The AI hiring stack in 2026: what's used, what's regulated

What the AI hiring stack is, and what it is not

You are not a recruiter, and in 2026 that is still a defensible position. “AI hiring stack” sounds like a checklist you are behind on; treat it as an inventory of what exists, of which a small team touches one layer or two on purpose. In SHRM’s 2024 Talent Trends survey, fielded on 2,366 US HR professionals in January 2024, 26% of organizations used AI for any HR activity at all, and among those that did, recruiting and hiring was where it landed most often, at 64%. One level deeper the name misleads: in that same 2024 data, among organizations using AI for recruiting, 65% generated job descriptions with it and 7% ran AI-led pre-screening interviews. In use, most of an AI hiring stack is a writing layer, not a judging layer. The part that can cost you money sits at the far end, where a hire becomes a paycheck under somebody’s employment rules.

What has a real account behind it here is narrow. Five tools were tested hands-on and scored between June and July 2026: Manatal 9, Breezy HR 9, Deel 8, Gusto 8, Tellent 8. Paid or self-registered accounts only, one real workflow built in each account, no vendor demo environment, no aggregate community rating behind any score. Prices carry a capture date and a source label: Manatal’s and Breezy’s were read in-app, Gusto’s and Deel’s on their published pricing pages in June and July 2026, and Tellent publishes no price at all, so no number is pinned to it. Where a test stopped short, the review says so: with Deel I set up a real account and posted a real role, but ran no cross-border EOR hire and no payroll run. AI video interviewing, assessment products and standalone sourcing engines have no account behind them at all, and I say so in the section instead of a footnote. The rubric is the 5-Signal Test, and what it refuses to score.

Applying for jobs and wondering whether AI is reading your resume? This page is written for the employers who buy these systems, but one number in it is worth your time: in SHRM’s 2024 survey, among organizations already using AI in recruiting, 34% used it to review or screen resumes while only 7% ran AI-led pre-screening interviews, so the system you are most likely to meet is screening your resume rather than interviewing you. For what that system is and what it keeps on file, see the glossary entry on what an ATS is.

Four layers follow, and the split is mine rather than an industry taxonomy. Each gets its dated evidence, a tested or not-tested label, and one question: what number comes out of it, and does that number survive being shown to your boss? Call it the Boss-Slide Test. For each layer, name the one number it puts on a slide, say where that number comes from, and say who answers for it when somebody questions it. A layer that fails this test can still be worth using. It is just not a budget line, and treating it as one is how small teams buy the wrong thing. Each layer ends in one decision: adopt it, buy nothing and use what you already have, or wait. No catalog, no ranking. Nor is this the automation map on this site, which charts rules-based work across the whole employee lifecycle; here it is the hiring workflow alone and the probabilistic layer on top. For that wider picture, start from where HR automation stops and AI starts.

Why one survey says 26% and another says 72%

Two numbers circulate whenever anyone asks how much AI there is in hiring, roughly 26% and roughly 72%, and the temptation is to average them. Neither one measures hiring the way it gets quoted, which is why the gap is the more useful object.

The 26% is SHRM’s 2024 Talent Trends survey, fielded 10 to 19 January 2024 on SHRM’s Voice of Work panel of 2,366 US HR professionals: unit, the organization; object, any HR activity, not hiring. The 72% is HireVue’s 2025 Global Guide to AI in Hiring, published 19 February 2025 on a panel of more than 4,000 HR leaders and employees worldwide, reporting adoption among HR professionals rising from 58% in 2024 to 72% in 2025: unit, the individual professional. HireVue sells AI video interviewing, so its panel over-represents organizations already inside the category. Both can be accurate. They do not measure the same object on the same population, which is why arithmetic between them produces nothing.

One number from the 2024 data cuts against the replacement story anyway: among organizations using AI in recruiting, interviewing or hiring, 88% said it saved them time or made them more efficient, and 2% said AI had displaced workers at their organization. None of that answers whether you are behind, so go down the layers instead.

The Who-Asked-Whom Test

Three questions, mine rather than a recognized method. They take twenty seconds against any AI hiring statistic in a deck.

1. Who was in the sample, and what is the denominator: organizations, or individual professionals? Does “among companies already using AI” quietly narrow it?
2. When was the fieldwork, as opposed to when the report was published?
3. Who sells something that depends on the answer? Being a vendor survey is not disqualifying. Going unlabeled is.

