Before AI Rejects a New Jersey Applicant, Build This File

New Jersey’s Law Against Discrimination already applies when software contributes to an unlawful employment decision. State civil-rights guidance has made the enforcement position explicit: using an algorithm or relying on a vendor does not remove the employer’s responsibility for discriminatory results. Proposed rules for automated employment decision tools may add duties later, but proposals are not current requirements.

Consider one concrete workflow. A retailer wants a vendor’s screening system to rank applicants for customer-support supervisor jobs. The system parses each résumé, scores a timed online assessment, combines those results into a number from zero to 100, and sends only applicants above a cutoff to a recruiter.

Before that score touches a live application, the employer should be able to open one folder and show why every input belongs there, what happens when the test is inaccessible, who can inspect or reuse the data, and what the applicant was told. Call it the deployment evidence file. It is not a substitute for legal advice; it is a structured set of questions for employment, privacy and procurement counsel.

Freeze the workflow before reviewing the claims

Start with a one-page decision map, not the vendor’s description of its product. Record each step from application to disposition: what data enters, what the model infers, which rule applies the cutoff, what the recruiter sees, and whether a person can reverse the result.

For the supervisor score, the map should distinguish résumé parsing from assessment scoring. A parser extracts fields such as prior titles or dates, while a scoring model converts responses or behavior into a prediction. Those components can fail differently. A parser may miss experience because a résumé uses an unusual layout; a timed test may penalize an applicant who uses assistive technology; a ranking model may reproduce patterns in historical hiring data even when protected traits are absent from its inputs.

Save the model name or product release identifier supplied by the vendor, the configured cutoff, enabled features, weights if available, integration settings and deployment date. A later audit is weak if nobody can reconstruct which configuration produced the disputed score.

The file should also identify the fallback. If the service is unavailable or an applicant contests a result, does a recruiter review the underlying application, administer a different assessment, or merely rerun the same model? Repeating the same automated step is not independent review.

Document why the score belongs in this job

Federal disability law restricts selection criteria that screen out a person with a disability unless they are “job-related for the position in question and consistent with business necessity.” Federal employment guidance also expects employers to examine whether a selection procedure causes adverse impact, meaning a materially lower selection rate for a protected group. New Jersey’s Law Against Discrimination covers protected characteristics under state law and applies regardless of whether the employer built the system or bought it.

For the customer-support supervisor role, begin with a current job analysis. Identify the tasks performed, conditions under which they occur, and competencies needed at entry. Then connect each scored feature to one of those requirements. If the assessment measures response speed, the evidence file should explain why speed under that particular time limit predicts supervisory performance rather than familiarity with online tests or the ability to use a standard keyboard.

Vendor statements that a model is “validated” are not enough. Ask what outcome it predicts, which population supplied the validation data, whether the studied role resembles the New Jersey position, and whether the configured version matches the evaluated version. A model validated for high-volume call-center agents may not support decisions about supervisors whose work centers on coaching, scheduling and escalation judgment.

Record alternatives considered. A structured interview, work sample or recruiter review may take more staff time, but it can be easier to connect to the job and easier to modify for an accommodation. If the automated score saves minutes at the top of the funnel while creating an opaque rejection that nobody can explain, the lower processing cost may not justify the governance burden.

Keep outcome data by stage. The federal Uniform Guidelines say employers should “maintain and have available for inspection records or other information which will disclose the impact” of selection procedures. The commonly used four-fifths comparison, where one group’s selection rate is less than 80 percent of another’s, is a screening signal rather than a safe harbor; sample size, job grouping and the relevant legal theory still matter.

Build an accommodation route that changes the test

A line reading “contact us if you need help” does not establish an accommodation process. The deployment file needs an owner, a monitored contact channel, a response procedure and a way to pause automated rejection while the request is reviewed.

Return to the timed assessment. An applicant using a screen reader may need more time, a compatible interface or a different format. The employer should test whether the vendor’s timer pauses, whether keyboard navigation works, and whether an accommodation flag is kept away from the people making the hiring decision. The alternative assessment must measure the same job requirement without reproducing the original barrier.

Document who can authorize the change and how the resulting score enters the workflow. If extra time automatically lowers a “speed” feature, the accommodation is cosmetic. If the vendor cannot suppress that feature, the practical fallback may be a recruiter-administered work sample, which costs staff time but produces a decision the employer can inspect.

