# The Pipeline-Step AI Hiring Read: Out of Scope, Notice, Rights, or Audit

*Score any sourcing or screening step on five dimensions and assign Out of Scope, Notice Only, Notice plus Rights, or Audit Required, with the obligations and evidence to file.*

- Canonical URL: https://www.refolk.ai/guides/ai-hiring-pipeline-step-classification
- Pillar: Process, data, and compliance
- Format: Framework
- Published: 2026-11-15
- Last reviewed: 2026-11-15
- Reading time: 22 min
- Keywords: does NYC Local Law 144 apply to sourcing tools, AI candidate ranking tool compliance checklist, CCPA ADMT pre-use notice for job applicants, Colorado AI hiring law human review exemption, is an AI sourcing tool high-risk under EU AI Act, AI recruiting tool inventory and risk classification

## Key takeaways

- Classify each pipeline step separately, not the vendor: the step's verdict is the highest tier any of the five regimes reaches, and its obligation list combines the duties of every regime that applies.
- NYC Local Law 144's bias audit and notice duties do not apply to people who have not applied for a specific position, but that sourcing carve-out exists only in NYC: Colorado reaches anyone whose opportunity in Colorado is evaluated, and the EU AI Act names targeted job ads.
- A recruiter's sign-off can take a ranking step outside California's ADMT rules if the reviewer is trained, actually reviews the output, and can change the decision, but Colorado's draft-rule presumption keeps the same step in scope.
- An NYC scoring-rate bias audit measures the share of each category scoring above the sample median, so moving a pass/fail cutoff changes nothing in the audit table and a filter disguised as a score goes unseen.
- Retention across the regimes runs from at least 6 months (EU deployer logs) to 4 years (California FEHA automated-decision records), so keeping every step's evidence for 4 years is the only setting that fails none of them.
- In Refolk's index there are 160 US recruiting-ops profiles against 26 US AI-governance profiles, about 6.2 to 1, so step classification will usually land on recruiting ops rather than a specialist.

This is a classification standard for each automated step in a sourcing and screening pipeline: search, rank, score, auto-filter and drafted outreach. It is for recruiting-ops and compliance leads who must decide which AI-in-hiring rules a step triggers, and what has to happen before that step touches a person. You score the step on five dimensions, read off one of four verdicts - Out of Scope, Notice Only, Notice plus Rights or Audit Required - and file the obligations and evidence alongside it, so that a second reviewer reaches the same answer. It is a method for preparing a classification you can hand to counsel. It does not replace counsel.

## What does a step read classify?

A step read classifies one automated step against five AI-in-hiring regimes. A step means one feature producing one output at one point in the funnel. It does not classify a vendor or a product, because one product usually contains several steps with different verdicts.

The five regimes are:

- **NYC Local Law 144 (LL144)**, which regulates an "automated employment decision tool" (AEDT): a tool used to screen people for employment decisions in New York City.
- **California's CPPA rules on automated decisionmaking technology (ADMT)**, meaning technology whose output replaces or substantially replaces a human decision.
- **Colorado SB 26-189**, which covers automated systems whose output materially influences a consequential decision such as hiring.
- **Illinois HB 3773**, which covers the use of AI in recruitment and employment decisions.
- **The EU AI Act, Annex III point 4(a)**, which makes recruitment and selection AI "high-risk". It places duties on the **deployer**: the organisation that uses the system, as opposed to the vendor that built it.

Two more terms carry the whole framework. A **sourced person** is someone your pipeline found who has never applied. An **applicant** is someone who has applied for a specific position. Most sourcing tools touch both, often inside a single product.

### The four verdicts

| Verdict | What it means | Typical source |
|---|---|---|
| Out of Scope | No regime attaches | Organize-only, no IL or EU link |
| Notice Only | Tell people AI is used | Illinois |
| Notice plus Rights | Notice plus individual rights | California, Colorado |
| Audit Required | A standing evaluation regime | NYC bias audit, EU deployer duties |

The tiers are my own grouping of duties, ordered by how much work each one creates. They are not a legal category. Each regime keeps its own obligation list, and you file that list with the verdict. Out of Scope still produces a record: the record is your proof that you looked.

