How does the EU AI Act apply if we are a US-only company?

Four ways. You hire EU residents working remotely. You operate any EU subsidiary that uses your hiring tools. Your vendor's tool is subject to Act conformity assessment, and your data flows trigger compliance obligations. Your AI governance posture becomes the de facto global standard once the Act is the strictest regime you operate under. Most US-only companies underestimate the first and last paths; both routinely apply.

Which AI Agents Should You Build for HR?

TL;DR

Agent	Verdict	Why
Job-description quality	Build now	Internal-only edits; zero candidate-decision surface
Interview scheduling	Build now	Logistics only; no selection or quality calls
Candidate communication (inbound)	Build now	Equal information access reduces adverse impact
Onboarding	Build now	Post-hire; selection has happened
Resume screening	Hold 12 mo	Mobley v. Workday case law unsettled
Video-interview analysis	Don’t build	Banned in IL, regulated in MD, prohibited in EU
Performance prediction	Don’t build	Disparate-impact problem on every promotion / PIP decision
Compensation recommendation	Hold 18 mo	NLRA exposure + unsettled case law
Always-on sentiment monitoring	Don’t build	Privacy, NLRA, trust cost too high
Internal-mobility matching	Hold 12 mo	Inherits historic-bias signal in training data
Termination recommendation	Don’t build	End the meeting if a vendor pitches this
”Ask HR” employee chatbot	Buy, don’t build	Fiduciary / Title VII exposure on hallucination

A CHRO who deploys an AI screening agent without a published model card and a queryable audit log is one EEOC complaint away from a board-level crisis. So the right question isn’t which HR agents could we build — it’s which ones will your legal team sign off on without a six-month review.

That filter cuts the universe of HR AI agents from hundreds down to four. This piece walks through those four — and through the eight that look ready but aren’t.

The legal frame, in 200 words

US employment law treats decisions that adversely affect protected classes as the employer’s responsibility, regardless of who or what made the decision. An AI screening tool that disproportionately filters out women, candidates over 40, or minority applicants exposes the employer — not the vendor — to disparate-impact liability under Title VII, the ADEA, the ADA, and (in the EU) the AI Act, which classifies HR AI as “high-risk” by default and requires a conformity assessment before market deployment.

Three things change the math:

Decision rights. An agent that recommends and is overridden by a human is treated differently from one that screens out without human review. The ratio of recommend-only to screen-out determines your exposure tier.
Auditability. If you can’t reconstruct why the agent made each decision, you can’t defend it in regulatory discovery. Black-box vendors lose this case.
Demographic monitoring. EEOC guidance (2023, updated 2025) explicitly expects employers to test their hiring tools for adverse impact quarterly. Most don’t.

These three constraints — decision rights, auditability, and demographic monitoring — define what’s deployable and what isn’t.

The four agents that pass legal review

(In order of build complexity, easiest first.)

1. The job-description quality agent

What it does: rewrites internal hiring requisitions for clarity, consistency with role family, and removal of language correlated with adverse impact (gender-coded adjectives, age-coded phrases, ableist framing). Read-only on the candidate side.

Why it passes: zero candidate-impact decisions. The agent edits internal artifacts before they reach the labor market. Output is reviewed by the hiring manager.

Realistic ROI: a mid-sized employer (5,000+ FTE) typically recovers 1.5–2 percentage points of qualified-applicant-rate on previously low-yield roles. The cost saving compounds: every requisition that closes faster is one less month of recruiter capacity.

Build cost: light. A well-prompted LLM with a 2,000-word internal style guide and an EEOC-language reference does most of this. Engineering effort: 2 person-weeks plus quarterly maintenance.

2. The interview-scheduling agent

What it does: owns the back-and-forth between candidate, recruiter, and hiring manager calendars. Doesn’t make any candidate-quality decisions. Doesn’t speak about the role.

Why it passes: zero adverse-impact surface. Scheduling is the most automatable, least legally exposed HR workflow that exists.

Realistic ROI: recruiters spend 20–40% of their week on scheduling friction. A working agent reclaims most of that. For a 30-recruiter team, that’s roughly $750K–$1.2M of recovered capacity annually.

