Automated Employment Decision Tools: Which of Your HR Tools Count

If you lead HR at a company of a few thousand people, you probably own more hiring software than you could list from memory. There’s the applicant tracking system (ATS), which may now rank candidates for you. There’s a sourcing tool that gives each profile a match score. There’s a chatbot that books interviews, a video-interview platform, and an HRIS that has started to flag people who might leave. One day a lawyer forwards you a note about New York City’s Local Law 144. It says any “automated employment decision tool” needs a bias audit. The note doesn’t say which of your tools it means.

That’s the gap this guide tries to close. The law doesn’t care what a vendor calls its product. It cares about what the tool produces and how your team uses that output. So the same feature can be covered in one company and not in the next. Below, we walk the legal test across the tools a typical HR stack holds. We write from the seat of people who build HR agents, so we look at where each tool’s output goes and who reads it.

Everything here is current as of 29 October 2026. It isn’t legal advice, and the rules are moving fast.

What counts as an automated employment decision tool

The term comes from New York City, so that’s where the definition lives. Other places use different words, such as “automated decision system” in California or “high-risk” AI in the EU. But the New York test is the clearest one to learn first, because the others ask similar questions.

The definition in plain words

The law describes a computer process built on machine learning, statistical modeling, data analytics or AI. The process has to produce a simple output, such as a score, a category or a recommendation. That output then has to “substantially assist or replace” a person’s judgment in an employment decision about a real person. That’s the definition in the text of Local Law 144 itself, in our plain words. The law also says what an “employment decision” is. It means screening candidates for a job or screening employees for a promotion within the city. So a tool used only for exits or pay doesn’t meet the New York test. Other rules further down this guide may still reach it.

So there are three parts. The tool has to use data-driven methods. It has to boil things down to a simple output. Finally, that output has to carry real weight in a hiring or promotion decision. Most modern HR tools meet the first two parts without trying. The third part is where the real question sits.

The three ways a tool “substantially assists”

The city’s final rules spell out what “substantially assist” means. A tool meets the test if any one of these is true:

1. Your team relies on the tool’s output alone, and no other factors are considered.
2. Your team gives the output more weight than any other criterion in a set.
3. Your team uses the output to overrule conclusions drawn from other factors, including a person’s own judgment.

Each of these describes a habit, not a feature. A ranking score doesn’t fire a trigger by existing. It fires one when a recruiter stops reading past the top ten, or when a hiring manager trusts the score over the panel. That’s why two companies with the same software can get different answers.

What is left out

The law names some tools that don’t count. The list covers junk email filters, firewalls and antivirus software. It also covers calculators, spreadsheets, databases and data sets. But there’s a catch. These tools are excluded only while they stay out of the decision itself. If one of them starts to support or replace a person’s judgment in a way that materially affects people, it can count. A spreadsheet that just stores interview notes is out. A spreadsheet with a formula that ranks candidates, where you hire from the top, may not be.

Gibson Dunn also notes that tools whose output falls outside the three triggers appear to fall outside the definition. If so, those tools can be used without a bias audit. That’s useful, because it means a lot of your stack may be fine. It just depends on how people use it.

Walking the test across your HR stack

Here is how we’d read the most common HR tools against the three triggers. This table is our judgment as builders, not a legal finding. The “typical use” column describes what we often see, and your team may work differently. The last column is the one that matters, and it nearly always turns on use.

Tool type	What it outputs	How it’s often used	Does a trigger fire?
ATS with candidate ranking	A rank order or fit score per applicant	Recruiters work from the top of the list and rarely open the rest	Likely, if the rank decides who gets reviewed (triggers 1 or 2)
AI interview agent with scoring	A score or pass/fail per interview	Used to decide who moves to the next round	Likely, if the score outweighs other inputs or overrides the panel
Sourcing tool with match scores	A match score per profile	Recruiters pick who to contact, often by score	Possibly, if contact lists are built from the score alone
Scheduling-only chatbot	A booked time slot	Books interviews for people already chosen	Unlikely, because it makes no judgment about the person
HRIS attrition or promotion flag	A risk flag or readiness score	Shown to managers before reviews or promotion talks	Possibly, if a promotion decision leans on the flag; exits fall outside the NYC test
Skills inference or matching module	Inferred skills or a match to a role	Used to shortlist people for internal moves	Possibly, if the shortlist is the only path to the role
Video-interview platform	Recordings, sometimes with a score	Hiring managers watch clips, or read a score	Unlikely for recordings alone; likely if a score drives the decision
Calculator or spreadsheet	Numbers you entered	Stores notes or adds up panel ratings	Unlikely, unless a formula ranks people and you hire from it

A pattern runs through the table. Tools that only move information around, like a scheduler or a recording, rarely count. Tools that produce a judgment about a person, like a score or a flag, can count. Whether a tool does count comes down to one question: does a person weigh that judgment alongside other evidence, or does the judgment decide?

