AI in HR and recruitment: hiring AI laws checklist (2026).
Where AI helps HR and recruiting teams today, where a person has to decide, and a checklist of the hiring AI laws in New York City, Colorado, the UK and EU as of October 2026.
For mid-size employers and recruitment agencies, the safest place to start with AI in HR is work that doesn’t decide anything about a person: answering candidates, scheduling interviews and answering employees from your own policies. Scoring, ranking or rejecting people is different. It is one of the most regulated uses of AI anywhere, and the rules below are current as of October 2026.
What can AI do in HR and recruitment today?
Answer candidates at any hour
Candidates ask the same few things: where their application stands, what the process involves, when the interview is, and whether the role is still open. An assistant connected to your applicant tracking system (ATS) can answer from the real record, in the candidate’s language, and hand over to a recruiter when the question is about an offer, a rejection or a complaint.
Schedule interviews
Finding a slot across a recruiter, two interviewers and a candidate is a constraint problem. The assistant offers times from the real calendars, books the room or video link, sends reminders and handles reschedules, then writes the result back to the ATS.
Run the employee helpdesk
Leave, benefits, payroll dates, expenses, policy wording. A retrieval assistant answers from your handbook and benefits documents, shows which document each answer came from, and passes anything personal or sensitive to the HR team. It should only see what the person asking is allowed to see, which is why we build the permission check outside the model. Our note on retrieval that admits when it doesn’t know covers why the “not covered by policy” answer matters most here.
Handle onboarding paperwork
Chasing and checking new-joiner documents, answering first-week questions, and drafting offer letters, contracts and policy acknowledgements from your approved templates for HR to review before anything is sent.
Help recruiters with the paperwork
Drafting job descriptions and outreach, turning interview notes into the structured scorecard your process uses, and keeping the ATS up to date. The recruiter reads and owns every output.
Run first conversations for a recruitment agency
Agencies placing large numbers of candidates can start on WhatsApp. An agent can collect the facts a recruiter would otherwise chase one by one: availability, notice period, location, right-to-work status, licences and preferred shifts. WhatsApp’s business messaging policy (last updated 23 September 2026) allows you to contact people only if they have opted in, and expects clear escalation paths, such as a transfer to a human agent, when automation is used. Our guide to WhatsApp AI agents covers how those flows work. The line to hold: the agent gathers facts, and a recruiter decides who moves forward.
Which kind of AI fits which HR job?
Not every HR problem needs a language model, and the risk class depends on what the output is used for, not on the technology.
| Use case | Right kind of AI | Systems and data | Risk under hiring laws |
|---|---|---|---|
| Candidate questions and status | LLM agent reading the ATS record | ATS, careers site, job descriptions | Low: informs, decides nothing |
| Interview scheduling | Scheduling logic with an LLM for the conversation | Calendars, ATS, video tool | Low |
| Employee policy helpdesk | Retrieval (RAG) over your documents | Handbook, HRIS for permissions | Low to medium: privacy and access control |
| Offer letters and contracts | Template-based drafting with review | Approved templates, HRIS | Low if HR approves each one |
| Reading CVs into structured fields | Document AI | CVs, ATS | Medium: safe as data entry, regulated once the fields are used to filter |
| Ranking, scoring or filtering candidates | Machine-learning or LLM scoring | CVs, past outcomes | High: bias audits, notices, human review |
| Job-ad targeting | Ad-platform optimisation | Ad accounts, applicant data | High in the EU: named in Annex III |
| Promotion, contract terms, task allocation, performance or dismissal | Not an AI decision | HRIS, performance records | High: a person decides |
Which systems does it need to connect to?
Most of the work is integration. The usual set is an ATS (Greenhouse, Lever or Workday Recruiting), an HR information system (Workday, BambooHR), calendars and video tools, email, an e-signature tool, background-check and right-to-work providers, and WhatsApp or web chat. Most expose APIs for the actions an assistant needs, but coverage varies, and some write operations are limited. Plan for the integration to take longer than the AI.
Two design points matter more in HR than elsewhere. First, permissions: the helpdesk should only retrieve what the specific employee may see, enforced in code. Second, records: every answer the assistant gives a candidate or employee should be logged with the source it used, so HR can explain it later.