Build cost: light to medium. The technical surface is calendar-API integration plus tone-controlled email composition. The non-technical surface — the escalation rules — is what most teams underbuild.

3. The candidate-communication agent (inbound only)

What it does: answers candidate questions about the role, company, benefits, application status, and interview logistics. Does not read or evaluate applications.

Why it passes: information access, not selection. The agent gives every candidate the same access to the same information — which is, if anything, an adverse-impact-reducing control. EEOC guidance has been favorable on this category specifically.

Realistic ROI: 60–80% deflection on candidate FAQs (status checks, scheduling, benefits), measured at companies with 1,000+ open requisitions at any time. Removes pressure from coordinator headcount.

Build cost: medium. The work is in the knowledge base, not the agent. The agent is a 4-week build; the knowledge base is a 12-week curation effort.

4. The onboarding agent

What it does: takes a hire from offer-accepted to day-30 productive. Owns the I-9, benefits enrollment, IT provisioning kickoff, manager handoff sequence, and the new-hire question queue.

Why it passes: post-hire. Selection has happened; the legal exposure surface for adverse-impact has closed. Privacy and data-handling rules are the dominant constraint instead, and they’re well-understood.

Realistic ROI: new-hire ramp-time-to-productivity is the second-most-expensive line in your talent budget. A working onboarding agent shaves 2–4 weeks. At an average loaded cost of $12K/month, that’s $6K–$12K per hire — visible in the next quarter’s people-cost line.

Build cost: medium-heavy. The integrations (HRIS, IT ticketing, benefits broker, payroll, learning system) are the work. The agent is the easy part.

The eight agents that look ready but aren’t

These will be on every vendor pitch deck you see this quarter. None of them should be your first deployment.

Resume screening agents. The thing every CHRO’s CTO is being asked to build, and the one most likely to land you in court. Vendors will tell you their model is “audited for bias.” Ask for the technical report. Ask for the most recent quarterly disparate-impact test results. Watch the pitch fall apart. Defer until your governance stack is mature.

Video-interview analysis agents. Some claim to assess “fit,” “communication skills,” or “engagement” from facial expressions. Illinois has banned this practice outright. Maryland regulates it. The EU AI Act prohibits it for hiring. Buying this category is buying a regulatory liability.

Performance-prediction agents. The conceit that an LLM can predict an employee’s future performance from their resume + assessment + first 90 days. Setting aside the methodological problems, the legal frame is brutal: any signal that correlates with a protected class becomes a disparate-impact problem the moment it’s used to inform a promotion or PIP decision.

Compensation-recommendation agents. Adjacent to performance-prediction, with the added complexity of NLRA exposure (collective compensation discussions are protected). Hold for 18 months. The vendors will get to you when the case law settles.

Always-on employee-sentiment agents. Companies are pitching agents that analyze Slack/email to flag attrition risk or burnout. The privacy-and-NLRA exposure is severe; some states (notably California, New York) require explicit notice and opt-out. Even when legal, the trust cost on your workforce is rarely worth the lift.

Internal-mobility matching agents. The plausible idea: match employees to internal openings using skills and history. The hidden risk: the historic-performance signal that powers the match is exactly the signal that encodes existing bias. Every promotion the agent recommends inherits whatever bias was in the data it learned from.

Termination-recommendation agents. The reductio ad absurdum of HR AI. Don’t build this. Don’t buy this. If a vendor pitches this, end the meeting.

AI-powered “Ask HR” chatbots for current employees. The reasonable-seeming one that breaks in production. Employee questions about leave, benefits, accommodations, and complaints are dense with legally protected categories. Hallucination on a benefits question is a fiduciary problem; hallucination on a harassment complaint is a Title VII problem. The auditability requirement here is severe enough that most teams should buy a vetted vendor solution rather than build.

The architectural decision under all of this

If you’re building any of the four legal-safe agents above, three architectural commitments determine whether you can scale them.

1. Decision rights are explicit and logged. Every agent action that touches a candidate, employee, or employment decision is tagged: recommend, prepare, schedule, answer. Anything that decides — screens, ranks, eliminates — requires a human override step that is also logged.

2. The audit log is queryable, not just stored. “We have logs” is what every vendor says. The question is whether you can answer “show me every decision the agent made about applicants in protected class X over the last quarter” in under an hour. If you can’t, your audit log is not legally useful.