The same feature, in and out

The following example is a composite made up for this guide. The people and figures are invented, but the situations are ones HR teams will recognize.

Dana runs talent acquisition at a 1,800-person logistics company with an office in New York. The company’s ATS gives every applicant a fit score from 1 to 100. A busy warehouse supervisor role draws 600 applicants in a week. The team is stretched, so the recruiters phone the top 40 by score and never open the other 560. So the score is the only thing deciding who gets a call. That looks like trigger one, because nothing else is considered at that step.

Now take the same ATS and the same score at a second company. There, the recruiters first filter for the hard requirements, such as a forklift license and shift availability. After that they read every application that passes. The score sits beside the résumé, and each recruiter writes a short note on why they moved someone forward. The score is one input among several, and it isn’t weighted above the others. On the three-trigger reading, that use may fall outside the definition. Counsel should still confirm it, but the facts counsel would look at are very different from the first company’s.

The second pair is about interviews. Say Dana’s company adds an AI interview agent that asks screening questions and returns a score. A hiring manager sees a candidate the panel liked but the agent scored low. If the manager rejects the candidate because of the score, the tool has overruled human judgment. That’s trigger three. Now say the manager reads the transcript instead. They see the low score came from a poor phone line, so they move the candidate on. In that case the score informed the decision without deciding it.

What we change when we build

This is where the builder’s view helps. When we build an HR agent on a client’s own systems, a few design choices affect how its output gets used. None of them makes a tool exempt by itself. But they make it easier for a person to do the reading the law expects.

First, we’d rather show the evidence than a single number. A recruiter who sees “strong on shift planning, no forklift license listed” has something to weigh. A recruiter who sees “62” has a number to sort by, and sorting is how trigger two happens. Second, we keep every application in front of a person, not only the top of the queue. The Application Screening Support Agent on our HR Agent Map orders the pool for review. Even so, the recruiter sees the full pool, and the agent never rejects anyone on its own. Third, we log every summary and every ordering the agent produces, and a recruiter can overturn any of them. If you ever need to show that a person weighed several factors, that log is where the proof starts.

These choices don’t take our own agent out of scope. It still orders the pool, so in New York City we treat it as an automated employment decision tool. That means it needs a bias audit and candidate notice before its first day of use, not after. Good design changes how people use the output, but it doesn’t change what the law calls the tool.

My advice is to treat these as habits you can check, not features you can buy. A vendor can add a “human review” button. But only your team’s daily practice decides whether the person who clicks it has actually read anything.

What happens once a tool counts

If a tool meets the test and you use it for decisions in New York City, the law asks for a few things. The law’s text sets them out, and the city’s consumer and worker protection agency (DCWP) lists them on its own page too. There are three:

• A bias audit of the tool no more than one year before you use it.
• Notice to candidates or employees who live in the city at least 10 business days before you use it.
• A summary of the latest audit results and the tool’s distribution date. It goes on your website before you use the tool.

The city began enforcing the law on 5 July 2023. The fines are set in the law’s penalty section. A first violation costs up to $500. Each later violation costs between $500 and $1,500. Each day a tool is used without a current audit counts as a separate violation, so the bill grows with every day the tool stays in use. Our guide to NYC Local Law 144 covers the audit math and the posting rules in more detail.

There’s one more point that’s easy to miss. Even if a tool falls outside the New York definition, your general exposure under anti-discrimination law doesn’t go away. Littler’s summary of Colorado’s new law makes the same point. It notes that the new law leans on existing state and federal anti-discrimination law. A biased screening tool is a problem whether or not an AI statute names it.

Beyond New York City

New York was first, but it isn’t alone. Here’s where the main rules stand as of 29 October 2026. Several of these dates have already moved once, so check each one before you plan around it.

