national origin or age group out at a noticeably higher rate.

The doctrine comes from the Supreme Court's 1971 decision in Griggs v. Duke Power Co., where a diploma requirement and two aptitude tests excluded Black applicants at far higher rates and the company could not show they predicted job performance. Congress wrote the framework into Title VII with the Civil Rights Act of 1991. The Age Discrimination in Employment Act also allows disparate impact claims (Smith v. City of Jackson, 2005), although the employer's defense there is a reasonable factor other than age rather than business necessity.

A claim generally runs in three steps:

1. The plaintiff identifies a specific employment practice and shows, usually with statistics, that it causes a significant disparity for a protected group.
2. The employer must show the practice is job related for the position and consistent with business necessity.
3. Even if the employer does that, it can still lose if the plaintiff shows an equally effective, less discriminatory alternative existed and the employer refused to adopt it.

That third step is the one hiring teams forget. If a shorter, less exclusionary assessment predicts performance just as well, keeping the harsher one is a liability.

Adverse impact analysis and the four-fifths rule

Adverse impact is the measurement; disparate impact is the legal theory. Adverse impact analysis compares selection rates at each step where candidates are screened out: knockout questions, assessments, interviews, background checks and final offers.

The Uniform Guidelines on Employee Selection Procedures (1978), adopted by the EEOC, the Department of Labor and others, use the four-fifths rule as a practical first screen: a selection rate for any group that is less than 80% of the rate for the highest group is generally regarded as evidence of adverse impact. The Guidelines also say the rule is a rule of thumb. A smaller gap can still count if it is statistically significant, and a bigger gap may not count when the sample is tiny.

That is why experienced analysts run both the four-fifths check and a significance test, typically a two-standard-deviation or Fisher's exact test. In high-volume hiring the sample sizes are huge, so even a 0.85 ratio can be statistically significant. In a ten-person professional search, one hire changes everything and the ratio alone means little.

Run the analysis by step, not just on the overall funnel. A clean end-to-end ratio can hide a biased assessment that a generous interview stage happens to offset. Courts have rejected that kind of bottom-line defense since Connecticut v. Teal (1982).

Business necessity and job relatedness

Business necessity is the employer's defense, and it is only as strong as the evidence behind it. Saying a requirement feels sensible is not enough. The employer needs to show the criterion measures something the job actually requires.

In practice that means:

• A current job analysis that documents the tasks and the knowledge, skills and abilities the role needs.
• Validation evidence for each test or screen: content validity (the test samples real work, like a call simulation for an agent), criterion validity (scores correlate with later performance or retention) or construct validity.
• A cut score set with a reason. Many teams copy a vendor's default passing score, which is hard to defend if it screens out more people than the job requires.
• Records showing you looked for less exclusionary alternatives and why you chose what you did.

We would put the most effort into the steps with the largest drop-off. If 60% of applicants fail a knockout question about weekend availability, that question needs a documented operational reason because it carries the most weight.

Disparate impact vs disparate treatment

These two get mixed up constantly. Disparate treatment is intentional: a candidate is treated differently because of a protected characteristic, such as a hiring manager who passes over mothers of young children. The question is motive, and evidence usually comes from comments, inconsistent reasons and comparisons with similar candidates.

Disparate impact needs no motive at all. A perfectly consistent process applied identically to everyone can still be unlawful if its results skew and it cannot be justified. Disparate treatment is proved with stories and documents; disparate impact is proved with numbers. A single selection step can raise both claims at once, for example when a recruiter applies a rule inconsistently and the rule itself also has adverse impact.

Why it matters more in high-volume and AI-assisted hiring

Volume turns small biases into big numbers. A screen that excludes one extra candidate per hundred barely shows up in a 20-hire year, but a BPO hiring 3,000 agents from 60,000 applicants will have a statistically obvious pattern, and the class size of affected candidates makes litigation attractive.

Automated tools are the newest source of risk. Resume screeners, chatbots that apply knockout logic, game-based assessments and AI interview scoring all make selection decisions at scale. Under Title VII the employer remains responsible for the outcomes of a tool it chooses to use, even when a vendor built it. New York City's Local Law 144 requires a bias audit of automated employment decision tools before use, including impact ratios by sex and race/ethnicity, and Colorado and the EU have passed broader AI rules.

Federal enforcement priorities shifted in 2025, when an executive order directed agencies to deprioritize disparate impact enforcement. That changes what the agencies chase, not what Title VII says. Private plaintiffs can still bring claims, and many state fair employment laws and state attorneys general apply the same theory. We would not change a monitoring program because of it.

Common mistakes

• Not collecting voluntary self-identification data. Without demographic data at the applicant stage you cannot run the analysis, and you will be reconstructing it later under subpoena.
• Only checking final hires. Adverse impact usually lives at one step. Check each screen separately.
• Treating vendor validation as your validation. A vendor study done on a different job, population or cut score does not prove business necessity for your role.
• Ignoring age. Many teams check race and sex but not applicants 40 and over, which matters for screens like years-of-experience caps or graduation dates.
• Running the numbers once. Applicant pools shift with sourcing channels and seasons. Quarterly monitoring is a reasonable cadence for high-volume roles.
• Fixing the ratio instead of the tool. Adjusting scores or setting different cut scores by group (race norming) is itself prohibited by the Civil Rights Act of 1991.

How to reduce disparate impact risk

Start with the steps that remove the most people. Ask whether each requirement is truly needed on day one or can be trained. A contact center that drops a two-year customer service experience requirement and replaces it with a 15-minute call simulation often gets a wider, more diverse pool and better first-90-day retention.

Then build a routine: collect self-ID data, compute impact ratios and significance at each step every quarter, document the job relatedness of every screen, and keep a log of alternatives you considered. When a step shows adverse impact, test less exclusionary options such as a lower but still valid cut score, a different assessment format or removing a requirement altogether.

This is general information about US federal law. State laws and local AI hiring rules add their own requirements, so involve employment counsel before redesigning a selection process.