3. The eval harness runs quarterly without engineering effort. EEOC expects quarterly disparate-impact testing. If running that test requires three engineers and a Jira ticket, it won’t happen. Bake it into a scheduled job before you deploy.

These three commitments are the same regardless of which of the four agents you build. They’re not a per-agent overhead; they’re the platform you build once before any agent goes live.

The counter-argument

A reasonable CHRO will push back: “My peers are deploying screening agents. Are we falling behind?”

Two things to know.

First, the screening-agent space is undergoing a slow-motion correction. Three of the largest enterprise vendors quietly rolled back screening features in 2025 in response to the Mobley v. Workday litigation. The market signal is conservatism, not aggression.

Second, the productivity gap from the four agents above is large enough that you’ll outpace any peer who’s playing in the high-risk categories. Onboarding-time-to-productivity alone is a six-figure win per 100 hires for a mid-sized employer. You don’t need to take on adverse-impact exposure to extract real ROI from AI in HR — you need to be more disciplined about which use cases you back.

What to do this quarter

Run a 60-minute legal-exposure inventory. What HR AI tools are already in your environment? (Hint: more than your CHRO knows.) Most companies discover a Slackbot, an outreach agent, and a sourcing tool that no one approved.
Pick one of the four legal-safe agents to ship by quarter-end. The job-description quality agent is the lowest-friction first deployment. It teaches your team the discipline you’ll need for the harder ones.
Build the platform, not just the agent. Decision-rights tagging, queryable audit log, scheduled eval. These are infrastructure, not features.
Defer the screening conversation by 12 months. The case law and the vendor stack will both be in better shape. Your peers who go early will pay the tuition for the rest of the market.

The companies that win the HR-AI cycle won’t be the ones who deployed the most agents. They’ll be the ones who deployed the right four, well, before regulation made the others obsolete.

FAQ

Will an AI resume-screening agent pass our employment counsel’s review in 2026? Almost certainly not, in any form that automates rejection. Counsel reviewing the Mobley v. Workday docket is now writing position memos that say “no autonomous adverse-action AI tools.” Tools that rank for human review still pass, but only if you can produce a queryable audit log and quarterly disparate-impact data. If the vendor can’t give you both, the tool fails counsel’s review on contact.

What does the EEOC actually expect us to do with our AI hiring tools? Three things, per the 2023 Technical Assistance document and 2025 update. (1) Run a four-fifths-rule disparate-impact test at least quarterly, broken down by sex, race, age 40+, and disability where you have data. (2) Document the test, the population tested, and the remediation taken. (3) Notify candidates that AI is in use, where state law requires (NYC, Illinois, Maryland, Colorado already do; expect more by 2027). The agency has signaled it will treat absence of testing as evidence of negligence in enforcement actions.

How does the EU AI Act apply if we’re a US-only company? Four ways. (1) You hire EU residents working remotely. (2) You operate any EU subsidiary that uses your hiring tools. (3) Your vendor’s tool is subject to Act conformity assessment, and your data flows trigger compliance obligations. (4) Your AI governance posture becomes the de facto global standard once the Act is the strictest regime you operate under. Most US-only companies underestimate paths (1) and (4); both routinely apply.

What’s the cheapest first AI agent to deploy in HR with a real ROI? The job-description quality agent. Engineering cost is roughly 2 person-weeks; ongoing maintenance is light; the legal exposure is negligible because nothing reaches a candidate before a hiring manager has reviewed it. The recovered recruiter capacity from improved qualified-applicant rates pays it back in the first quarter for any employer over 1,000 FTE.

How do we test our existing AI hiring tools for bias? Run an EEOC four-fifths-rule analysis on the last 12 months of hiring outcomes — selection rates by protected class at each stage where the AI tool acts. If any class group’s selection rate falls below 80% of the highest-rate group, you have a presumptive disparate-impact finding that requires either remediation or a documented business-necessity defense. Most companies that have not tested their tools quarterly discover at least one finding on first analysis. Run it before regulators ask.

Working with JAIN on HR AI? We design and build the four legal-safe agents above with decision-rights instrumentation and quarterly disparate-impact testing baked in. Book a 30-minute call.

Related reading: