New Jersey Employers Need a Map of Every AI Hiring Filter

Take a warehouse-supervisor requisition. An applicant uploads a résumé to an applicant tracking system, which ranks the application against the job description. A video platform scores recorded answers. A scheduling service offers interview slots, but only after checking availability rules.

An online assessment produces a percentile score and sends candidates below a configured cutoff to the rejection queue.

Procurement may describe those as four separate products. The candidate experiences one selection pipeline.

That distinction matters in New Jersey because the state’s Law Against Discrimination, or LAD, already covers employment decisions made with software. Guidance published by the New Jersey Division on Civil Rights in January 2025 explains that automated decision-making tools can produce unlawful discrimination through their design, training data, deployment, or interaction with people. Buying a tool from a vendor does not remove the employer’s exposure under the law.

The inventory is the missing control. It does not decide whether a system is lawful, and the LAD does not expressly require employers to maintain an AI inventory. It gives legal, recruiting, procurement, accessibility, and security teams a shared record of where automated outputs enter a hiring decision, which is necessary before anyone can sensibly assign a notice, request a bias review, or preserve evidence.

Begin with a requisition, not a vendor list

A vendor list starts in the wrong place. It captures contracts but misses features embedded in broader recruiting suites, administrator-enabled recommendations, free assessment plug-ins, and tools purchased by a staffing agency rather than the employer.

Instead, follow the warehouse-supervisor requisition from publication to disposition. Record every point where software sorts people, recommends an action, changes what a recruiter sees, or determines whether the applicant reaches the next stage. “Disposition” is the status explaining what happened to an application, such as withdrawn, rejected after assessment, or selected for interview.

At the résumé stage, inspect the recruiter’s screen rather than the product’s marketing page. A conventional applicant tracking system may store applications without evaluating them. The same system may also parse work history, infer skills, rank applicants, recommend candidates from an existing database, or suppress applications that fail knockout rules. Those functions belong in the inventory even if the vendor avoids the term artificial intelligence.

Next, inspect the interview workflow. A platform that records and transmits video is different from one that transcribes answers, compares language with a scoring rubric, evaluates speech patterns, or creates a recommendation. Facial, vocal, and language analysis deserves explicit attention because camera access, speech disabilities, accents, lighting, and bandwidth can affect the input before the scoring model evaluates anything.

Scheduling looks administrative until it changes access. A calendar tool that displays mutually available times may have no selection role. A system that prioritizes candidates, withholds slots based on availability, treats delayed responses as withdrawal, or routes only certain applicants to a recruiter has moved into the decision path. Put the rule in the warehouse-supervisor record, not merely the product name.

Finally, trace the assessment. Note whether it uses fixed questions, adapts later questions to earlier answers, checks response time, detects suspected cheating, or combines several measures into one score. Then identify the configured cutoff and the fallback. If a candidate requests an accommodation or the platform fails, the record should show whether a person can administer an alternative assessment and who can authorize it.

Record the output and what it controls

“Uses AI” is not a useful inventory field. The record needs enough detail to reconstruct the hiring effect without requiring a recruiter to understand the model’s code.

For each system, capture its owner, vendor, product and module, covered jobs and locations, input data, generated output, configured threshold, downstream action, human review point, accommodation route, deployment date, last material change, and evidence location. A material change is one that can alter an output or how the organization uses it, such as enabling résumé recommendations, replacing an assessment, or moving a cutoff.

Write the central fields as a causal statement: “The assessment converts candidate responses into a score; scores below the employer-set threshold enter a rejection queue; a recruiter does not review them unless an accommodation or technical-failure flag is present.” That sentence reveals more than an “AI assessment” label. It identifies who set the consequential rule and whether human review is routine or exceptional.

The warehouse-supervisor row may also expose duplicated filtering. If résumé ranking favors prior supervisory titles and the assessment separately rewards experience with management scenarios, the pipeline could weight similar experience twice. A standalone vendor report may not show that interaction because each supplier sees only its own stage.

Do not collect sensitive demographic information merely to make the spreadsheet look complete. Bias testing may require protected-class data or a defensible proxy methodology, but access, purpose, retention, and separation from selection decisions need their own controls. The inventory can point to an approved analysis environment rather than copying candidate-level data into a general governance file.