## Where does each regime draw its lines?

The regimes disagree on three lines: whether people who never applied are covered, what kind of output triggers coverage, and whether a human reviewer takes a step out of scope. A step can land on opposite sides of the same line in two jurisdictions.

| Regime | Never-applied people | Trigger | Human review exempts? |
|---|---|---|---|
| NYC LL144 | Out (per FAQ) | Substantially assists screening | No |
| CA ADMT | Not established | Replaces or substantially replaces | Yes, if 3-part test met |
| CO SB 26-189 | Possibly in | Non-de minimis factor | No (draft-rule presumption) |
| IL HB 3773 | Not established | Any AI use | No threshold |
| EU 4(a) | Targeted ads are in | Place, filter, evaluate | No (oversight is a duty) |

The dates matter because they decide which duties you must build now and which you can stage:

| Regime | Applies from | Status note |
|---|---|---|
| NYC LL144 | 5 Jul 2023 | Enforced by DCWP |
| CA ADMT | Operative 1 Jan 2026 | Existing uses compliant by 1 Jan 2027; later uses from day one |
| CO SB 26-189 | 1 Jan 2027 | AG not enforcing until rulemaking concludes |
| IL HB 3773 | 1 Jan 2026 | Rules withdrawn; statute unchanged |
| EU Annex III | 2 Dec 2027 | Stand-alone systems, per Regulation 2026/1744 |

Check these dates against the primary sources in the last section before you file a verdict. Delays and rule withdrawals have happened in this area, and a classification built on a stale date is wrong in a way nobody notices.

### Three consequences of the table

**"Sourcing is out of scope" is a New York City position only.** It rests on the 6 RCNY 5-300 definition: a candidate is someone who has "applied for a specific employment position by submitting the necessary information or items in the format required by the employer or employment agency." The DCWP FAQ says: "If AEDT use is to assess someone who is not an employee being considered for promotion and who has not applied for a specific position for employment, the bias audit and notice requirements do not apply." The FAQ is not law. Colorado's definition of consumer reaches "any individual whose access to, eligibility for, or opportunity in Colorado is evaluated". EU point 4(a) names targeted job advertisements, which by their nature reach people before they apply.

**California keys on replacement, Colorado keys on influence.** In California, a ranked list that a qualified recruiter reviews and can override is not ADMT, because a human made the decision. Colorado's draft rules presume that an output the decision-maker reviewed, and which matches the outcome, materially influenced the decision. So the same list is out in one state and in the other.

**Illinois has no threshold and no settled reach.** Any use of AI counts. However, nobody has publicly established whether "recruitment" covers people who never applied, and sources disagree on who is owed notice: one analysis says applicants and employees, another says the text says only "employee".

## The five dimensions, and how to score each

Score every step on population, output type, decision weight, human override and location. Each dimension has a small, fixed set of values, so two reviewers have nothing to interpret differently.

### 1. Population: applied or never applied

Values: **PRE** (only never-applied people), **POST** (only applicants), **BOTH**. A step whose boundary is not logged is scored BOTH.

What it proves: whether LL144 can attach at all, and whether Colorado and the EU are on notice for a sourcing step. What a mislabel looks like: a record still tagged "sourced" after the person has clicked apply. A search run in [Refolk](/) across GitHub, LinkedIn and the open web returns people who never applied, so that search step scores PRE. Once any of those people applies, every re-rank that touches them is POST, even if the record still carries its old tag.

#### Where a sourced person becomes an applicant

1. **Sourced** - Found by search, never applied, LL144 does not attach
2. **Contacted** - Outreach invites them to apply, still pre-application
3. **Apply event** - They submit the items in the format the employer requires
4. **Applicant** - ATS logs an application timestamp, post-application from here
5. **Re-rank** - Any score after that timestamp is a post-application step

*A re-rank's verdict changes the moment the apply event lands, whatever label the record still carries.*

### 2. Output type

Values: **ORGANIZE** (summarize or organize for human review), **RANK**, **SCORE**, **CLASSIFY**, **FILTER** (removes or advances people automatically). Check each one against Colorado's verbs - "constraining, ranking, scoring, recommending, classifying" - and against the Annex III verbs: place targeted ads, filter applications, evaluate candidates.

