New Jersey Employers Need to Map Every AI Hiring Cut

Consider a customer-support analyst opening. The employer’s applicant tracking system rejects people who answer a knockout question a certain way. A separate assessment vendor scores the remaining applicants, another service ranks them, and a generative model turns résumés and interview notes into short summaries. The recruiter sees the ranked list and summaries, then advances several names.

That is one hiring workflow, even though no single vendor can describe all of it.

The practical governance artifact is a decision-path sheet for that opening. It should let an internal reviewer start with one rejection and work backward: which systems touched the application, what each system received and returned, which threshold changed the candidate’s status, and whether an authorized person could have produced a different result.

This is an operational method, not legal advice. Employment counsel should determine which laws apply, what records the company must retain and whether a particular tool requires notice, an accommodation process or another control.

Inventory decisions, not products

A software inventory answers, “What did we buy?” An employment-decision inventory answers, “What changed an applicant’s chance of advancing?” The second question catches far more of the customer-support analyst workflow.

Start with one active requisition rather than sending a questionnaire to every department. Open the applicant tracking system, or ATS, which is the database and workflow software used to manage candidates, and follow the configured path from application submission to final disposition. Review integration settings, scoring rules, recruiter templates and status-change automations. Do not rely on a sales description that labels a feature “assistive.”

A step belongs on the sheet if its output removes a person, changes their order, controls what a reviewer sees or compresses source material into a representation the reviewer is expected to trust. That includes deterministic rules, such as rejecting an applicant who selects “no” on an availability question, even when the rule contains no machine-learning model. It also includes generated interview summaries that never assign a score but can omit a qualification or present an uncertain statement as settled fact.

For the customer-support analyst opening, the sheet might contain these columns:

| Field | What to record | |---|---| | Sequence | The step’s position in the real workflow | | Input | The application field, document, recording or earlier score received | | System and owner | The vendor or internal service, plus the team accountable for it | | Output | A rejection, score, rank, label, summary or recommendation | | Decision rule | The threshold or instruction that changes what happens next | | Human authority | The role allowed to review, override and restore a candidate | | Evidence | The logs, configuration and source material needed to replay the step |

Write down the configured rule, not a vague purpose. “Helps recruiters find strong candidates” is not enough. “Sorts applicants by assessment score, displays only the first page by default and moves scores below the configured threshold to inactive status” describes behavior that can be tested.

Find the cut points hidden between vendors

The most important entries are cut points, where an output narrows the pool or information available to the next reviewer. Some are visible status changes. Others happen in an interface.

Return to the ranked customer-support list. If the ATS technically allows the recruiter to open every application, but its default view shows only a small top-ranked group and the recruiter’s workload leaves no time to inspect the rest, the ranking has practical force. Calling the final click “human review” does not explain the decision path.

For each cut point, capture the rule that operated at the time. A current settings screenshot cannot establish what the threshold was months earlier if an administrator has since changed it. Version the configuration, record effective periods and connect each applicant event to the applicable version. If a vendor silently updates a model or prompt, the employer may be unable to reproduce an earlier output; that gap belongs in the inventory rather than being papered over with the vendor’s current documentation.

Generated summaries need their own line. Record the source documents, system instruction or template, model or service identifier when the vendor exposes it, output and downstream use. A summary can vary after a model update, and reconstructing it may require more than resubmitting the same résumé. If exact replay is unavailable, preserve the original output that the recruiter saw.

The fallback matters too. When the ranking service fails or cannot process a résumé, does the candidate enter a manual queue, receive a zero or disappear from the default view? An error-handling rule can become an employment rule, particularly when parsing failures cluster around unusual résumé formats, assistive-technology exports or missing fields.

Test whether a human override is real

Put a named role beside every cut point. “Recruiting can override” is too loose. The inventory should state who has permission, what that person can see and whether reversing the result returns the candidate to the ordinary workflow.

A meaningful reviewer needs access to the underlying material rather than only the model’s score or summary. The reviewer also needs enough context to recognize an error, a route for changing the outcome, and a log that records what changed. If recruiters may advance a low-ranked applicant but the system continues hiding that applicant from the hiring manager, the apparent override does not reverse the whole path.

Test the control with a nonproduction record or another approved method. Ask the designated reviewer to locate a screened-out application, inspect the source, restore it and confirm that downstream users can see it. Record the result. This costs staff time and can slow a high-volume workflow, but an untested override is weak evidence that people exercise independent judgment.

Do not force reviewers to invent explanations after the fact. Use reason categories tied to observable problems, with room for a short note: parsing error, inaccurate summary, accommodation review, wrong threshold or relevant experience omitted from the source extraction. Access to those notes should be limited because they may contain sensitive employment information.

