What Is an Automated Employment Decision Tool (AEDT)?

The Renewal Notice That Started This Question

The email from your applicant tracking vendor lands on a Tuesday. Subject line: "Exciting new AI-powered screening features now included in your plan." Somewhere in the feature list is a phrase like "smart candidate ranking" or "predictive fit scoring." You forward it to yourself with a note: is this a problem?

That question — is this a problem — is really a narrower one: does this tool meet the legal definition of an automated employment decision tool under New York City's Local Law 144? Because if it does, you're now on the clock for an annual independent bias audit, a public results summary, and advance notice to candidates. If it doesn't, none of that applies, and you can move on.

Most HR teams never get a clean answer to that question. They guess, based on how "AI" the tool sounds in the marketing copy, and either over-comply on tools that don't need it or under-comply on tools that do. Neither is a good use of your time.

By the end of this article, you'll be able to read any hiring tool's function against the actual statutory language and reach a defensible answer of your own.

What Is an Automated Employment Decision Tool, Under LL144?

Start with the text, not the marketing. Local Law 144 defines an automated employment decision tool as any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues a simplified output — a score, classification, or recommendation — that is used to substantially assist or replace discretionary decision-making for employment decisions (Perkins Coie, 2023).

That's four separate conditions, and a tool has to clear all four to be an AEDT:

1. It runs on a computational process — not a person's judgment alone.
2. That process is built on machine learning, statistical modeling, data analytics, or AI — not a static rule someone wrote by hand with no learning or modeling behind it.
3. It produces a simplified output: a score, a classification, or a recommendation. Not raw data. A conclusion.
4. That output is used to substantially assist or replace a human's discretion in an employment decision — hiring, promotion, or similar.

The law took effect January 1, 2023, but enforcement didn't begin until July 5, 2023, after an initial April 15, 2023 start date was pushed back to give employers and vendors more time to prepare (Epstein Becker Green / Workforce Bulletin, 2023). If a tool clears all four conditions above and is used on candidates or employees who work in New York City, three obligations follow: an annual independent bias audit, a public summary of the most recent audit results posted to the employer's website along with the AEDT's distribution date, and notice to candidates or employees at least 10 business days before the tool is used, including a way to request an alternative process or accommodation (Crowell & Moring LLP, 2023; Epstein Becker Green, 2023).

"Any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues a simplified output... that is used to substantially assist or replace discretionary decision-making." (Perkins Coie, 2023)

That's the whole test. Everything else — the tool-by-tool judgment calls, the vendor conversations, the documentation — comes back to whether a specific feature satisfies that sentence.

Breaking Down "Substantially Assist or Replace Discretionary Decision-Making"

The phrase doing the most work in that definition is "substantially assist or replace." It's also the phrase with the least amount of bright-line guidance attached to it, which is exactly where most classification disputes live.

Here's the practical read. If a tool produces a ranked list, a percentile score, a pass/fail flag, or a "top match" recommendation, and a recruiter or hiring manager is expected to rely on that output to decide who moves forward — even if they technically retain the ability to override it — that's substantial assistance. The presence of a human "in the loop" does not automatically take a tool out of scope. What matters is whether the tool's output is functionally driving the decision, not whether a human's name is on the final approval.

Compare that to a tool that simply organizes information for a human to review with no scoring, ranking, or recommendation attached — a resume parser that extracts years of experience into a spreadsheet column, for instance, with no computed fit score layered on top. That's closer to assistive infrastructure than a decision tool, because there's no simplified output steering the outcome.

This is the exact judgment call that trips people up, and it's also exactly where a documented rationale you can defend matters more than a confident guess. You don't need to be right by accident. You need to be able to show your reasoning if DCWP, a candidate, or your own legal counsel ever asks how you reached your conclusion.

None of this is legal advice, and it isn't a substitute for it. Whether a specific feature in your stack meets the statutory definition is a fact-specific determination. If you're not certain, that's the moment to verify with DCWP's published guidance or with outside counsel — not to guess and move on.

Resume Screeners, Video Scorers, and Ranking Algorithms: Tool by Tool

Applying the four-part test to real categories of hiring software:

Resume/applicant screening software. If the tool uses machine learning or statistical modeling to compute a match score, a fit percentage, or a ranked order of candidates, and recruiters use that score to decide who advances, it's very likely an AEDT. If the tool only does keyword matching against a fixed list a human wrote — no learning, no statistical model, no computed score — the "derived from machine learning, statistical modeling, data analytics, or AI" prong is harder to satisfy, and the analysis gets more complicated. This is a case-by-case read, not a category-wide default in either direction.

