New Jersey Employers Need a Map of Every Hiring AI Tool

Start with one applicant’s path through a real requisition. A sourcing platform finds the person, an applicant tracking system stores the application, a screening module assigns a match score, and an interview service produces a transcript and summary for the hiring manager.

Those products may appear under one software contract. They do not perform the same function, create the same evidence or raise the same compliance questions.

That distinction matters in New Jersey because the state’s Law Against Discrimination, or LAD, already prohibits discriminatory employment practices, including those produced through automated tools. New Jersey’s Division on Civil Rights has made clear that using an algorithm does not displace the employer’s existing obligations. More specific state requirements have also been proposed, but a proposal is not an enforceable rule unless it is enacted and takes effect.

Employers recruiting in New York City face another layer. The city’s Local Law 144 regulates certain automated employment decision tools, or AEDTs, used in hiring and promotion. Its definition centers on a computational process that produces a score, classification or recommendation and is used to “substantially assist or replace discretionary decision making.” Covered use requires a recent independent bias audit, publication of specified audit information and advance notice to affected candidates or employees.

A procurement list cannot answer whether that definition applies. A functional inventory can.

Trace the decision, not the product category

Return to the applicant moving through the requisition. The useful question is not whether each vendor advertises “AI.” Record what happens immediately before and after every system touches the application.

The sourcing platform searches public profiles and recommends prospects. Its output affects who enters the funnel, even though it never rejects an application. The screening module compares resume content with job criteria and gives the recruiter a ranked queue. That output directly changes review order and may determine which applicants receive attention before the position closes.

The interview service is more ambiguous. A transcript that reproduces spoken words differs from a summary that selects themes, omits details and labels an answer as weak or incomplete. If the service scores communication, predicts traits or recommends advancement, it has crossed further into evaluation. Accuracy also changes by audio quality, accent, specialized vocabulary and overlapping speech, so the employer needs a way to recover the recording or transcript and correct the record.

The applicant tracking system may only store names, resumes, dispositions and interview notes. Storage alone is not candidate evaluation. Yet the same platform may include optional ranking, knockout questions, recommendation models or generated candidate summaries, and a module enabled by an administrator can change the classification without changing the vendor name on the invoice.

Capture modules separately. One row labeled with a large software suite conceals the setting that matters.

Build each row around an employment action

For every module, name the employment stage, the input, the output and the action a person takes with it. Include the system owner and the person who can alter its configuration, since responsibility often sits outside recruiting.

A useful row for the screening module might read: resumes and application answers enter; a match score and ranked queue leave; recruiters normally open the first group shown; no written rule requires review of lower-ranked applicants; the vendor hosts the model; talent operations controls weighting and knockout settings. That description exposes more than a checkbox marked “AI used.”

Add the population and geography. Record whether the tool touches applicants for New Jersey jobs, New Jersey residents, New York City positions or remote roles whose location may change during recruitment. Jurisdiction should be reviewed with qualified counsel because coverage can turn on facts that a software administrator cannot resolve.

Then record the fallback. If the ranking module is disabled tomorrow, can recruiters review applications in submission order, search by job-related criteria or use a documented random batch? A manual fallback costs staff time and may slow hiring, but a system that cannot be paused leaves the employer dependent on a vendor’s legal interpretation and release schedule.

The final fields should identify available evidence: model or rule documentation, configuration history, output logs, validation material, accessibility information, retention periods and prior bias testing. Do not write “vendor handles compliance.” Record the document, its date, its scope and whether it covers the version and customer configuration in use.

Classify by influence on the decision

The first category is decision output. Scores, rankings, classifications, recommended dispositions and generated shortlists belong here when they guide screening, advancement, selection or promotion. These systems deserve the earliest review because they may fit an AEDT definition, create measurable selection effects and expose the employer to discrimination claims under existing law.

The second category is decision support. An interview summary, generated job-fit narrative or recruiter copilot may not issue a numeric score, but it can still shape a decision if managers rely on it instead of reviewing the underlying material. The inventory should record actual use rather than the vendor’s label. A summary becomes consequential when managers treat omitted information as absent or copy its conclusions into a rejection record.