Where AI is the wrong tool, and where a person stays in charge
Four places we advise against automating, whatever a vendor demo shows:
- The decision to reject or hire. A model can organise evidence. A named person should decide, with the reasons recorded.
- Inferring who someone is. The UK data regulator’s audit of AI recruitment tools found some that let recruiters filter out candidates by protected characteristics and some that guessed gender and ethnicity from names. Collect what a person tells you directly, and don’t infer the rest.
- Reading emotions or personality from faces and voices. The EU prohibits AI that infers emotions in the workplace (with narrow medical and safety exceptions), in force since 2 February 2025. Whether a given interview tool falls under it is a question for a qualified professional.
- Promotion, contract terms, performance and dismissal decisions. These sit in the EU’s high-risk list and in most employment law. AI can assemble the record. It should not conclude.
A person “in the loop” has to mean something. The ICO’s March 2026 report, drawing on conversations with more than 30 employers, found that many are likely relying on solely automated decisions, and that where human involvement exists it should be applied consistently to all candidates within a hiring stage. A recruiter who approves every suggestion in two seconds is not meaningful involvement. Build the review step so the reviewer sees the evidence, can change the outcome, and the system records that they did.
What do the hiring AI laws require?
This is a checklist of what the rules say, with the official text linked. It is not legal advice: a qualified employment or data protection lawyer decides how it applies to your hiring. Dates are as of October 2026.
| Where | Rule | What it asks of an employer | Status |
|---|---|---|---|
| New York City | Local Law 144 | Bias audit within a year before use, published summary, notice to candidates | Enforced since 5 July 2023 |
| Colorado | SB 26-189 | Notice, a plain-language explanation of adverse outcomes within 30 days, human review, data correction | Applies to decisions made on or after 1 January 2027 |
| EU | AI Act, Annex III | Human oversight, logs, informing workers and affected people | High-risk rules apply from 2 December 2027 |
| UK | UK GDPR, Articles 22A to 22D | Safeguards for solely automated significant decisions | In force since 5 February 2026 |
| Italy | Law 132/2025, Article 11 | Tell workers and union representatives when automated decision or monitoring systems are used in staff management | In force since 10 October 2025 |
New York City
The city’s enforcement page for automated employment decision tools says the tool must have had a bias audit within one year of use, a summary of the results must be public, and candidates must get notice 10 business days before the tool is used. Enforcement is thin: the State Comptroller’s December 2025 audit found the city’s enforcement department reviewed 32 companies and found one issue, while the auditors found at least 17 instances of potential non-compliance in the same companies (NY State Comptroller). Thin enforcement is not a reason to skip the audit.
Colorado
The signed text of SB 26-189 (signed 14 May 2026) covers automated decision-making technology that materially influences consequential decisions, employment included, and treats employees and Colorado-resident applicants as consumers. Deployers must give clear notice at the point of interaction. After an adverse outcome they must explain the system’s role within 30 days and offer human review and data correction. The attorney general is to adopt clarifying rules by 1 January 2027. The act applies to consequential decisions made on or after that date, and it carves out routine scheduling and tools that only summarise or organise information for human review without scoring or ranking.
The EU
Under Annex III, AI used to recruit or select people (placing targeted job adverts, analysing and filtering applications, evaluating candidates) and to decide on promotion, termination, task allocation or performance monitoring is high-risk. After the 2026 AI omnibus, the European Commission’s AI Act page gives 2 December 2027 as the date these rules apply. As the employer you are the deployer. Article 26 asks deployers to assign oversight to people with the competence and authority, keep the system’s logs (at least six months unless other law says otherwise), and inform workers’ representatives and affected workers before using a high-risk system at work. Our EU AI Act guide covers what that means for builders.
The UK
Section 80 of the Data (Use and Access) Act replaced Article 22 of UK GDPR with Articles 22A to 22D, and commenced on 5 February 2026. For significant decisions made solely by automated processing, the employer must tell the person, let them make representations, give them human intervention and let them contest the decision. “Solely” means no meaningful human involvement, which is why the point above about review matters. Decisions using special category data face a tighter restriction. The ICO’s March 2026 recruitment report adds expectations to be open with jobseekers, explain how to challenge a decision and monitor for bias, and it consulted on draft guidance (the consultation closed on 29 May 2026), so check whether the final version has been published. See also our UK AI rules guide.
