HR-Built AI Agent Compliance Risks Nobody Reviews

HR-Built AI Agent Compliance Risks Nobody Reviews

The biggest HR-built AI agent compliance risks come from agents that quietly shape decisions about employees: who gets flagged for attendance, whose PTO request waits, how harsh a performance plan reads. If HR builds that agent in a no-code tool, the employer carries the same discrimination risk as if it had bought a vendor product. The difference is that nobody reviews it, because there's no purchase to trigger a legal check and security tools don't flag an approved app doing its job.

Who this is for: HR operations, people analytics and IT governance leaders whose teams can build agents and workflows in low-code platforms they already license.

Key takeaways

  • AI hiring rules were mostly written around tools an employer buys. Agents HR builds itself often skip those reviews entirely.
  • Shadow-AI monitoring looks for unauthorized tools and data flows, not authorized tools making risky decisions.
  • Disparate impact doesn't require intent. A rule applied evenly can still fall harder on people on protected leave or with accommodations.
  • Govern by what the agent's output touches, not by which platform built it.
  • Before go-live, check for proxies, confirm a real human decision point, save the agent's rules and plan an outcome check using the four-fifths comparison.

Why don't HR-built agents trip legal or security review?

Both review systems are built around different triggers. Legal review usually starts at procurement, and security review starts when something unapproved shows up. An agent built inside a licensed platform hits neither.

A quick definition. A no-code (or low-code) agent is an automated assistant someone assembles from drag-and-drop blocks and plain-language instructions, without writing software. An HR generalist can build one in an afternoon to triage PTO requests, flag attendance patterns, draft a first pass at a performance improvement plan (PIP) or route accommodation requests. None of that looks like "AI in hiring." It looks like workflow automation. If you've worked through a governance playbook for shadow IT and business-built apps, you'll recognize the pattern: the tool is approved, the use is not reviewed.

The employment law review. The emerging rules on AI in employment assume a moment where a tool gets selected, reviewed and contracted. New York City's Local Law 144 requires a bias audit within one year before an automated employment decision tool is used, public information about that audit and notice to candidates and employees 10 business days before use. The city's consumer and worker protection agency began enforcing it on July 5, 2023. Illinois's Artificial Intelligence Video Interview Act requires notice, an explanation of how the AI works and the general characteristics it evaluates, and consent before AI analyzes a video interview. An agent built on a platform you already license has no selection moment, so nobody asks whether these rules apply. Our overview of state AI employment laws in Colorado, Illinois and NYC covers how far each one reaches.

Many of these laws focus on hiring and promotion. Internal agents used for adjacent work, such as attendance flags, scheduling or discipline drafts, may sit in ambiguous territory under them. But the ambiguity only concerns the AI-specific laws. Title VII, the ADA and the FMLA apply no matter what tool shaped the decision.

The security review. Shadow-AI monitoring is generally built to catch unapproved tools and unapproved data flows. An agent built inside an approved platform, used exactly as intended, doesn't raise those flags, even if it's shaping decisions that touch a protected characteristic.

Review systemWhat normally triggers itWhy an HR-built agent misses it
Legal or compliance reviewBuying or contracting a new AI toolNo purchase; the platform is already licensed
Security or shadow-AI monitoringUnapproved apps or unusual data movementApproved app, approved data, normal use
Bias audit under local AI lawsA tool identified as used for hiring or promotionAgent framed as "workflow automation," not a decision tool
Change managementIT release cycleBuilder edits rules directly with no version control

Where does the real legal exposure sit?

The exposure sits in disparate impact: a consistent rule that falls harder on a protected group. The agent doesn't need to be malicious or even usually wrong. It applies the same rule to everyone, so any pattern it creates is consistent, and a consistent pattern that tracks a protected characteristic is exactly what this analysis looks for.

Disparate impact means a neutral-looking practice that disproportionately harms people in a protected group. Under Title VII (42 U.S.C. 2000e-2(k)), a disparate impact claim is established when a complaining party shows a particular employment practice causes a disparate impact on the basis of race, color, religion, sex or national origin, and the employer fails to show it is "job related for the position in question and consistent with business necessity." The claim can also succeed if a less harmful alternative practice exists and the employer refuses to adopt it. Intent isn't part of the test.

