AI Can Help Recruit. It Can’t Own the Hiring Decision.

AI can help a staffing firm move faster on sourcing, resumes, and candidate communication, but it doesn’t remove the employment-law responsibilities that come with deciding who gets hired. Speed only helps if a person can still explain the criteria, review the result, and own the decision.

A recruiter receives 180 applications for a client opening before lunch.

An AI tool can summarize them, identify common qualifications, draft candidate emails, and perhaps rank likely matches in minutes. That can be genuinely useful for a 25-person staffing firm where recruiters are balancing candidate conversations, client requests, interviews, documentation, and business development.

The problem begins when “assist the recruiter” quietly becomes “make the decision.”

Employment law still applies when software is involved

Federal anti-discrimination laws do not create an exception for decisions made with algorithms or AI.

In 2022 and 2023 the Equal Employment Opportunity Commission published technical assistance explaining how the Americans with Disabilities Act and Title VII apply to AI-assisted employment decisions. Those documents were removed from the agency’s website in January 2025, following Executive Order 14179 and a change in federal AI policy. The Department of Labor’s October 2024 workplace AI best-practices materials were pulled at the same time.

The underlying laws did not change. Title VII, the ADA, and the ADEA still apply to decisions made with algorithms. What changed is that an employer looking for a current federal playbook will not find one. That vacuum is being filled by state law, which is covered below.

The ADA remains especially relevant when a tool evaluates applicants in ways that may screen out a person with a disability or creates barriers to requesting a reasonable accommodation. Title VII and other federal employment laws can also be implicated when selection systems intentionally discriminate or create unlawful adverse effects on protected groups.

Agency guidance was never the source of the obligation. The legal obligation comes from the underlying employment laws, and it did not move. The practical lesson is straightforward: a staffing firm cannot outsource accountability to the vendor that built the tool.

Separate low-consequence assistance from consequential decisions

Not every recruiting use needs the same level of control.

AI is relatively easy to supervise when it is used to:

1. rewrite a job description for clarity
2. summarize recruiter notes
3. draft a scheduling email
4. generate interview-question ideas
5. organize a candidate pipeline
6. create a first draft of a client update

The stakes rise when AI is used to:

1. reject applicants
2. score resumes
3. recommend who advances
4. infer personality or suitability
5. evaluate video interviews
6. match candidates to jobs using opaque criteria
7. prioritize some candidates while effectively hiding others

That second group deserves a much more deliberate review process because the tool is influencing access to employment.

Know what the system is evaluating

A staffing manager should be able to ask a simple question: Why did this candidate move forward and that candidate not?

If the answer is “the system scored them that way,” management does not yet have enough information.

Most of that information comes from the vendor, and it should be requested before purchase rather than after a complaint. Five questions are enough to start:

1. What criteria does the tool evaluate, and are they job-related?
2. Has it been tested for adverse impact, and can we see the results?
3. Can we get the decision rationale for an individual candidate?
4. What are our contractual protections if the tool produces a discriminatory outcome?
5. What happens to candidate data submitted to the system, and how long is it retained?

Every state AI employment framework enacted so far places responsibility on the deployer, not only on the developer. The firm using the tool is the firm on the hook.

A recruiter does not need to understand machine-learning mathematics. The firm does need to understand the business criteria being used. Are candidates being evaluated on required skills, years of experience, schedule availability, certifications, geography, employment history, test results, keywords, or something else? Are those criteria job-related? Are proxies creeping into the process that could create unfair outcomes?

The American Staffing Association’s April 2026 guidance, A Practical Path to AI Adoption, frames this as a “crawl, walk, run” progression, built on a principle worth quoting: as AI moves closer to decisions that affect people’s careers, governance must expand proportionally. ASA also warns against all-or-nothing governance, which is the trap most small firms fall into. That is industry guidance, not law, but it points to a sensible operating principle: a firm should understand enough about a tool to supervise its use.

Human review has to mean more than clicking “approve”

A nominal human-in-the-loop process is not much protection if recruiters routinely accept the system’s recommendation without review.

“Meaningful human review” is also the phrase state AI employment laws use, so this is not only good practice. It is the standard those laws are written against. Meaningful human review means the recruiter has enough information, authority, and time to challenge the output. That may include checking whether a resume was parsed correctly, reviewing candidates who fell just below a cutoff, verifying that required accommodations were considered, and documenting why the recruiter agreed or disagreed with the AI-assisted recommendation.

