AI screening tools in public sector hiring: what to check first

Reference article · 5 min read

Applicant tracking systems have filtered and ranked candidates for a long time. What is newer is that the ranking is often produced by a model trained on past hiring data, and that assessment and video platforms score how a candidate writes or speaks. These tools sit inside a hiring process governed by civil service rules, employment law, and public records duties written for human examiners.

The first question is not whether the tool is accurate. It is whether the agency knows where scoring happens, who approved it as a selection procedure, what candidates are told, and what record survives. Requirements differ by state and locality and are changing, so confirm current rules with counsel.

Find the scoring before you evaluate it

Ask HR to walk through one requisition end to end and list every point where an applicant is filtered, scored, ordered, or removed: knockout questions, resume parsing with a match score, ranked candidate lists, scored assessments, analysis of recorded interview responses. Scoring features often arrive inside an existing applicant tracking system in a routine release, so the agency never made a purchase decision about them.

Civil service and merit rules come first

In most merit systems, an examination is anything used to rate and rank applicants for a classified position. Rules commonly require that examinations be job-related, uniformly administered, approved by the personnel director or civil service commission, and used to produce an eligible list in a defined order, with appeal rights over results.

An automated screen that eliminates or reorders applicants is part of that examination. Check three things:

Approval. Whether the tool and its configuration were approved the way any other examination instrument would be, and whether that approval covers later vendor changes to scoring.
Preference points. Where veterans preference and other statutory preferences apply relative to algorithmic ranking. A model that reorders a list after preferences are added can quietly undo them.
Labor agreements. Whether a collective bargaining agreement covers selection procedures, promotional examinations, or new evaluation methods.

Notice to candidates and accommodation requests

Tell candidates at or before the point of application that an automated tool is used, what it evaluates, what information about them it uses, and how to request an alternative. Several jurisdictions now require some version of this. New York City requires notice and an independent bias audit for automated employment decision tools, Illinois regulates artificial intelligence analysis of video interviews, Maryland restricts facial recognition in interviews without consent, and Colorado has enacted broader requirements covering consequential decisions including employment. Coverage and effective dates vary and have shifted, so confirm the current text with counsel.

Accommodation needs a working route, not a policy statement. Publish how to request an alternative evaluation method, train recruiters to grant one without requiring a diagnosis, and make sure the request does not delay or disadvantage the candidate. Timed tests, game-based assessments, and anything scoring speech or facial movement deserve particular care, since they can penalize disability and speech differences without measuring job ability.

Validation, adverse impact, and independent audits

Federal employment law reaches the tool exactly as it reaches a paper test. Under the Uniform Guidelines on Employee Selection Procedures, a selection procedure with adverse impact must be job-related and consistent with business necessity, and the guidelines treat a selection rate below four-fifths of the highest group's rate as evidence of adverse impact. Title VII, the Americans with Disabilities Act, and the Age Discrimination in Employment Act all apply, and federal guidance on automated tools has been revised, so check what is current.

Vendor validation studies are usually generic. Ask for evidence tied to your job family and an applicant population resembling yours, and analyze selection rates by group at each stage, with and without the tool. Where an independent audit is required, confirm the auditor is not the developer or otherwise financially interested.

Records the agency has to be able to produce

Scores are employment records. They fall under your retention schedule, can be requested under public records law subject to applicable exemptions, and are evidence in an appeal or a discrimination complaint. Federal equal employment rules require keeping application and selection records for a period after the action, and state schedules are often longer.

Retain enough to reconstruct a decision: inputs used, score, thresholds in force, tool and configuration version, who reviewed the result, and any override. If those records live only in a vendor dashboard, the agency may be unable to answer a records request or defend a ranking at a hearing.

What to require from the vendor

Disclosure of every scoring function, including features added in later releases, with agency approval before material changes to scoring logic.
Inputs and features in writing, with a prohibition on proxies for protected characteristics and on inferring them.
Validation and adverse impact evidence, refreshed periodically, plus cooperation with agency testing or an independent audit.
Per-candidate record export covering inputs, scores, versions, and overrides, during the term and at exit.
Data terms covering retention, deletion, subprocessors, and no use of agency applicant data to train models for other customers.
Cooperation with public records requests, civil service appeals, audits, and litigation holds.

How to start

Turn off automatic rejection first, since it carries the most risk and is the easiest setting to change. Run the tool alongside human screening and compare outcomes by group before relying on it. Then get the examination plan approved and write the candidate notice and accommodation route in plain language.

HQT works with state and local agencies on system integration, data handling, and the logging that makes automated steps reviewable.

Frequently asked questions

Do civil service rules apply to an AI resume screening tool?

In most merit systems, an examination is anything used to rate and rank applicants for a classified position, so a tool that filters or reorders candidates is usually part of the examination. That generally means the same approval path, job-relatedness expectations, and uniform administration required of any other instrument, and candidates keep their appeal rights over results. Rules differ by jurisdiction, so confirm with your personnel department and counsel before deployment.

Do we have to tell applicants that AI is used in screening?

Several jurisdictions require notice in some form, including New York City for automated employment decision tools and Illinois for artificial intelligence analysis of video interviews, with coverage and effective dates that vary and have changed. Even where no statute applies, notice is what makes an accommodation request possible and supports the merit record. Useful notice states that an automated tool is used, what it evaluates, what information about the candidate it uses, and how to request an alternative.

What is a bias audit of a hiring tool, and do we need one?

It is a review of whether the tool produces different selection rates across groups, performed by a party that did not develop the tool and does not have a financial interest in it, often with a published summary of results. Some jurisdictions require one for automated employment decision tools; elsewhere agencies commission them voluntarily. Regardless of any mandate, compare your own selection rates by stage with and without the tool, because vendor validation studies are usually generic rather than tied to your jobs and applicant pool.

How do we handle an accommodation request when the assessment is automated?

Treat it as the ordinary interactive process under disability law: publish how to request an alternative evaluation method before the assessment begins, let recruiters grant one without requiring a diagnosis, and make sure the request does not delay or disadvantage the candidate. Timed tests, game-based assessments, and anything scoring speech, video, or facial movement are the most common problem areas. Define the alternative method in advance rather than improvising one when a request arrives.

How long do we keep the scores an AI screening tool produces?

Scores are employment records. They fall under your agency's retention schedule and under federal equal employment recordkeeping rules that require keeping application and selection records for a period after the action, and state schedules are often longer. Retain enough to reconstruct a decision, including inputs, thresholds in force, tool and configuration version, reviewer, and overrides, and confirm you can export all of it from the vendor system.