Three examples of how this happens:

  • The attendance agent. It flags "absence patterns of concern" based on frequency and clustering, applied to everyone. Employees on intermittent FMLA leave produce exactly the pattern it was built to find. FMLA regulations say employers "cannot use the taking of FMLA leave as a negative factor in employment actions, such as hiring, promotions or disciplinary actions; nor can FMLA leave be counted under no fault attendance policies" (29 CFR 825.220(c)). If the agent's flag reaches a manager's desk, that's an interference problem nobody intended.
  • The PTO triage agent. It deprioritizes "last-minute" requests. Some disability-related absences and some religious observances can't be scheduled weeks ahead, so the rule lands harder on those employees.
  • The PIP drafting assistant. It pulls language from an employee's history. If that history includes accommodation-related absences it was never told to exclude, the draft may come out harsher for exactly the wrong people, and a busy manager may not change it.

No-code builders are designed to remove the need to involve IT, legal or a data scientist. That's also why nobody with disparate impact training is in the room when the rules are set. For a sense of how leave data patterns draw scrutiny on their own, see the FMLA data patterns that trigger a DOL audit.

How do you tell if an agent's output is skewed?

Compare outcome rates by group. The federal Uniform Guidelines give a simple starting test, the four-fifths rule: a selection rate for any race, sex or ethnic group that is "less than four-fifths (4/5) (or eighty percent)" of the rate for the group with the highest rate "will generally be regarded by the Federal enforcement agencies as evidence of adverse impact" (29 CFR 1607.4(D)).

That rule was written for selection decisions, and leave status isn't itself a Title VII category. But the same arithmetic is a fast screen for any flag an agent produces.

Worked example: an attendance-flag agent

Say your attendance agent ran for one quarter across 240 hourly employees. Treat "not flagged" as the favorable outcome.

GroupEmployeesFlaggedNot flaggedFavorable rate
No protected leave on file2002018090%
Intermittent FMLA or accommodation on file40122870%

Divide the lower rate by the higher one: 70 / 90 = 0.78, or 78%. That's under the 80% line. Even before a lawyer weighs in, you know the agent is treating the leave group differently and the rules need a look, most likely to exclude protected absences from the count. Run the same comparison by sex, race and age band if you have the data, and keep the numbers small groups produce in perspective: with 40 people, two or three flags can swing the ratio. Our guide to self-auditing an ATS for disparate impact covers the statistical caveats in more depth.

What does a reasonable governance line look like?

Draw the line by what the agent's output touches, not by which platform built it. Banning HR from no-code AI tools trades a compliance gap for a productivity loss, and it rarely holds, because the tools are already licensed for other work.

Agent typeExamplesReview level
InformationalAnswers policy questions, summarizes a handbook, finds a formStandard IT and data review
Drafting with real human reviewFirst drafts of letters or job postings that a person rewritesLight review of inputs and prompts; spot-check outputs
Decision-shapingPrioritizes requests, flags employees, scores candidates, sets the tone of discipline documentsPre-deployment HR, legal and data review; logging; quarterly outcome checks

For a broader structure, the NIST AI Risk Management Framework, released January 26, 2023 and intended for voluntary use, is organized around four functions: govern, map, measure and manage. NIST added a Generative AI Profile on July 26, 2024. Mapping each HR agent to those four verbs is a practical start. If you're building that process from scratch, start with how leaders set up an AI governance framework step by step.

What should you check before a decision-shaping agent goes live?

Five checks, done by the builder with one HR compliance reviewer, cover most of the risk. Run them in this order:

  1. Proxy check. Does the agent use anything correlated with a protected characteristic, such as leave status, accommodation history, tenure (which can stand in for age) or schedule constraints? Remove it or justify it in writing.
  2. Human checkpoint. Is there a person downstream who sees the reasoning, can override it and actually does? A reviewer who approves everything is a rubber stamp.
  3. Rule snapshot. Save the agent's instructions, rules, data sources and version at go-live and at every change. "What did the agent do six months ago?" is hard to answer for a no-code agent.
  4. Outcome monitoring. Schedule the four-fifths comparison above for 30 days after launch, then quarterly.
  5. Registry. Add the agent to a list of HR automations with an owner, purpose, agent type and review date.

