NYC's AI Hiring Law Skips Sourcing. The First Scorecard Doesn't
Local Law 144 exempts AI sourcing until a candidate applies and a tool scores them. Where Tandem and Hiring Assistant 2 cross that line, and how to stay clear.
NYC's enforcement agency has told employers that Local Law 144 doesn't cover tools used only to find and contact people. That exemption was easy to stay inside when sourcing tools did nothing but source. In November, Joveo's Tandem and LinkedIn's Hiring Assistant 2 put scoring and pre-screens into the same product as search. Leaving the exempt zone used to take a separate, deliberate step. Now it can take one more prompt.
Does Local Law 144 apply to sourcing?
No, as long as the tool only finds, scans, or contacts people who haven't applied. NYC's Department of Consumer and Worker Protection (DCWP) says the law's obligations don't apply to using a tool to "scan a resume bank, conduct outreach to potential candidates, or invite applications."
Two definitions matter here:
- Local Law 144 is New York City's law on automated employment decision tools used in hiring for NYC roles.
- An AEDT (automated employment decision tool) is a tool whose output substantially assists a hiring decision.
DCWP's reasoning turns on one phrase. A "candidate for employment" is someone who has applied for a specific position. A passive prospect you found on GitHub hasn't applied to anything, so a tool that ranks and messages them is outside the law.
Two caveats. First, the DCWP FAQ is guidance, not statute. Second, the same commentators who cite the exemption add that the rules can apply once those people apply and a tool scores them. Searching is exempt. A score attached to an applicant is not.
What Local Law 144 requires once it applies
Once an AEDT is used on NYC candidates, the employer owes three things, and the vendor isn't liable under the law. Per the rules:
- An independent bias audit of the tool, completed within the past year.
- A public summary of the audit results.
- Notice to candidates at least 10 business days before the tool is used on them.
Penalties run up to $1,500 per violation per day. The part that catches buyers out is that vendors are not liable under Local Law 144 itself. If Joveo or LinkedIn ships a feature that turns your workflow into an AEDT, the audit and notice obligations land on you, the employer. Your vendor can give you documentation, but the compliance obligation stays with you.
The exact step where a sourcing tool becomes an AEDT
A sourcing tool becomes an AEDT when a person applies and the tool's output then substantially assists the decision about them. Where the tool sits in the funnel doesn't matter. What matters is how much you lean on its output.
Under DCWP's rules, a tool "substantially assists" a decision when the employer:
- relies exclusively on the tool's output,
- weighs that output more heavily than any other criterion, or
- uses it to overrule human judgment.
Being early in the funnel won't protect you. A first-round filter that you follow without question meets test one just as well as a final-round score would.
Here is how a typical agent workflow maps against those rules:
| Workflow step | Who it touches | Local Law 144 status |
|---|---|---|
| Agent searches profiles, repos, open web | Passive prospects | Exempt per DCWP FAQ |
| Agent ranks prospects and sends outreach | Passive prospects | Exempt per DCWP FAQ |
| AI pre-screen of sourced, not-yet-applied people | Unclear | Open question for counsel |
| Sourcing score carried into the applicant record | Applicants | Covered if it substantially assists |
| Chat application, recorded AI interview, scorecard | Applicants | Covered if it substantially assists |
If you want to classify each step of your own funnel, the pipeline-step AI hiring read sorts each one into out of scope, notice, rights, or audit.
Tandem: one chat, two legal regimes
Joveo launched Tandem on November 17, 2026. CEO Kshitij Jain calls it "vibe recruiting." At the sourcing stage, Tandem scores each candidate out of 100 and shows the evidence behind the score. That part is exempt, because those people haven't applied.
Then the workflow keeps going inside the same chat. Candidates apply by chatting with an AI assistant, then complete a recorded AI interview, and every candidate gets a scorecard. Those people are applicants now, and a scorecard is exactly the kind of output the substantial-assistance tests look at. I covered the vibe recruiting pitch behind Tandem when it launched. The compliance view is narrower: the line sits in the middle of a single product, and the recruiter never sees it.
Hiring Assistant 2: the pre-screen you can't turn off
LinkedIn's Hiring Assistant 2 adds memory of how you hire, personalized fit signals, identity and employment verification, and voice pre-screening. The pre-screen runs by voice or text, can be extended to sourced candidates, and takes a median of six minutes. Existing English-language customers get the update automatically from November 2026.
The detail that matters most for AI recruiting agent compliance: pre-screening is on automatically for recruiters who hold the Hiring Assistant permission, and it can't be turned off for a single recruiter or for the whole contract. You don't opt in, and you can't opt out.
Why "never rejects" doesn't get you out
"Never rejects" doesn't settle anything, because the law tests whether you rely on the output, weight it, or let it override a person. Rejection isn't one of the tests. A ranked list can meet the weighting test without ever turning anyone down.
Both vendors describe their guardrail in rejection terms. Joveo says Tandem never rejects anyone. Hiring Assistant 2 is described as making no automatic decisions and not screening out applicants. Neither statement says anything about ranking.
Ranking is the problem. Say a recruiter has 200 applicants and works only the agent's top 20. No one was rejected by the tool. But in practice the tool's score outweighed every other criterion, because the other 180 never got a human look. That is the "weighted more heavily" test, met by habit rather than by settings.
The trigger isn't the AI. It's the application, and the score that follows it.
The sourcer layer is thin, so the handoff became a feature
Agents are taking over a small specialist job, and they hand off straight to screening, with no person in between. In Refolk's index, US profiles with sourcer titles are outnumbered roughly 22 to 1 by recruiter and TA-partner titles.
| Segment | Profiles |
|---|---|
| US sourcer titles (Sourcer, Talent Sourcer, Sourcing Specialist, Sourcing Recruiter) | 5,477 |
| UK sourcer titles (same filters) | 461 |
| US recruiter titles (Recruiter, Technical Recruiter, TA Partner) | 122,598 |
| US recruiters per US sourcer | ~22.4 |
| US vs UK sourcer pool | ~11.9x |
Treat the sourcer counts as upper bounds. "Sourcing Specialist" also picks up procurement roles, and Senior Buyer and Procurement Specialist profiles showed up in the samples. In a sample of 25 US sourcer profiles, New York City Metro tied for the top region with 3. That sample is too small to prove anything, but it hints that a fair share of sourcers work in the city where this law applies.
Here's why this matters for the law. The sourcer used to be the handoff. A person found prospects, judged them, and passed a shortlist to a recruiter. That judgment happened before anyone applied, and it left a record of human reasoning. When an agent replaces the sourcer and feeds its own ranking into its own screen, the handoff disappears. Nobody makes a cut that the score didn't make first.
The fix is to keep discovery and screening in separate tools, with a person between them. That split is easier when your search tool only searches. Refolk works that way: you describe the person in plain English and get a ranked list of people from GitHub, LinkedIn, and the open web, ready for outreach. It stops at discovery, the part DCWP's FAQ exempts, and it doesn't spill into an applicant scorecard. If you're comparing tools on these lines, the people-search bake-off playbook shows how to score them against reqs you've already filled.
Low enforcement is not low risk
Local Law 144 enforcement has been weak, but the risk is moving to lawsuits and discovery. There, an agent's logs show exactly how much you leaned on its scores.
The enforcement record so far:
- DCWP received only two AEDT complaints between July 2023 and June 2025.
- When DCWP reviewed 32 companies, it found one instance of non-compliance. The New York State Comptroller's auditors, in audit 2024-N-6 (released December 2, 2025), found at least 17.
- 75% of test calls to NYC 311 about AEDT issues were routed to the wrong place.
The study's explanation is the useful part. Most employers decided for themselves that they were out of scope. Bundled agents make that self-assessment more likely to be wrong. An employer that bought a "sourcing tool" in 2024 may now be running chat applications and scorecards through the same login, and the original scoping memo still says "sourcing only."
If a rejected candidate sues, the agent's records are evidence: who was scored, what score they got, and whether anyone below the top of the list was ever opened. A gap between what your notice says and what your logs show is the kind of evidence a plaintiff's lawyer looks for.
A workflow where AI finds and a named human makes the cut
Let the agent find and contact people, and have a named person make the shortlist decision, with written reasons, before any AI-scored screen. This list isn't legal advice, but it's the AI candidate screening setup in New York I'd defend:
- Automate discovery fully. Search, rank, and outreach to people who haven't applied all fall in the DCWP exemption. This is where agents save the most time.
- Wall off the sourcing score at application. When a prospect applies, don't carry the 0-to-100 sourcing score into the applicant record. A sourcing score reused after application is the easiest way to slip into AEDT territory without noticing.
- Make the human cut before the AI screen. A named recruiter reviews applicants and writes a sentence or two on each one advanced or passed. A "human approves" checkbox won't pass the reliance test. Notes that show independent judgment might.
- Treat any scored screen on applicants as an AEDT. If you use Tandem's AI interview scorecards or Hiring Assistant 2's pre-screen on applicants for NYC roles, get the independent audit, publish the summary, and send notice 10 business days before use.
- Take the sourced-candidate pre-screen to counsel. Hiring Assistant 2 can pre-screen people you sourced, and you can't switch that off. Whether a sourced person becomes a "candidate" once they're screened is an open legal question, not settled law.
- Keep the records on purpose. Logs can protect you or hurt you depending on what they show. Set a policy before you need one, using something like the candidate retention standard to decide what to hold and what to delete.
The split also makes sourcing better. A tool like Refolk that returns people with the evidence behind each match gives the recruiter something to reason about, so the human cut becomes a real judgment and not a rubber stamp on someone else's score.
FAQ
Is the DCWP sourcing exemption legally binding?
No. It comes from DCWP's FAQ, which is agency guidance, not the text of Local Law 144. It reflects how the enforcing agency reads the law, which matters in practice, but a court could read "candidate" or "substantially assists" differently. Rely on it for sourcing that ends before anyone applies, and don't stretch it past that point.
If my vendor runs a bias audit, am I covered?
Not automatically. Vendors are not liable under Local Law 144 itself, so the obligation to have an independent audit from the past year, publish a summary, and give 10 business days' notice sits with the employer. Ask vendors for their audit documentation, but have counsel check that it matches how you actually use the tool.
Does having a human approve the AI's shortlist make the tool exempt?
Only if the human's judgment is real. A tool still substantially assists if you rely on it exclusively, weigh it above every other criterion, or use it to overrule people. A recruiter who approves the agent's top 20 without looking at anyone else probably meets the second test. Written reasons for each advance or pass are your best evidence that a person actually decided.
Does LinkedIn Hiring Assistant 2's voice pre-screen count as an AEDT?
Used on applicants for NYC roles, where its output substantially assists the decision, it very likely falls in scope. Used on sourced people who haven't applied, it's a gray zone that counsel should rule on. Because the feature is on by default for recruiters with the Hiring Assistant permission and can't be disabled, assume it's running and document how you use its output.
The data behind this
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.