Preserve the receipt without building a second risk

A final hiring decision is not enough to reconstruct the customer-support analyst rejection. Preserve the input as received, any transformed version sent to a vendor, the returned output, the active rule and the resulting status change. Attach timestamps, system identifiers and the authorized user behind an override.

Logs should be tamper-evident. One common control is a cryptographic hash, a digital fingerprint that changes when a file changes, paired with access controls and an audit trail. That can help show that a stored output was not edited later, although it does not prove the model’s output was accurate or lawful.

Retention creates a tradeoff. Keeping every résumé, recording, prompt and generated summary indefinitely increases privacy and security exposure, while deleting intermediate outputs can make a complaint impossible to investigate. Counsel and records staff should set a schedule that accounts for applicable employment rules, litigation holds, contractual deletion terms and the company’s stated privacy practices. The hiring team should not improvise that schedule inside the ATS.

Demographic information deserves separation. An employer may need protected-class data to examine selection rates and possible adverse impact, meaning a selection practice disadvantages a protected group, but ordinary recruiters generally do not need that data in their decision view. A controlled analytics environment can restrict who sees it and keep the analysis separate from day-to-day selection.

A vendor’s bias-audit report is supporting evidence, not a reconstruction of the employer’s workflow. It may cover a different customer population, job family, configuration or period. Ask what system was tested, which outcomes and groups were measured, whether sample sizes limited the conclusions, and whether the audited version matches the one on the decision-path sheet.

Separate current New Jersey enforcement from proposals

The enforceable baseline is not an AI-specific permission slip. The New Jersey Law Against Discrimination prohibits covered employment discrimination “because of” protected characteristics, and the New Jersey Division on Civil Rights stated in January 2025 that “the LAD applies to algorithmic discrimination just as it has long applied to other forms of discriminatory conduct.” The agency’s guidance explains its enforcement position; it did not create a safe harbor for employers that buy third-party software.

That distinction matters in the customer-support workflow. If a vendor’s ranking or résumé extraction contributes to unlawful discrimination, saying the employer did not build the model does not establish that the resulting employment action was lawful. Federal employment discrimination requirements may apply as well, including rules enforced by the Equal Employment Opportunity Commission.

New Jersey lawmakers have also considered AI-employment bills with duties such as notices or bias audits. A proposal is not an enacted statewide requirement, and its language can change during the legislative process. Teams should track current legislative status rather than copying a pending bill into a compliance checklist. Employers recruiting for jobs in New York City may separately encounter the city’s automated employment decision tool law, which has its own scope, audit and notice questions.

The decision-path sheet still has value where no AI-specific audit rule applies. It gives counsel, investigators and system owners a shared account of what happened, while exposing operational gaps before a rejected applicant, regulator or court asks for the same reconstruction.

Make the map change with the workflow

Assign one owner to the customer-support analyst sheet and trigger review when a vendor, model, prompt, threshold, integration or fallback changes. Annual review alone will miss a ranking feature switched on midway through a hiring cycle.

Procurement should require vendors to disclose subprocessors that handle applicant data, material system changes, available event logs, retention behavior and procedures for correcting outputs. If the vendor cannot preserve the output shown to a recruiter, record that limitation and decide whether the feature’s convenience justifies the evidentiary gap. For a generated summary that saves only a modest amount of reading, it may not.

Finally, reconcile policy with observation. Compare the sheet with a small set of completed hiring records and confirm that status histories, reviewer actions and vendor events follow the documented route. The useful endpoint is concrete: one rejected application, one chronological record and no unexplained jump from submitted to declined.

Questions people ask

Does

New Jersey currently require every employer to run an annual AI bias audit?

The Law Against Discrimination does not itself impose a general annual AI-audit filing rule on every New Jersey employer. Agency guidance explains how existing discrimination law reaches automated systems, while proposed AI-employment bills are not enforceable unless enacted; separate rules may apply to particular locations, contracts or employers.

Does a generated résumé or interview summary belong in the inventory?

Yes, when a reviewer relies on it or sees it instead of the complete source. Record what material entered the generator, the output shown, the applicable template or instruction, and any available service version, because omissions and unsupported statements can change who advances without producing a formal score.

Can a vendor’s audit replace the employer’s own decision map?

No. A vendor audit can describe the tested system and measured outcomes, but it usually cannot show the employer’s thresholds, interface defaults, later integrations or recruiter behavior. Keep the report, then connect its tested version and population to the specific system entry on the employer’s decision-path sheet.

How much evidence should an employer retain?

There is no single retention period for every artifact and employer. Counsel and records staff should set the schedule under applicable employment, privacy and litigation-hold requirements; operationally, the retained package should be enough to trace one applicant from original input through each score, summary, status change and override.