Rough time budget

These are planning estimates from typical HR ops work, not benchmarks. A proxy check and human-checkpoint design usually take one to two hours for a simple agent. A rule snapshot takes 15 minutes if you export the configuration and store it with a date. A quarterly outcome check takes one to three hours depending on how clean your HRIS data is. Under a day per agent per year is a small price against defending a disparate impact claim.

What to ask the builder

  • What decision does a person make after seeing this output?
  • Which fields does the agent read, and could any reveal leave, disability, religion, age or pregnancy?
  • What happens if the agent is wrong, and who would notice?
  • How often do reviewers change the agent's recommendation?
  • Where is the current version of its instructions saved?

Our guide to governing AI agents in HR goes deeper on permissions, approval gates and audit logs.

Common mistakes

  • Treating "internal" as "exempt." Local Law 144 doesn't carve out tools you build, and federal anti-discrimination law never cared who wrote the rule.
  • Counting all absences the same. FMLA leave can't be counted under no-fault attendance policies, so an agent that counts it is building in a violation.
  • Calling a rubber stamp "human review." An override rate near zero means the agent is making the decision.
  • Editing live rules without a record. If you can't reconstruct last quarter's logic, you can't defend last quarter's outcomes.
  • Writing a policy nobody reads. Put the five checks into the automation request form itself. An AI acceptable use policy for HR teams sets the rules, but the form is where builders actually see them.

FAQ

Are employers liable for AI tools they build themselves?

Yes. Title VII's disparate impact test asks whether a practice causes the impact and whether the employer can justify it, not who built the tool. A homegrown agent removes even the option of pointing at a vendor. The FMLA and ADA apply the same way.

Does NYC Local Law 144 apply to an agent we built ourselves?

It can. The law covers automated employment decision tools used for hiring or promotion decisions about candidates and employees in New York City, and it doesn't exempt tools you build. Whether a particular internal agent qualifies depends on what it does and how much it influences the decision. Ask counsel to review any agent that scores, ranks or screens people.

If a manager makes the final call, are we protected?

Only if the review is real. If managers routinely accept the agent's flag or draft without changes, the agent is effectively making the decision, and a disparate impact claim can still focus on the practice. Train reviewers, show them why the agent flagged someone and track how often they override it.

What is the four-fifths rule and does it apply to internal agents?

It's a guideline that treats a group's favorable outcome rate below 80% of the top group's rate as evidence of adverse impact. It was written for selection decisions such as hiring and promotion. For flags, prioritization or discipline drafts, it's still a practical early-warning screen, though not a legal safe harbor.

Should we stop HR from building AI agents?

Usually not. Informational and drafting agents can save real time with modest risk. Focus controls on agents that prioritize, flag, score or shape documents about individual employees, keep a registry and check outcomes regularly.

What to do next

  • Ask IT for a list of agents and automations built in your licensed no-code platforms that touch HR data.
  • Sort them into informational, drafting and decision-shaping, and start with the last group.
  • Run the four-fifths comparison on any decision-shaping agent that has been live for a quarter or more.
  • Add the five-point checklist and the builder questions to your HR automation request process.
  • Agree between HR, legal and IT on who owns the registry and the quarterly review date.

This article is general information, not legal advice.

Sources

Last reviewed: November 2026.

Comments

Popular Posts

Why Workday Onboarding Breaks Down for Frontline Employees

Are Biometrics Better Than Passwords? What NIST Says

How to Improve the Customer Experience (CX)

How to Market Your Handyman Services to Small Business Owners

Interview Tips for Beginners: A Step-by-Step Prep Plan

Are Mobile Apps Tracking Your Every Move? How to Check

Predictive Scheduling Laws: Where They Apply and How to Comply

How Do I Log In and Sign In to Workday HCM

Multi-State Remote Employees: Payroll Tax Registration Steps

End-of-Day Routines That Boost Productivity at Work