Employers and human resources departments are relying on artificial intelligence across the employment lifecycle. Screening tools rank applicants. Interview platforms score responses. Performance systems flag “low performers.” Compensation models recommend pay. Some of those outputs now feed discipline and termination decisions.
The efficiency is real. So is the discrimination risk. A flawed model does not create a new kind of claim. It can create the same claim employers already face under Title VII, the Americans with Disabilities Act, and the Age Discrimination in Employment Act — applied to every person the tool touches, and often with a thinner record of why the decision was made.
State rules are filling in around those federal statutes, and they do not all say the same thing. California made one of them concrete on September 30, 2026, when Governor Newsom signed SB 947, the No Robo Bosses Act. Beginning July 1, 2027, California employers may not use an automated decision system as the sole basis for discipline or termination. Where the system plays a primary role, a human has to independently corroborate the result. Colorado’s narrower automated-decision framework, SB 26-189, applies to consequential employment decisions made on or after January 1, 2027. New York City already requires a bias audit and notice for certain automated employment decision tools.
Those statutes are narrower than the full set of AI uses in HR. The compliance problem is not. The practices below are becoming the practical baseline — whether or not a given state has written a dedicated AI statute yet.
1. Inventory and assess the AI ecosystem
You cannot manage a risk you have not identified. Start by mapping where algorithms touch employment decisions, not where a vendor’s marketing says “AI.”
Cover the full lifecycle:
- Recruiting and sourcing
- Resume screening, ranking, and knockout questions
- Interviews, including video and automated scoring
- Promotions and internal mobility
- Performance assessment and productivity monitoring
- Compensation recommendations
- Discipline, attendance enforcement, and termination
Include tools embedded in a broader HR platform, not only systems built in-house. For each one, record what it does (scores, ranks, flags, recommends, or decides), what data it uses, who sees the output, and whether that output is effectively driving the decision. A manager who always follows the score is not a control. That person is the last click.
This inventory is also the starting point for the tiered AI policy many organizations still do not have. Low-risk drafting and high-stakes employment decisions should not sit in the same bucket.
2. Implement human review that can actually override the tool
Meaningful oversight means three things. The reviewer understands how the system works well enough to question it. The reviewer has authority to override the recommendation. And the reviewer can evaluate whether the outcome fits the organization’s criteria and its legal duties.
Rubber-stamping does not meet that test. “The manager clicked approve” is a log entry, not a review.
Build an escalation path for adverse decisions in particular: rejections, denied promotions, pay decisions that fall outside the range, discipline, and separations. The reviewer should see the underlying information — evaluations, work product, interview notes, personnel records — not only the score. If the output cannot be corroborated, or it is inaccurate, incomplete, or misleading, it should not be the basis for the decision.
California will require a version of this for discipline and termination once SB 947 takes effect. Employers outside California should not treat that date as the reason to wait. A discrimination claim does not need a state AI statute. It needs an adverse action and a record that cannot explain it.
3. Validate tools for adverse impact — and keep the file
Conduct regular testing to see whether an AI system disproportionately affects protected groups. Selection-rate gaps by race, sex, age, disability, and other protected characteristics are the question. A vendor dashboard labeled “fair” is not the answer.
Document the methodology, the results, and any corrective action. Keep that file current. It serves two purposes: it is how you find a problem while you can still fix the tool or the workflow, and it is what you will want if an agency or a plaintiff’s lawyer later asks why a screen-out happened.
Test before broad deployment, and again on a set schedule or after a material change in the model, the data, or the job. A validation study from the vendor’s other customers is a starting point. It is not a substitute for results in your applicant pool and your roles.
4. Rebuild the ADA accommodation path around AI assessments
An algorithmic assessment can screen out a candidate or employee with a disability even when the person can do the job with an accommodation. Timed simulations, video analysis of speech or facial movement, chatbot interviews, and keystroke or productivity monitors are common pressure points.
Build a clear path to request an accommodation before the assessment, not only after a rejection. That path may include an alternate format, extra time, a bypass of the algorithmic screen, or guaranteed human review. Tell candidates the tool is being used, and tell them how to ask. A process that exists only in an internal policy no one sees is not a process.
Train recruiters and the vendor’s implementation team on the same path. A tool that cannot be paused or bypassed is a tool that will create ADA exposure the first time someone asks.
5. Plan for regulatory fragmentation
Multi-state employers should expect the map to stay uneven.
New York City already imposes bias-audit and notice duties on certain automated employment decision tools. California’s SB 947, effective July 1, 2027, is aimed at discipline and termination, not every hiring screen. Colorado’s SB 26-189, effective January 1, 2027, focuses on transparency, notice, recordkeeping, and human review when automated decision technology is used in a consequential employment decision. Other states are writing narrower — or broader — versions of the same idea. A single national posting can trigger more than one of them.
Two operating choices are common. Adopt the strictest standard across the footprint, which is simpler to administer and may over-comply in some states. Or maintain jurisdiction-specific protocols, which tracks the law more tightly and is harder to run. Both carry trade-offs in complexity and risk. What works in one state can fall short in another. Pick deliberately, write it down, and revisit it when a statute or a worksite changes.
6. Scrutinize vendor contracts — and do not accept “fair” at face value
Most of these tools are bought, not built. The contract is where risk either stays with you or is shared.
Review agreements for:
- A plain explanation of what the system does and what data it uses
- Compliance with anti-discrimination and accessibility laws, stated as an obligation rather than a marketing line
- Data use, retention, and deletion
- Audit rights, and the right to receive validation results
- Indemnification if the tool produces a discriminatory outcome
- The vendor’s duty to provide the information you need for your own notices, testing, and defense file
- What happens if a law changes or a model is updated without notice
Ask for documentation of validation studies, information about training data, and an explanation of how the system reaches a decision. If the vendor cannot explain the output, your reviewers cannot corroborate it, and you will not be able to answer a later question from a candidate, an agency, or a court.
Courts examining AI hiring tools have already been willing to treat a software vendor as the employer’s agent for some liability purposes. A contract that leaves every obligation on the employer, and every explanation with the vendor, is not a neutral allocation.
Use the tool. Do not let it police itself.
AI offers real gains in speed, consistency, and scale. Those gains are worth taking where the task is low-risk and the human stays in charge of the decision. They are not a reason to treat a score as the decision.
The working strategy is straightforward. Inventory the tools. Require review that can override them. Test for adverse impact and keep the file. Give people a real ADA path. Choose how you will handle a fragmented map. And make the vendor produce the documentation your compliance program needs.
The right depth for a 12-person employer using one attendance flag is not the right depth for a multi-state employer whose performance model feeds every termination recommendation. The tool, the job, and the states where people work still matter.
Follow Synergy HR on social media for practical updates as more states write rules on AI in hiring, discipline, and termination — and what those rules mean for employers outside any one state.