What it proves: Colorado excludes tools "used solely to summarize or organize information for human review", so ORGANIZE is the only type that can escape Colorado. What a mislabel looks like:

- A "summary" that labels people "strong" or "weak" is a classification. Score it CLASSIFY.
- A "score" with a threshold that auto-advances or auto-rejects is a filter. Score it FILTER, whatever the vendor calls it.
- Drafted outreach is not named in any of the verb lists. What matters is who chooses the recipients. If the step also picks who gets the message, score that choice as RANK or FILTER. This is my reading, not settled law. Mark it ASSUMED.

### 3. Decision weight

Values: **NONE** (the output is not consulted for the decision), **FACTOR** (a non-de minimis input among others), **DECIDES** (the output becomes the decision with no human step in between).

What it proves: Colorado's first prong, which asks whether the output is a non-de minimis factor in the decision, and California's replacement test. What a mislabel looks like: an "advisory" score that sets the order of the queue a recruiter works from the top down. That output is constraining and ranking. Score it FACTOR at minimum. Be aware that Colorado's draft rules leave "materially influence" itself undefined, which is why FACTOR is the safe default for anything a decision-maker sees.

### 4. Human override

Values: **PASS** or **FAIL** against California's three-part test. To score PASS, all three parts must be documented for the reviewer:

1. They know how to interpret and use the output.
2. They actually review and analyze the output and any other relevant information.
3. They have authority to change the decision.

What it proves: whether California's rules attach. It proves nothing for NYC, where the scope test applies regardless of review, or for Colorado, where review triggers the presumption rather than an exemption. In the EU, human oversight is itself a deployer duty, not an exit. What a mislabel looks like: a log with zero overrides. A 0% override rate points to rubber-stamping, and a rubber stamp fails part 2.

#### Decision weight against human review

Horizontal axis runs from Review fails the three-part test to Review passes the three-part test. Vertical axis runs from Output only organizes for review to Output weighs in the decision.

| Quadrant | What it means |
| --- | --- |
| Weighs, no real review | In scope for CA and CO, Notice plus Rights at least |
| Weighs, real review | Out of CA ADMT, still in CO under the presumption |
| Organizes, no real review | Check for hidden labels or thresholds before trusting the tag |
| Organizes, real review | Likely outside CA and CO, but Illinois and EU may still apply |

*California keys on replacement and Colorado keys on influence, so the same ranked list can sit inside one regime and outside the other.*

### 5. Location

For each of NYC, California, Colorado, Illinois and the EU, record the job location and the candidate's residence as separate tags. Start with the job: the published LL144 scope test puts job location first.

What it proves: which regimes can attach at all. What a mislabel looks like: keying NYC only to where the candidate lives. LL144 scope turns on where the job is. Sources describe the 10-business-day notice as going to candidates who reside in the city, so record both tags and have counsel confirm how the two interact for remote roles. For Colorado, a job based in Colorado brings in non-resident sourced people whose opportunity there is evaluated.

## How do scores map to the four verdicts?

Check each regime's trigger against the five scores. The step's verdict is the highest tier any regime reaches, and its obligation list combines the duties of every regime that applies. Never average, and never let one regime's exemption cancel another regime's duty.

| Regime | Applies when | Tier |
|---|---|---|
| NYC LL144 | NYC job, POST or BOTH, not ORGANIZE, weight not NONE | Audit Required |
| EU 4(a) | EU link, used to target ads, filter or evaluate | Audit Required |
| CO SB 26-189 | CO link, not ORGANIZE, weight FACTOR or DECIDES | Notice plus Rights |
| CA ADMT | CA link, weight not NONE, override FAIL | Notice plus Rights |
| IL HB 3773 | IL link, any AI use | Notice Only |

The NYC trigger approximates the "substantially assist" definition in 6 RCNY 5-300. Have counsel read that definition against any step you score FACTOR. If no regime applies, the verdict is Out of Scope.