Video interview scoring. Tools that analyze video or audio responses and generate a competency score, a personality profile, or a recommendation ranking are squarely aimed at the kind of simplified output the law describes. If a hiring manager relies on that score to decide who continues, this is one of the clearer AEDT cases.

Applicant ranking algorithms. By design, these tools produce exactly the output the statute names — a ranked list functioning as a recommendation. If that ranking substantially assists the decision of who gets contacted or interviewed first, treat it as in scope by default and confirm otherwise, not the reverse.

Skills or personality assessments with automated scoring. If the platform computes a score or a fit classification from candidate responses using a model, it fits the same pattern as video scoring above.

Chatbots that only schedule interviews or answer FAQs. These typically don't issue a score, classification, or recommendation about a candidate — they route logistics. Absent a scoring layer, this is a weaker case for AEDT status, though it's worth checking whether the vendor has quietly added a "fit" or "engagement" score to the same tool.

What Falls Outside the AEDT Definition

The definition has edges, and it's worth being precise about where they sit rather than assuming every piece of "AI-powered" marketing copy describes an AEDT.

A tool is not automatically an AEDT just because a vendor calls it "AI-powered." Marketing language is not statutory language. The test is the four conditions above — computational process, ML/statistical/data-analytics/AI basis, simplified output, and substantial assistance or replacement of discretion — not the adjectives on a pricing page.

A tool that only stores, organizes, or displays candidate data — with no computed score, classification, or recommendation attached — sits outside the definition, because it never issues the simplified output the statute requires.

A process where a human makes the actual decision using their own independent judgment, and any software output is informational rather than determinative, is a harder case to classify as substantially assisting or replacing discretion — though "informational" is doing a lot of work in that sentence, and it's worth scrutinizing whether that's really how the tool is used in practice versus how it's described.

This is genuinely gray territory in places, and treating it as gray — rather than picking a comfortable answer and moving on — is the more defensible posture. That's the posture this site's Local Law 144 compliance guide and the deeper walkthrough on whether your specific hiring tool is an AEDT are both built around: work the definition against the tool, don't work backward from what's convenient.

Why Classification Comes Before Every Other LL144 Step

Every other LL144 obligation is downstream of this one question. You cannot schedule a bias audit for a tool you haven't confirmed is in scope. You cannot post a public audit summary for a tool that was never audited because nobody classified it. You cannot send candidates 10 business days of notice for a tool you didn't know needed disclosure.

That ordering has a real cost when it's skipped. New York City's own comptroller reviewed enforcement of Local Law 144 for the period from July 2023 through June 2025 and found it largely ineffective — the city's own agency found only 1 of 32 companies reviewed non-compliant, while the comptroller's auditors, reviewing the same companies, found 17 (Office of the NY State Comptroller, 2025). A separate academic review of 391 employers found only 18 had posted audit reports and only 13 had posted the required transparency notices (ACM FAccT, Wright & Muenster et al., 2024). Weak enforcement doesn't make the underlying obligation optional — it makes an accurate self-classification more important, not less, because you're largely operating on the honor system right now.

If you get the classification wrong in the direction of "not an AEDT" and you're incorrect, the exposure is real: civil penalties run up to $500 for a first violation, then $500 to $1,500 for each subsequent violation, and those penalties accrue per violation, per day (Office of the NY State Comptroller, 2025). Getting the call right, once, and being able to show your reasoning, is a much better position than reaching a comfortable answer under deadline pressure.

Document the Call, Then Act on It

You don't have to make this determination from memory every time a vendor pushes a new feature. Run each tool through the same four-part test, in writing, and keep the record — what the tool does, what output it produces, who relies on that output and how, and which prong of the statute it does or doesn't satisfy. That written record is what "substantially assist or replace discretionary decision-making" looks like when someone asks you to defend it later, instead of re-litigating it from scratch.

The "Is This Tool an AEDT?" Classification Decision Workbook is built to run exactly that process across every tool in your hiring stack — a structured worksheet that walks each tool through the statutory test and produces a documented, defensible determination you can point to later. It doesn't perform a bias audit and it doesn't score any candidate; it only helps you decide, and document, whether a bias audit is required in the first place. Once a tool clears the test, the next step is scheduling the audit itself — covered in the bias audit requirements guide — and an independent auditor, never this workbook, performs and signs that audit.

If you're managing this across more than a handful of tools or more than one entity, see current pricing for the full workbook set. And if you're waiting on the always-on version of this — a live tool where classifications, audit deadlines, and vendor status update automatically — that product isn't built yet; join the waitlist to hear when it ships.