AI Agents for Recruiting: Screening Without Bias Traps

AI agents for recruiting earn their place on four jobs that surround the hiring decision rather than make it: drafting job descriptions, scheduling interviews, answering candidates’ routine questions, and summarizing applications against the criteria you published. In each, the agent drafts and a person decides. The job to keep away from an agent is screening and ranking people. In the US, a selection tool that screens out a protected group can be unlawful under Title VII even when nobody meant it to, and one that screens out people with disabilities can breach the ADA, whoever built the tool. A person decides on every candidate, every time.

This page is about a US employer’s own hiring. The wider set of places where a person steps in, including a hiring screen, is in human-in-the-loop AI examples; the EU’s rules for AI in recruitment are covered in human oversight under the EU AI Act. Once someone accepts an offer, AI agents for HR takes over.

Four jobs worth handing over

1. Job-description drafts

Give the agent the hiring manager’s notes, the current description if there is one, and your house template, and ask for a draft that separates the essential functions of the role from the nice-to-haves. That split matters later: it is what you will measure every applicant against, and what a reviewer will check a summary against. Ask the agent to flag requirements that may not be needed for the job, such as a degree for a role nobody has needed one for, or a physical requirement the work does not involve. A person approves the final wording and the list of requirements.

2. Interview scheduling

The agent reads the interviewers’ calendars, proposes times, drafts the invitation with the format and length of each interview, and sends reminders once a person has confirmed the slot. Put one line in every invitation that tells candidates how to ask for an adjustment to the process, and route those requests straight to a person. An agent that tries to handle an accommodation request itself is handling medical information it should never see.

3. Candidate FAQs

Most candidate emails ask the same things: where the process stands, what the interview involves, whether the role is remote, when they will hear back. The agent drafts answers from a short document of approved answers. Anything outside that document, and anything about pay, visas, a disability, or why someone was turned down, goes to a person without a draft. Candidates may be talking to a machine without knowing it, so say plainly in your process page that some replies are drafted with AI and reviewed by a person.

4. Summarizing applications

This is the job closest to the line. The safe version: the agent reads each application and writes a short summary against the published criteria, quoting the evidence for each one and saying “not found” where it found none. It does not score, rank, sort or recommend, and it never sees a field it should not weigh, such as a photo or a date of birth. The reviewer reads the summary next to the application, not instead of it. How to check an agent’s summary quickly is in verifying AI-generated work.

Why automated screening and ranking are risky

This is a plain-language summary, not legal advice. State and local laws add to it; talk to employment counsel before any tool filters or ranks applicants.

Title VII and disparate impact

Title VII does not only forbid deliberate discrimination. Section 703(k) of the statute as the EEOC publishes it↗ (opens in a new tab) sets out when “an unlawful employment practice based on disparate impact” is established: broadly, when an employment practice causes a disparate impact on the basis of race, color, religion, sex or national origin and the employer cannot show it is job related and consistent with business necessity. A screening model trained on past hires can do exactly that, with nobody intending it.

The federal Uniform Guidelines on Employee Selection Procedures, 29 CFR 1607.4↗ (opens in a new tab), give the usual rule of thumb, the “four-fifths rule”: a selection rate for any race, sex or ethnic group that is less than four-fifths of the rate for the group with the highest rate will generally be regarded as evidence of adverse impact. The same section adds that smaller differences may still count. Passing the four-fifths check is not a safe harbor.

The ADA

The Justice Department’s guidance on algorithms, AI and disability discrimination in hiring↗ (opens in a new tab), dated May 12, 2022, is still published on ada.gov. It says employers must provide reasonable accommodations during the hiring process unless doing so would create an undue hardship, that employers violate the ADA if their hiring technologies unfairly screen out a qualified individual with a disability, and that this includes “when an employer uses another company’s discriminatory hiring technologies.” Buying the tool does not move the responsibility to the vendor. The page also notes that its guidance has no force of law; the ADA itself does.

Where the federal guidance stands today

Two changes are worth knowing. First, the EEOC’s own technical assistance documents on AI in hiring, one under Title VII and one under the ADA, are no longer published: their former addresses on eeoc.gov return “page not found” as of September 29, 2026. Second, Executive Order 14281↗ (opens in a new tab) of April 23, 2025 tells federal agencies to deprioritize enforcement of statutes and regulations to the extent they include disparate-impact liability, and names 42 U.S.C. 2000e-2, the Title VII section above. An executive order does not change the text of the statute, and section 703(k) is still in it. Less federal enforcement is not the same as less risk.

Local rules

Some cities and states have their own rules. New York City’s Local Law 144↗ (opens in a new tab), enforced since July 5, 2023, bars employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within a year, the audit results are public, and candidates have been given notice. Check where you hire before any tool touches a candidate.

What a human must approve

• Every decision about a candidate: advance, hold, reject or offer. An agent may prepare the paperwork; it never makes the call.
• The requirements in every job description, and every change to them after the role opens.
• Any message that tells a candidate an outcome, and every offer, including pay and start date.
• Every accommodation request, from the first reply onward.
• Any tool, feature or setting that filters, scores or ranks applicants, before it is switched on, with counsel.
• What data the agent can read: the application and the criteria, not photos, dates of birth, or anything else you would not want weighed.

Start with every item approved one at a time and keep it that way for decisions about people. The patterns for building that approval step are in AI agent approval workflows.

A person decides on every candidate

Saying a human decides is easy; making it true takes design. Three habits help:

1. No automatic rejections. If your applicant tracking system can reject on a knockout question or a score, turn that off or have counsel review it first. A summary that says “not found” is a prompt for a person to look, not a verdict.
2. Decide before you read. Write the criteria, and what counts as evidence for each, before the first summary arrives, so the reviewer is not anchored by the agent’s wording.
3. Look at the numbers. Every few weeks, compare how many people from each group advance at each stage. It is the check the four-fifths rule describes, and it catches a problem in your process whether it came from the agent or a person.

A board for the hand-offs

Candidate data belongs in your applicant tracking system, not on a shared board. What a board is good for is the work around it that needs a person: “approve the job description for the support lead”, “review 30 application summaries for the analyst role”, “reply to an accommodation request”. On fenbs, an AI assistant files each as a task with the link to the tracking system in the note and a priority from 1 to 10, and a person moves it through To Do, Next Up, In Progress and Completed. Refer to candidates by their tracking-system ID, never by name or with their details.

The rule that matters most, “the assistant never scores, ranks or rejects a candidate”, goes on the Decisions and rules page. A rule is a decision that holds from now on, the decider is always a person, and every connected AI assistant reads the rules first. History records who changed each task, so you can see which approvals a person gave. Roles are set per company, and moving a task between lanes is its own permission, so the assistant can file and comment while only the hiring team moves work to Completed. Be clear about what fenbs is not: it is not an applicant tracking system, it has no due dates and no assignee you can set, so interview dates stay in the calendar and the owner’s name goes in the note.