### The obligations behind each tier

- **NYC LL144 (Audit Required):** a bias audit within one year before use, then annually. Publish the audit summary. Give candidate notice at least 10 business days before use.
- **EU deployer (Audit Required):** follow the instructions for use, assign human oversight, check input data, monitor operation, keep logs for at least six months, and inform workers' representatives and the affected workers before using the system at the workplace.
- **Colorado (Notice plus Rights):** a pre-use notice, a post-adverse-outcome notice within 30 days, rights to access and correct personal data and to request meaningful human review, and records kept for at least three years after the decision.
- **California (Notice plus Rights):** a pre-use notice, an opt-out or human appeal, and answers to access requests. I could not retrieve the hiring-specific opt-out exceptions in section 7221. Treat them as unverified until counsel reads them.
- **Illinois (Notice Only):** notice that AI is used. Who receives it, and in what form, is not settled. See the failure modes.

> **Rule:** Unresolved reach resolves to notice, not silence
>
> Where a regime's reach is not publicly established, score the step as in scope for that regime's notice duty, mark it ASSUMED with a date and a name, and re-read it when the regime's rules are published. Never let "not established" become "out of scope" without a written reason.

**Step classification record**

```
STEP ID:
Vendor / feature:
Funnel position:
Output (exact form, e.g. ranked list, 1-100 score, pass/fail):

D1 Population: PRE | POST | BOTH
   Apply trigger event logged: yes / no (where)
D2 Output type: ORGANIZE | RANK | SCORE | CLASSIFY | FILTER
   Labels or thresholds found in sample of __ outputs: yes / no
D3 Decision weight: NONE | FACTOR | DECIDES
D4 Human override (CA three-part test): PASS | FAIL
   (1) Reviewer trained to interpret output: evidence
   (2) Reviewer reviews output plus other information: evidence
   (3) Reviewer has authority to change decision: evidence
   Logged override rate over review window: __%
D5 Location (job / candidate): NYC __/__  CA __/__  CO __/__  IL __/__  EU __/__

Regimes that apply: NYC | CA | CO | IL | EU   (mark ASSUMED where applicable)
VERDICT: Out of Scope | Notice Only | Notice plus Rights | Audit Required
Obligations (all applicable regimes combined):
Evidence to keep, and retention period:
Reviewer 1 / date:
Reviewer 2 / date / same verdict: yes / no
Next re-read trigger:
```

*One record per step. Copy it into your inventory sheet or ticket, fill every field, and write ASSUMED wherever reach is not established.*

## What does Audit Required involve under LL144?

An LL144 audit reports an impact ratio for each demographic category. It uses selection rates for steps that pass or fail people, and scoring rates for steps that score them. The first audit falls within one year before use, and the audit is repeated every year after that. Most of the classification errors in this framework show up here.

The impact ratio is "either (1) the selection rate for a category divided by the selection rate of the most selected category or (2) the scoring rate for a category divided by the scoring rate for the highest scoring category." The scoring rate is "the rate at which individuals in a category receive a score above the sample's median score."

That median has a consequence most teams miss. Because a scoring-rate audit counts who lands above the median, moving a pass/fail cutoff changes nothing in the audit table. A filter that is labelled a score will pass a scoring-rate audit without the table ever showing the filtering. That is why output type is scored by what happens downstream, and why any FILTER step should be audited on selection rate.

Four more rules shape the audit:

- **No pass mark.** LL144 does not set a threshold for the ratio. Teams often compare against the EEOC four-fifths rule in 29 CFR 1607.4(D). I have not verified that text for this guide, so read it at [eCFR, 29 CFR Part 1607](https://www.ecfr.gov/current/title-29/subtitle-B/chapter-XIV/part-1607) before you cite it.
- **No minimum sample, but a 2% rule.** An auditor may exclude a category that makes up less than 2% of the audit data. The excluded category's count and rate must still be published.
- **Your data first.** If you have historical data, the audit must use it unless you can explain why it is insufficient.
- **The employer audits.** A vendor is not responsible for conducting a bias audit of its tool. The employer or agency using the tool is.