Separate enforced duties from proposed controls

The current baseline is the LAD. Its employment provision makes it unlawful, because of protected status, “to refuse to hire or employ or to bar or to discharge” a person, or “to discriminate against such individual in compensation or in terms, conditions or privileges of employment.” An automated ranking, cutoff, or recommendation does not sit outside that rule merely because no person intended a discriminatory result.

New Jersey proposals have also targeted automated employment decision tools more directly. A3854, introduced in 2024, proposed controls involving bias audits and candidate notices. Proposed legislation can change, stall, or be replaced, so employers should verify legislative status rather than entering a bill requirement in the inventory as though it were currently enforced.

Other jurisdictions may attach enforceable controls to the same workflow. New York City’s Local Law 144, for example, restricts use of a covered automated employment decision tool unless it “has been the subject of a bias audit conducted no more than one year prior to the use of such tool,” with a summary made publicly available. It also carries notice obligations. A New Jersey employer recruiting for a New York City position may therefore need a different control mapping from the one used for the New Jersey warehouse role.

This is why location belongs at the requisition level. A companywide product label cannot establish which rule applies to which candidate, job, or stage.

Ask vendors for evidence tied to the configured use

A generic assurance that a product is unbiased is not evidence. Send the vendor the inventory statement describing the employer’s use, then request material that matches it: validation studies for the relevant job and output, subgroup results and methodology, accessibility documentation, known input limitations, change logs, data-retention terms, and records showing when scores or recommendations were generated.

Ask who controls the threshold. If the employer chooses it, a vendor’s test of the underlying score may not evaluate the employer’s rejection rule. If the vendor silently adjusts it, the employer needs change notification and a way to identify affected applications. The same issue applies when a résumé model changes its job-matching logic while the recruiting interface and contract name remain unchanged.

Bias audits also have boundaries. A selection-rate comparison can reveal disparities among groups included in the analysis, but it does not establish accessibility, validate a score for a particular job, detect every form of discrimination, or prove that recruiters use the output as documented. Record the population, period, version, categories, exclusions, and configured workflow covered by each report. An audit that cannot be matched to the deployed warehouse-supervisor pipeline should not close that inventory item.

Where a supplier will not provide sufficient evidence, the options are operational rather than rhetorical. The employer can narrow the tool to a non-decisional function, add documented review before rejection, replace the feature, or accept a gap for further evaluation. Calling the system “decision support” does not resolve the issue if recruiters routinely follow its ranking.

Keep receipts when the workflow changes

Assign each inventory row an operational owner and a review trigger. Contract renewal is too late. A new module, altered cutoff, revised job family, vendor model update, or change from recommendation to automatic rejection should reopen the record.

Preserve the job configuration, notice version where one applies, vendor evidence, assessment rubric, override record, accommodation path, and dates of material changes under the organization’s approved retention schedule. The goal is not to log every mouse click. It is to retain enough evidence to show which system produced an output, what that output controlled, and which version governed the candidate at that time.

Return once more to the warehouse-supervisor requisition. If the spreadsheet cannot show why an applicant disappeared after the assessment, who set the cutoff, and whether a failed video interview had a human fallback, the employer has found a governance gap before debating the quality of the model.

Questions people ask

Does every recruiting tool belong in an AI inventory?

Include a tool when it sorts candidates, generates a score or recommendation, controls access to a stage, or supplies information people routinely use in selection. A system that only stores résumés or displays open calendar slots can be recorded as non-decisional, with that boundary documented, rather than subjected automatically to every AI control.

Is a human reviewer enough to remove an automated tool from scope?

Usually, the presence of a reviewer does not answer the operational question. Record what the reviewer sees, whether rejected candidates ever reach that person, how often recommendations are overridden, and whether the interface anchors the reviewer to a score. Applicable legal definitions and duties still need to be assessed for the particular location and use.

What should an employer request from an assessment vendor?

Request evidence for the deployed assessment and job context, including validation methods, subgroup analyses, accessibility testing, known limitations, version history, retention practices, and the party controlling cutoffs. A broad fairness statement or audit for another configuration cannot show how the employer’s threshold affects applicants in its own workflow.

Does

New Jersey currently require an annual AI hiring audit?

The LAD currently prohibits employment discrimination and applies when automated tools contribute to a decision, but it does not itself create a general annual AI-audit mandate. New Jersey proposals have contemplated more specific controls, while rules such as New York City’s may apply to covered roles. Legislative status and geographic scope should be checked separately.