Sourcing needs its own classification. A recommendation engine can determine who hears about a job and who remains invisible to recruiters, while traditional selection-rate analysis may begin only after someone applies. Ask the vendor how it defines the eligible pool, which signals influence recommendations and whether recruiters can inspect candidates outside the recommended set.

Records-only tools sit at the other end. File storage, scheduling and transmission generally do not rank or recommend people, although they still carry security, retention, accessibility and privacy obligations. Keep them in the inventory. Mark them as non-evaluative only after confirming that scoring and summarization features are disabled.

Classification is not a legal safe harbor. It gives counsel, compliance staff and procurement a factual record from which to evaluate a rule’s definition.

Test the use that happens in practice

Written policy often says a recruiter makes the final decision. That fact alone says little about how much the tool assists.

Observe several ordinary reviews or inspect system logs where lawful and available. If recruiters see only the top-ranked applicants by default, rarely move beyond the first page and must override a warning to advance someone with a low score, the tool has more influence than a policy describing it as optional. Conversely, a search aid that returns documents in response to recruiter-selected terms may have less decision authority, though the terms themselves still need job-related justification.

Record thresholds and defaults. A model may rank everyone, while a customer-configured cutoff converts that ranking into an automatic rejection. The vendor built the score; the employer chose the operational consequence. Both facts belong in the row.

Check generated interview summaries against the source recording or transcript. The goal is not to declare the model accurate after a few examples. It is to identify predictable failure points, establish who may correct an error and prevent the summary from becoming the only retained account of the interview.

Attach controls after classification

Once the inventory reflects actual use, map controls to each row. For a decision-output tool, preserve the audit or assessment supplied by the vendor, then determine whether it matches the legal standard that applies to the employer’s use. New York City’s bias-audit requirement, for example, cannot be satisfied by a generic fairness brochure that omits the required calculations, covered categories or audit date.

For a decision-support summary, require access to the underlying material, prohibit unsupported trait inference and tell reviewers which source controls when the summary conflicts with the transcript. For sourcing, retain enough information to understand the candidate pool and recommendation logic. For records-only modules, monitor settings so an update does not quietly activate evaluation features.

Vendor questions should follow the row. Ask which inputs affect the output, whether customer activity retrains a shared model, which subcontractors receive applicant data, how versions are logged, and whether the employer can export individual outputs and configuration history. A refusal to provide model internals does not end the inquiry; it changes the procurement decision, because the employer must decide whether the available evidence supports continued use.

Set a review trigger rather than relying only on an annual calendar. Reclassify a tool when a new module is enabled, a model or threshold changes, the workflow expands to promotion, or a recruiter begins using an output for a purpose absent from the original assessment. Return to the same applicant path after each change. If the path now includes a recommendation that was previously just storage, the inventory should show it before a regulator, candidate or auditor does.

This is operational guidance, not legal advice. New Jersey employers should have counsel verify current state and local requirements, particularly where proposed legislation, remote hiring and multi-jurisdiction recruiting overlap.

Questions people ask

Does

New Jersey already ban discriminatory hiring AI?

New Jersey’s LAD already prohibits unlawful employment discrimination, and using an automated system does not remove that obligation. The state’s civil-rights guidance addresses algorithmic discrimination under existing law, while AI-specific bills or requirements must be checked for their current legislative and effective status.

Is an interview transcription tool an automated decision tool?

A transcription service that only reproduces speech is different from a system that scores answers, infers traits or recommends advancement. A generated summary may still influence a decision, so record its output, whether reviewers consult the source material and how errors can be corrected before assigning a legal classification.

Does a vendor’s bias audit cover the employer?

Not necessarily. An employer must determine whether the audit covers the relevant product version, configuration, job population and legally required categories, and whether the auditor meets any applicable independence standard. A vendor document may be useful evidence without satisfying a particular jurisdiction’s requirement.

Should records-only software stay in the inventory?

Yes. Mark it as non-evaluative only after confirming that ranking, filtering, scoring and summarization functions are absent or disabled. Keeping the row makes later changes visible, especially when a routine platform update adds an AI feature or an administrator enables a module without a new procurement review.