The 6 RCNY 5-301 columns are the minimum you need to rebuild the table: count of applicants, count selected, selection or scoring rate, impact ratio, and the count of applicants whose category is unknown. Log those columns per step from the first day of use, or you will not be able to produce the audit later.

Finding an auditor who has done this before takes more than a keyword search, because the useful proof is a published audit summary with a name on it. [Refolk](/) lets you search for that proof directly.

Ask me this: `Independent auditors who have published NYC Local Law 144 bias audit summaries` - [run the search](https://www.refolk.ai/start?q=Independent%20auditors%20who%20have%20published%20NYC%20Local%20Law%20144%20bias%20audit%20summaries).

*Returns people who have put their name to public LL144 bias audit summaries, with the profiles that show it, so you can read their method before you engage them.*

## The step-by-step procedure

Run these seven steps in order for each step in your inventory. In total they take two to three weeks for a first pass, most of it in the inventory.

#### Classify every automated step in the pipeline

1. **Inventory every automated step** - List every automated step, including features built into the ATS. Owner is recruiting ops, 1 to 2 weeks; done when each step has a row with vendor, feature, output and funnel position.
2. **Fix the population boundary** - Tag each step pre-application, post-application or both, and log the event that turns a sourced person into an applicant. Recruiting ops plus counsel, about a day.
3. **Classify the output type** - Map each output to ORGANIZE, RANK, SCORE, CLASSIFY or FILTER, then to Colorado's and the Annex III verbs, from sampled outputs. Ops with the vendor, 2 to 3 days.
4. **Assess decision weight and human override** - Score weight NONE, FACTOR or DECIDES, then test review against California's three-part test. Compliance owns this; done when training, review evidence, authority and override rate are documented.
5. **Tag job and candidate location** - For NYC, California, Colorado, Illinois and the EU, tag job location and candidate residence separately, starting with the job.
6. **Assign the verdict and obligation list** - Check each regime's trigger; the highest tier wins and the obligations of every applicable regime are combined. Done when a second reviewer reproduces the verdict.
7. **Set the evidence and audit cadence** - Schedule the annual LL144 audit and set retention for every step at the longest period that applies anywhere, which is 4 years.

Step 6 is the test of the framework. If two reviewers reach different verdicts, the cause is almost always a dimension value someone inferred rather than evidenced. Go back to that dimension. Do not argue the verdict itself.

Plan for who will actually do this work. In Refolk's index of professional profiles, 160 US profiles carry recruiting, talent acquisition or recruitment operations titles, against 26 with AI governance, responsible AI or AI compliance titles.

**6.2:1 - US recruiting-ops profiles per US AI-governance profile, in Refolk's index**

Counts are title-matched, so treat them as a lower bound; in practice the classification lands on recruiting ops, not a specialist.

The EU is thinner still. Refolk's index holds 18 recruiting-ops profiles across Germany, France, the Netherlands, Ireland, Spain and Italy, about 8.9 times fewer than the US, and KPMG Deutschland accounts for 5 of them. If you need a second reviewer with EU deployer experience, start that search before the inventory is finished. Refolk can run it in plain English, for example: talent acquisition operations leads in Germany or the Netherlands at companies preparing for EU AI Act high-risk obligations.

## How the step read goes wrong

Each failure below produces a verdict that looks right and is wrong. Each one comes with a check you can run against your own logs.

1. **Population drift.** A sourced person clicks an apply link, and a re-rank runs on them while they are still labelled "sourced". LL144 now applies and nobody noticed. *Check:* compare ATS application timestamps against score timestamps. Any score timestamped after an application belongs to a POST step.
2. **Human review recorded as a universal exemption.** "A recruiter approves" can satisfy California's test and still fail Colorado, where a human's final say does not remove a weighed score from scope. *Check:* look for logged reviewer changes. A 0% override rate suggests rubber-stamping and also fails California.
3. **A summary that labels people.** A summary that calls people "strong" or "weak" is a classification, not an organizing tool. *Check:* sample outputs for labels, tiers and adjectives before accepting ORGANIZE.
4. **Location keyed to residence.** Teams key NYC to where the candidate lives. LL144 scope turns on where the job is. *Check:* every step must carry a job-location tag before its candidate tag.
5. **The 2% exclusion done silently.** Small categories get dropped without a trace. The rule does not make them disappear: their count and rate must still be published. *Check:* every excluded category appears in the audit summary with its numbers.
6. **Reliance on the vendor's audit.** A vendor-commissioned audit is treated as covering the employer even though the employer's own historical data exists. Historical data must be used unless the employer can explain why it is insufficient. *Check:* confirm which dataset the audit used, and record why yours could not be used if it was not.
7. **Treating the withdrawn Illinois draft as binding.** IDHR proposed notice rules and then withdrew them. The withdrawn draft would have required notice within 30 days of adopting a new or substantially updated AI system, notice in job postings, a contact person and recordkeeping. None of that is binding. The withdrawal also did not suspend, delay or modify the statute. *Check:* your Illinois notice cites the statute, and any element taken from the draft is marked as a voluntary choice.

> **Watch out:** A recruiter's approval is not a universal exemption
>
> The most common error is a single "human in the loop" tick box applied across every jurisdiction. File the three-part evidence for California and score Colorado separately, because the same evidence that exempts a step in one state shows influence in the other.

> A recruiter's sign-off is an exemption in California and evidence of influence in Colorado.

One more failure is quieter. Teams classify the product once instead of each step. A sourcing product with a search step and a post-application re-rank has at least two verdicts. If the whole product carries the lower one, the re-rank goes unaudited.

## What evidence to keep, and for how long

Keep a five-layer evidence file for each step, and keep all of it for 4 years. Retention across the regimes runs from at least 6 months to 4 years, and only the longest period satisfies all of them.

| Item | Period |
|---|---|
| NYC candidate notice | 10 business days before use |
| NYC bias audit | Within 1 year before use, then annual |
| CO adverse-outcome notice | 30 days |
| CO records | 3 years after the decision |
| EU deployer logs | At least 6 months |
| CA FEHA automated-decision records | 4 years |

No public source specifies what a per-step log must contain. The 5-301 audit columns are the floor. The layers below are my working answer to the question each reviewer, auditor or regulator will ask.

#### The evidence file for one step

1. **Verdict record** - Five scores, verdict, obligations, two reviewer names
2. **Review log** - Reviewer training, what they saw, every change they made
3. **Output samples** - Dated samples checked for labels and thresholds
4. **Boundary log** - Application timestamps against score timestamps
5. **Inventory row** - Vendor, feature, output, funnel position

*Each layer answers a different reviewer's question, and the outer layers are worthless without the inner ones underneath.*

Retention starts at different points under different regimes. Colorado counts three years from the date of the consequential decision, not from when the record was created. Record the decision date on every record so the clock can be computed later.

## How do you keep the read current?

Re-read a step whenever its inputs change, and re-read every step whenever a regime's rules change. Treat the classification as a living record with named triggers, not a one-off project.

Triggers that reopen a single step:

- The vendor ships a new feature or changes how an output is produced.
- A threshold is added to a score, which turns it into FILTER.
- The step starts touching applicants, or a new office opens in NYC, California, Colorado, Illinois or the EU.
- The override rate falls to zero over a review window.

Triggers that reopen every step:

- **Colorado:** the Attorney General's final rules, which should define "materially influence". Watch the [SB26-189 bill page](https://leg.colorado.gov/bills/sb26-189).
- **Illinois:** a new IDHR rulemaking on notice. Check the IDHR rulemaking agenda for first-notice dates rather than relying on the withdrawn draft.
- **California:** the section 7221 hiring opt-out exceptions, once counsel has read them.
- **EU:** any further amendment to when Annex III applies. Check the dates in the Official Journal.
- **NYC:** changes to the DCWP rules at [6 RCNY 5-301](https://codelibrary.amlegal.com/codes/newyorkcity/latest/NYCrules/0-0-0-138530) or to the FAQ.

> **Note:** Deferred enforcement is not a deferred deadline
>
> Colorado's Attorney General has said there will be no enforcement until rulemaking concludes, but the law still takes effect on 1 January 2027. Build the notices and the review route to that date, and use the enforcement gap for testing, not waiting.

#### Before you file a step's verdict

- [ ] The step has an inventory row with vendor, feature, output and funnel position
- [ ] The population tag is backed by a logged apply event, not inferred
- [ ] Output type was set from a dated sample of real outputs, checked for labels and thresholds
- [ ] Decision weight is FACTOR or higher for any output a decision-maker sees
- [ ] All three parts of California's human-involvement test have written evidence, or the step is scored FAIL
- [ ] The logged override rate is above zero, or the rubber-stamp risk is written down
- [ ] Job location and candidate residence are tagged separately for all five jurisdictions
- [ ] The verdict is the highest tier reached, and the obligation list combines every applicable regime
- [ ] Every not-established regime is marked ASSUMED with a date and an owner
- [ ] A second reviewer reached the same verdict independently
- [ ] Retention is set to 4 years and the decision date is recorded
- [ ] Audit Required steps under LL144 have an audit date within one year before use and an annual renewal on the calendar

If you are starting from nothing, inventory first and classify the post-application rank and filter steps before anything else. Those are where LL144, Colorado and the EU overlap, and an error there is the most expensive kind.

## Frequently asked questions

### Does NYC Local Law 144 apply to sourcing tools?

Not when the tool only assesses people who never applied. The DCWP FAQ says scanning a resume bank, doing outreach and inviting people to apply are out of scope, because a candidate is someone who has applied for a specific position. Once a sourced person applies, any ranking or scoring of them is in scope. The FAQ is guidance, not law, and this carve-out is NYC's alone. Colorado and the EU AI Act can still reach the same sourcing step.

### Does human review exempt an AI hiring tool under Colorado's law?

No. Colorado covers outputs that materially influence a decision, including by ranking, scoring, recommending or classifying. Both definitions proposed during rulemaking share a presumption: if the decision-maker reviewed the output and the outcome matches it, the output materially influenced the decision. The exclusion that does exist covers tools used solely to summarize or organize information for human review. Note that the draft rules leave the term 'materially influence' undefined, so re-check when the final rules are published.

### When do job applicants need a CCPA ADMT pre-use notice?

You need one when automated decisionmaking technology replaces or substantially replaces human decisionmaking on a significant decision, such as hiring. A step falls outside the rules if a human reviewer can interpret the output, actually reviews it alongside other information, and has authority to change the decision. If the rules do apply, the duties are a pre-use notice, an opt-out or human appeal, and answering access requests. Uses that began before 1 January 2027 must comply by that date. Uses starting later must comply from day one.

### Is an AI sourcing tool high-risk under the EU AI Act?

Often, yes. Annex III point 4(a) covers AI intended for recruitment or selection, and names placing targeted job advertisements, filtering applications and evaluating candidates. The narrow-task exit in Article 6(3) rarely opens, because a system that profiles people is always high-risk. It is not publicly established whether a passive search of people who never applied counts as evaluating candidates. Mark such steps as assumed in scope until that is settled.

### Who has to commission the LL144 bias audit, the vendor or the employer?

The employer or employment agency using the tool. The FAQ states that a vendor is not responsible for conducting a bias audit of its tool. A vendor-commissioned audit may help, but if you have your own historical data, the audit must use it unless you can explain why it is insufficient. Before relying on a vendor audit, check that the audit used your data, or record why it could not.

### How long should I keep evidence for an AI hiring step?

Keep it for 4 years. The regimes range from at least 6 months for EU deployer logs, through 3 years after the decision for Colorado, to 4 years for California FEHA employment records that include automated-decision data. Using the longest period for every step means no regime is ever under-served, and you avoid managing a different deletion schedule for each step.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/guides/ai-hiring-pipeline-step-classification*