For a staffing firm, this is also a client-service issue. If a client asks why one candidate advanced and another did not, the recruiter should be able to point to job-related criteria and meaningful review rather than an unexplained score. That clarity protects the quality of the recruiting process as much as it supports compliance.

Candidate information is also a data-governance issue

Resume files can contain names, addresses, phone numbers, employment history, education, and other personal information. Recruiter notes may include additional sensitive details.

Before uploading that information into an AI system, a staffing firm should know which tools are approved, how the provider handles submitted data, what contractual protections apply, and whether the proposed use fits the firm’s privacy commitments and client agreements.

The rule should be concrete. “Do not put sensitive information into AI” is less useful than showing recruiters which fields and documents may be used in which approved systems.

Train recruiters for the decisions they make

Generic prompt training is not enough.

Recruiters need to know how AI affects the responsibilities already attached to their role. Training should cover:

1. what the approved tools may be used for
2. how to verify summaries and extracted qualifications
3. how to recognize when screening criteria may be problematic
4. when disability accommodation issues need human attention
5. what candidate data can be entered into which system
6. when an AI recommendation must be documented or challenged
7. which decisions require a human owner

The Department of Labor’s October 2024 AI best-practices materials emphasized meaningful human oversight for significant employment decisions, worker data protection, transparency, and training. Those materials were withdrawn in January 2025 and no longer reflect current federal policy, but the design principles remain a reasonable checklist for a staffing operation, and they closely resemble what several state laws now require.

Where you place people matters more than where you sit

North Carolina has not enacted a general AI employment law for private employers. If a firm sources, screens, and places candidates only in North Carolina, its obligations come from existing anti-discrimination law and from professional and contractual duties.

That changes the moment a placement crosses a state line, because these obligations generally follow the worksite, not the headquarters.

North Carolina is not silent, only quiet. Governor Josh Stein’s Executive Order No. 24 (September 2, 2025) directs state government toward trustworthy AI, and NCDIT’s Responsible Use of Artificial Intelligence Framework, though written for state agencies, is the closest local reference standard available.

House Bill 1161, the Omnibus Artificial Intelligence Protections bill (filed April 30, 2026; referred to committee May 4, 2026), is the one to watch. Part III would create a Fair AI Hiring Act and defines “automated employment decision tool,” “bias audit,” “independent auditor,” and “covered employment decision.” Three details in the draft are worth knowing now.

First, a covered employment decision includes a remote position performed primarily by someone who resides in North Carolina. Coverage would follow where the person sits, which is the same principle that already governs multi-state exposure, running in the other direction.

Second, the bill would require written notice to an applicant at least 10 business days before an automated employment decision tool is used, disclosing what the tool evaluates, linking to a bias audit summary, and stating the applicant’s right to request an alternative evaluation process. A firm that cannot describe what its tool evaluates cannot write that notice.

Third, enforcement over private employers would run through the Attorney General, with civil penalties assessed per violation per day, and private civil actions would remain available. The Commissioner of Labor’s role in the draft is to set the audit standards private employers are measured against, not to bring the case.

It has not been enacted, and most bills of this size do not pass in their first form. It does show what a future compliance request will look like, which is a reason to start the documentation habit now rather than later.

A note on Colorado. Colorado passed an AI act in 2024, delayed it twice, saw enforcement paused by a federal court in April 2026, then repealed and replaced it with SB 26-189 in May 2026 with a new January 1, 2027 date. The tempting lesson is to wait for the law to settle. That is the wrong lesson. The underlying anti-discrimination exposure never moved. Only the compliance paperwork did.

A practical control model for a 25-person staffing firm

A small or midsize staffing company can start with five management decisions:

1. Inventory: Identify where AI is already being used in sourcing, screening, communications, matching, and administration.
2. Classify: Separate low-risk assistance from tools that influence consequential employment decisions.
3. Approve: Decide which tools and use cases are permitted and what data may be used.
4. Review: Define where meaningful human review is required and what evidence reviewers need.
5. Monitor: Periodically check outcomes, complaints, overrides, errors, and changes to the vendor or model.

The purpose is not to turn recruiters into compliance officers. It is to help them use faster tools without losing control of the decision process.

AI can help a recruiter see more information, draft faster, and reduce repetitive work. It can even help surface candidates a recruiter might otherwise miss. But when the decision affects a person’s opportunity to work, someone in the organization needs to be able to explain the criteria, review the result, and own the decision.

That is a human responsibility.

Before you talk to a vendor, know what to ask. The five questions above are a good starting script, and our AI Assessment can help you see where your firm’s AI use already needs more structure. Or book a free call and we’ll walk through your vendor questions and your state footprint together.