Italy
Article 11 of Law 132/2025 has applied since 10 October 2025. It requires employers to tell workers when AI is used, in the cases and ways set out in Article 1-bis of Legislative Decree 152/1997. That article covers automated decision or monitoring systems that inform hiring, management, termination, task allocation or evaluation, and requires the information to go to workers and to their union representatives.
How do you check a hiring tool for bias?
The legal minimum differs by place, but the engineering is the same. NYC requires a bias audit. The ICO suggests asking developers how they test for bias and considering monthly reviews. A workable routine:
- Decide what is being measured. For each stage (screen, interview, offer), the share of each group that moves on.
- Collect group data separately and voluntarily. Keep it out of the model’s inputs, prompts and logs, and use it only to measure outcomes.
- Test before launch with matched pairs. Run the same CV with only a name, an address or a career gap changed, and compare scores. Any gap is a defect to fix before launch.
- Monitor after launch. Review stage pass-through by group on a schedule, and re-test whenever the model, prompt or scoring changes. Our note on evals before features explains how to make these tests part of the release process.
- Keep the evidence. Audit reports, notices sent, review decisions and test results, in one place a lawyer or regulator can read.
What does a defensible HR system look like?
| Requirement | How it’s built |
|---|---|
| Decisions stay with people | AI summarises and organises; recruiters and managers decide, and the record shows who decided and what they saw |
| Bias testing | Outcomes tested across groups before launch and monitored after, with independent audits where the law requires them |
| Transparency | Candidates and employees told when and how AI is used, before it is used |
| A route to a person | Anyone can ask for human review and get it, with a named owner and a time limit |
| Data minimisation | Only job-relevant data used; protected characteristics kept out of prompts, features and logs |
| Answers grounded in your policies | The helpdesk answers from your documents with citations, and says when they don’t cover a question |
| Logs and retention | Inputs, outputs, reviewer actions and notices kept for an agreed period, with personal data minimised |
How to start, and what to measure
Candidate communication, scheduling and an employee policy helpdesk are the best first projects. They save the most time, help people directly, and sit outside the high-risk categories because they don’t make decisions about people. Our services cover building these end to end, including the ATS and HRIS integrations.
- Pick one flow, such as status questions and scheduling for one job family, or policy questions from one handbook.
- Collect 100 to 200 real questions from inboxes and tickets, and write the correct answer for each. This becomes the test set.
- Connect read access first to the ATS or document store, then add write actions such as booking a slot.
- Run alongside the team for a few weeks, with a person approving anything beyond a factual answer.
- Widen only on the numbers, and add a handover rule for anything about an offer, a rejection, a complaint or health.
Five numbers to track:
- Time to first response for candidates and employees.
- Share of questions resolved without HR staff.
- Answer accuracy against your test set, rechecked after every change.
- Candidate drop-off between application and first interview.
- Human-review requests and how long each takes to close.
Questions HR teams ask
Can we use AI to screen CVs?
Legally, yes, in most places, but it is the most regulated use there is: bias audits in New York City, notices and human review in Colorado from 2027, high-risk obligations in the EU from December 2027, and safeguards for solely automated decisions in the UK. If you do it, keep a person deciding, test for bias, and tell candidates before the tool is used.
Does a recruiter clicking “approve” make it compliant?
Not by itself. UK law treats a decision as solely automated when there is no meaningful human involvement, and the ICO has said human involvement should be applied consistently to every candidate at a stage. The reviewer needs the evidence, the authority to overrule, and a record that they did.
Will it work with our HR and applicant tracking systems?
Usually, through their APIs: Workday, BambooHR, Greenhouse, Lever and most others expose what an assistant needs. Read access is easy; write actions need testing system by system.
Can a recruitment agency screen candidates on WhatsApp?
An agent can run the first conversation and collect facts such as availability and right-to-work status, provided candidates have opted in and a clear route to a person exists. Treat anything that ranks or rejects candidates as a regulated decision, not a chat feature.
How do we keep employee data private?
Give the assistant only the documents and records each user is allowed to see, keep personal data out of logs, and process data only in approved regions with approved providers. Tell staff and candidates what is kept and for how long.
Modulus Labs AI builds candidate assistants, employee helpdesks and onboarding automation, with people kept in charge of decisions. Tell us where your team’s time goes, and we’ll reply within one business day.