Accessibility testing and bias testing answer different questions. An interface can conform to accessibility standards while its scoring logic still disadvantages people with disabilities. Conversely, group-level selection rates may look stable while an individual applicant cannot operate the test. Preserve evidence from both reviews.

Pin down the vendor’s role in writing

Procurement should ask what the vendor does with applicant information after producing the supervisor score. The contract and technical documentation should identify data fields collected, subprocessors, storage locations, retention periods, security controls, deletion mechanics and any use for product improvement or model training.

Legal labels need factual support. A vendor may call itself a processor, but that label does not answer whether it chooses new purposes for the data. A screening provider that assembles information about applicants for employment decisions may also raise questions under the federal Fair Credit Reporting Act, depending on what information it supplies and how it operates. If the product generates a consumer report, employers can face authorization, disclosure and pre-adverse-action duties that are more specific than a general AI notice.

New Jersey’s comprehensive privacy law took effect in 2025, but its definition of consumer excludes a person acting in an employment context. Employers should not assume that every applicant interaction therefore falls outside every privacy obligation, especially if data is reused for an unrelated purpose, combined with consumer profiles, or collected from other sources. Counsel should map the actual data flow rather than paste a consumer privacy policy onto the careers page.

Require usable records from the vendor: input and output logs, model or configuration identifiers, reasons or feature information available for a score, override history, access logs, deletion confirmation and incident notification. Some vendors reserve detailed audit exports for a higher service tier. That price belongs in the deployment decision because a cheap screening tool that cannot produce evidence may become expensive during an investigation.

The contract should also address change control. If the vendor can alter weights, assessment content or model behavior without notice, the employer cannot rely on an earlier validation review. Set a trigger for reassessment when a material component changes.

Separate required notices from useful notice

New Jersey does not currently supply one blanket, generally applicable AI-hiring notice that resolves every workflow. State proposals concerning automated employment decision tools could change that position, while other laws may already require particular disclosures based on the data, vendor and location involved. A bill under consideration is not an enforceable requirement.

The evidence file should contain a jurisdiction matrix for the job location, applicant location and employer operations. A New Jersey company recruiting in New York City may encounter the city’s automated employment decision tool rules, including bias-audit and notice provisions. Video interviews, biometric features, background information and applicants in other states can trigger separate analyses.

Even where counsel finds no AI-specific mandate, a short operational notice can reduce confusion. It should state that an automated system evaluates specified application or assessment information, explain how the result affects review, identify the accommodation route, and provide a contact for correction or reconsideration. Do not promise “human review” unless a trained person can access the underlying material and change the outcome.

Keep the exact notice version, where it appeared, when it was delivered and any acknowledgment. For the supervisor workflow, the notice should arrive before the timed assessment, not inside a rejection email after the score has already closed the application.

Put a stop condition in the file

Approval should expire. Set review triggers for a vendor update, a changed cutoff, a new data source, a different job family, a material selection-rate disparity, repeated accommodation failures or an unexplained rise in recruiter overrides.

Name the person authorized to pause the system and preserve the logs. Without that control, an employer can accumulate more contested rejections while procurement, human resources and the vendor debate ownership. The first response to a credible failure in the supervisor score should be to route applications through the documented fallback, not to erase the evidence by resetting the integration.

Questions people ask

Does

New Jersey require a bias audit before an employer uses AI in hiring?

New Jersey’s discrimination law applies to automated decisions, but that does not create one universal bias-audit format for every employer and tool. Other jurisdictions may impose specific audit rules, and state proposals may change the requirements. Counsel should determine which rules cover the job, applicant and workflow before deployment.

Can an employer rely on the vendor’s validation report?

A vendor report is supporting evidence, not the complete file. The employer still needs to determine whether the evaluated model, population and job resemble its configured use, then monitor its own selection results. A report for another occupation or an earlier product configuration may not establish job relevance for the deployed score.

What should happen when an applicant requests an accommodation?

The employer should pause automated rejection, route the request to a responsible person, and offer a modification or alternative that measures the same job requirement without the barrier. Running the applicant through the unchanged tool again does not provide a meaningful accommodation or independent review.

Should rejected candidates be told that AI affected the decision?

The answer depends on applicable employment, consumer-reporting, local and state rules. Even when no AI-specific notice is required, employers should consider a clear pre-assessment notice describing the tool’s role, the information evaluated, the accommodation path and the contact for correction, then preserve proof that the applicant received it.