Automated Employment Decision Tool (AEDT): What Counts and What NYC Requires

An automated employment decision tool (AEDT) is any computational process that substantially assists or replaces a discretionary decision about hiring or promotion. That definition comes from NYC Local Law 144, and it is the one that matters, because NYC is the jurisdiction actively enforcing it. If a tool is an AEDT and you use it on a candidate for a job in New York City, you need an independent bias audit from the prior 12 months, a published summary of that audit, and at least 10 business days notice to the candidate.

The phrase sounds like it only covers exotic AI. It does not. A spreadsheet-driven scoring model can be an AEDT. A resume ranker almost certainly is. This piece explains what actually counts, what the law requires, what happens if you get it wrong, and why the enforcement picture changed in December 2025.

What is an automated employment decision tool?

Local Law 144 defines an AEDT as a computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues a simplified output, such as a score, classification, or recommendation, and that substantially assists or replaces discretionary decision making about employment decisions.

Three parts of that definition do the work, and they are worth reading slowly:

• Simplified output. A score, a rank, a tag, a yes or no. A tool that returns a transcript is not producing a simplified output. A tool that returns "82, recommend advance" is.
• Substantially assists or replaces. This is the phrase everyone argues about. The rules say it means relying solely on the output, or using it as one factor weighted more than any other, or using it to overrule a human conclusion.
• Employment decision. Screening for hiring, or deciding promotion. Not performance management, not compensation, not scheduling.

The middle test is the one that catches people out. Employers often assume that because a human reviews the output, the tool falls outside the law. That is not what the rule says. If the score is the heaviest single factor a recruiter weighs, it substantially assists the decision even though a person clicks the button.

Is it an AEDT? Some worked examples

That last row is the one that surprises people. Nothing in the definition requires a neural network. A weighted scoring model built in a spreadsheet, if it produces a score that drives the decision, fits the definition on its face.

What does NYC Local Law 144 require?

Four obligations, and they are separate. Meeting three of them is not partial credit.

• Bias audit within the prior 12 months. Not once. Every year, on a rolling basis, for each tool you use.
• Conducted by an independent auditor. Not you, and not the vendor who sold you the tool. Independence is the point, and it is the requirement most commonly fudged.
• Published summary of results. Publicly available on your website. The audit must test for impact by sex, by race and ethnicity, and by intersectional combinations of those categories.
• Notice to candidates at least 10 business days before use. Business days, not calendar days. This one bites teams running fast funnels.

Note that the bias audit obligation sits with the employer, not the vendor. A vendor can hand you an audit of their tool, and a good one will, but the accountability for using it on your candidates is yours. This is also the EEOC's position on selection procedures generally: if a tool produces adverse impact on a protected group, the employer answers for it regardless of who built the thing. Liability does not transfer with a purchase order, which is why a serious vendor conversation starts with what the audit found rather than whether an audit exists.

What are the penalties for using a non-compliant AEDT?

The headline numbers are small and the multiplier is not. A first violation is $500. Each additional violation on that same day is another $500. Each subsequent violation runs $500 to $1,500. The detail that matters is that each day you use a non-compliant AEDT is a separate violation, and failing to give the required notice is a separate violation again.

So a single non-compliant tool, running for a month, is not a $500 problem. It is roughly $15,000 to $45,000 for that tool in that month before anyone counts per-use violations, and a high-volume employer screening hundreds of candidates a week accumulates exposure quickly. The reputational cost of a published enforcement action against your careers page tends to exceed the fine anyway.

Is anyone actually enforcing this?

Until recently, the honest answer was barely. That changed on December 2, 2025, when the New York State Comptroller published an audit of how the Department of Consumer and Worker Protection (DCWP) had enforced Local Law 144 between July 2023 and June 2025. The findings are unusually blunt for a government audit, and they are worth knowing because they are a preview of what happens next.

Read that table as a forecast rather than a scandal. A regulator that has publicly committed to proactive investigation, formal complaint tracking, better bias-audit review, and technical support from the NYC Office of Technology and Innovation is a regulator that will find more than one issue in the next 32 companies it looks at. The gap between 1 and 17 was not a gap in compliance. It was a gap in looking.

There is a second, quieter signal in the methodology. The Comptroller found those companies by reading published academic and advocacy research, not by waiting for complaints. Your compliance posture is discoverable by anyone with a browser, because the law requires you to publish the audit summary. That cuts both ways.

Which other states regulate AI in hiring?

NYC is the strictest, but it is not alone, and the map moves fast enough that a 2026 article you find elsewhere may already be wrong.

• Illinois has had the Artificial Intelligence Video Interview Act since 2020: if you use AI to analyze a video interview, you must tell the applicant, explain how the technology works and what it evaluates, and get consent before it runs. Separately, HB 3773 amended the Illinois Human Rights Act and took effect January 1, 2026, making discriminatory use of AI in employment decisions an unlawful practice.
• Maryland requires an applicant's signed waiver before you use a facial recognition service to create a facial template during an interview.
• Colorado is the one most articles get wrong. The much-discussed Colorado AI Act (SB 24-205) was delayed to June 30, 2026, and then replaced: Governor Polis signed SB 26-189 on May 14, 2026, which substantially rewrites the framework and takes effect January 1, 2027. If a compliance guide tells you to prepare for SB 24-205 in June 2026, it is out of date.

The practical implication for a multi-state employer is that you cannot run one screening process for NYC and a looser one everywhere else, because the rules differ on different axes: notice in Illinois, consent for facial templates in Maryland, audits and notice in NYC. It is cheaper to build the strictest version once and run it everywhere. Teams with obligations across several jurisdictions usually end up needing something that keeps every deadline and control mapped to the rule it satisfies, because an annual audit that lapses quietly is the most common way a compliant program stops being compliant.

How to comply without stalling your hiring

The work is more boring than the legal writing suggests. In order:

• Inventory your tools. Every scoring, ranking, or knockout mechanism in your funnel, including the spreadsheet model somebody built three years ago and the ATS feature nobody switched off.
• Decide which are AEDTs. Apply the substantially-assists test honestly rather than hopefully. Get counsel on the contested ones.
• Get an independent audit for each, and diary the renewal. Twelve months is a rolling deadline and it expires without telling you.
• Publish the summary where a candidate can find it. Not buried in a PDF nobody links to.
• Fix your notice timing. Ten business days is two full weeks. Build it into the job posting rather than bolting it on at screening.
• Keep a human on every rejection. This does not remove the AEDT obligation, but it is the single biggest reduction in your adverse-impact exposure.

Is InterviewAgent.ai an AEDT?

Yes, and we would rather say so plainly than play definitional games. Our agent produces a score and a rank that substantially assist your decision about who advances, which is what the definition describes. So we run bias audits against EEOC guidance and Local Law 144, disclose to candidates that they are interviewing with AI, and take consent before the interview starts.

What we do not do is reject anyone. The AI screening agent conducts the interview, scores each answer against your rubric with the transcript attached as evidence, and advances candidates to a recruiter. A human makes every decision, sees why each score landed where it did, and can overrule it. That design is partly compliance and partly conviction: an automated rejection is the one action in hiring you can never explain to the person on the other end.

If you want the mechanics of the audit itself rather than the definitions, our guide to the NYC Local Law 144 bias audit covers what auditors test and what the summary has to contain. For how structure reduces the underlying bias in the first place, the structured vs unstructured interview comparison is the place to start, and our candidate screening software page shows what an audited screen looks like in practice.

This is general information about US hiring regulation, not legal advice. Rules change and the contested edges of the AEDT definition are genuinely contested. Talk to employment counsel about your specific tools and jurisdictions.