Jurisdiction	Instrument	Status or date	What an employer does
New York City	Local Law 144	Enforced since 5 July 2023	Bias audit within a year before use, notice 10 business days ahead, published summary
California	Civil Rights Council rules on automated-decision systems under FEHA	In effect since 1 October 2025	Keep automated-decision data for at least four years
Illinois	HB 3773, amending the Human Rights Act	In effect since 1 January 2026; proposed rules temporarily withdrawn 2 June 2026	Give notice when AI is used in employment decisions; don’t use zip codes as a stand-in for protected groups
Colorado	SB 26-189, which replaced SB 24-205	Signed 14 May 2026; takes effect 1 January 2027; attorney-general rulemaking pending	Give notice before use; run an adverse-outcome process with a right to correct and meaningful human review; keep records for three years
European Union	AI Act, Annex III high-risk list	Hiring AI is high-risk; under the AI Omnibus the Annex III rules apply from 2 December 2027	Prepare for the high-risk duties before 2 December 2027

Two details sit behind the Colorado row. A federal magistrate judge paused enforcement of the old law on 27 April 2026, according to Squire Patton Boggs. The new law is also narrower. Littler and Norton Rose Fulbright both report that it drops the impact assessments and the risk-management program. Norton Rose adds that the old duty of care is gone too. Only the attorney general can enforce the new law.

The EU row rests on the European Commission’s own notice. It says the AI Omnibus entered into force on 27 July 2026. It also says the high-risk rules for Annex III systems now apply from 2 December 2027. Hiring and recruitment are on that list, so that’s the date to plan around.

The California row comes from the state’s Civil Rights Department. The Illinois row comes from Covington’s round-up of these laws. Illinois also reaches further than New York, because its notice duty covers decisions such as discharge and training as well as hiring.

How strong is enforcement?

You might wonder whether any of this gets enforced. In New York the honest answer so far is “not much”. The New York State Comptroller’s audit, published 2 December 2025, covered July 2023 to June 2025. It called the city’s enforcement “ineffective”. The city received only two complaints about these tools in that time. When auditors made test calls to the 311 line, 75% were sent to the wrong place. The city reviewed 32 published bias audits and flagged one issue, while the Comptroller found at least 17 potential issues in the same set.

But weak enforcement in the past doesn’t mean low risk now. The city agreed to adopt most of the Comptroller’s recommendations. DLA Piper reads the audit as a sign of rising risk for employers. So I’d plan as if the rules will be enforced, because enforcement is moving that way.

What to do this quarter

You don’t need to settle every legal question this quarter. You need to know what you have and how it’s used. That’s an inventory, and HR can run it without waiting for legal.

1. List every tool that touches a hiring, promotion or exit decision. Include the small ones, like a sourcing extension a recruiter installed or a spreadsheet with a ranking formula. Ask each team what they actually use, because it won’t all be on the software contract list.
2. Name one owner for each tool. This is the person who can answer how it’s used day to day. It’s often a recruiting lead or an HR business partner, not the person who signed the contract.
3. Write down what each tool outputs. It might be a booked slot or a recording. It might also be a summary, a score or a flag. Tools that produce a judgment about a person go to the top of your list.
4. Record who reads the output and what else they look at. This is the step that answers the three triggers. Mark a tool for legal review if its output is the only input or the heaviest one. Do the same if people let it override their own view.
5. Note where each tool is used. A tool used for New York City roles faces different rules from one used only in Illinois or the EU.

Once you have the list, the next decisions get much easier. Some tools will be clearly fine. A few will need a bias audit or a change in how your team uses them. Our page on HR agents for talent acquisition shows where an agent fits in each step of hiring. The HR Agent Map lays out the wider set of HR tasks an agent can take on, so you can see which ones produce a judgment about a person and which don’t.

The limits of this guide

This is not legal advice. It’s a builder’s reading of how the rules meet the tools in an HR stack, and every rule above is dated as of 29 October 2026. Several of them are still changing, including Colorado’s rulemaking and Illinois’s withdrawn rules. The definition and triggers here are our plain-words reading of New York’s law and rules. Read the text itself before you quote the exact wording in a policy. Before you decide that any tool is in or out, check with your employment counsel.

Talk it through with us

If your inventory turns up tools that score or flag people, you have a choice to make. You can change how your team uses them, or you can rebuild the step so a person reads the evidence and the system records who decided. We build HR agents on your own systems, so the output lands where your team already works and the decision record stays with you. Book a 30-minute call and bring your tool list. We’ll go through it with you against the HR Agent Map and show you where an agent would sit relative to the three triggers.