A clean case of the first question: SHRM surveyed 129 chief HR officers between 7 October and 5 November 2025 and 92% anticipated greater AI integration. That is a leadership panel rather than a market, and anticipating is not deploying.

The writing layer: job descriptions, postings, outreach

In SHRM’s 2024 data, among organizations using AI for recruiting, 65% used it to generate job descriptions, 42% to customize or target postings and 33% to communicate with applicants during the interview process. That is the top of the list by a wide margin, and all of it is drafting.

Not tested here as a product, for a structural reason: it is rarely a line on an invoice. It is the general assistant your company already pays for, or a button inside the applicant tracking system you already bought. One guardrail needs no test: do not paste a candidate’s personal data into a general-purpose chatbot for a summary. The Boss-Slide Test fails here, and that is simply the honest answer, because almost no number comes out of this layer. Nobody walks into a management meeting with “we wrote the ad faster.”

Decision: adopt, but do not buy. The value is already inside a subscription you hold. If you are weighing a paid AI feature here, which AI add-ons actually earn their price covers it, and what these tools cost, vendor by vendor sits in the pricing index.

The sorting layer: screening resumes and searching candidates

In the same 2024 SHRM data, among organizations using AI to recruit, 34% used it to review or screen applicant resumes and 33% to automate candidate searches. A third, not everyone. It is also the first layer doing something you could not reasonably do in a spreadsheet.

Tested. On a real junior full-stack role run through Manatal on its free trial in June 2026, the AI returned 106 ranked candidates and scored each against the job, the top two at 78% and 75%, each showing its working as a required and preferred skills breakdown: 106 AI-scored candidates on a single role. That is the full extent of the first-hand AI evidence here, and its boundary sits next to it. Breezy HR also scored 9 out of 10 across June and July 2026, but its AI layer, Breezy Intelligence, is a paid add-on I did not buy, and on the plan I tested the candidate AI-score column stayed empty. The 9 is for the core product: the AI matching add-on I did not buy.

This is the layer that answers to a boss. Candidates per role, time in each stage, source of hire and disqualification reasons all come out of the tracking system, not the AI riding on it. One condition attaches, and it sets up the regulatory section below: insist on visible reasoning, because a score you cannot explain is a score you cannot defend.

Decision: adopt if you hire more than occasionally, knowing you are buying the applicant tracking system first, and the AI comes either bundled in it, as Manatal’s was on the trial I ran, or as a paid add-on, as Breezy’s is. Start from the ATS shortlist I tested for in-house teams, or on a zero budget, every free ATS plan and the wall it hits. Below that frequency a line per candidate in a spreadsheet is a rational system: no cost, no outage, and you can see all of it at once. A tool has to beat that on something nameable, usually not losing people, and not on scoring. If you fill roles for clients rather than your own team, the agency desk trades off differently.

The judging layer almost nobody runs: AI interviews and assessments

This is the layer the category is named after in the popular imagination, and it is the smallest thing on the board. In SHRM’s 2024 data, among organizations already using AI to recruit, 7% used it for pre-screening interviews such as AI-powered video interviews and 3% to analyze applicants’ interview performance. The population does the work: that is 7% of the subset already using AI in recruiting, not 7% of employers.

The counterweight needs its label. HireVue’s 2025 report found HR leaders’ confidence in AI systems rising from 37% in 2024 to 51% in 2025, and 41% of HR professionals saying they use skill assessments in hiring. HireVue sells this layer and is also the party measuring confidence in it, and that 41% covers skill assessments in general, not AI-driven ones. The layer with the loudest numbers is largely measured by the company selling it, while SHRM’s 2024 measurement of actual use, taken by a party with nothing to sell in this layer, says 7%.

I have not tested this layer. No AI interviewing or assessment product has been run on an account here, so none gets a score or a recommendation from me, only sourced facts carrying their dates. The Boss-Slide Test returns a liability here. The number is a score whose derivation you may not be able to explain, computed on a pool you narrowed by asking for it.

Decision: wait, at your size. It solves a volume problem that a company hiring four times a year does not have: a continuous req load where applications arrive faster than anyone can read them. A structured question set and two interviewers is not a stopgap in the meantime. It is the control condition anything else has to beat.

The number you cannot see

Pew Research Center surveyed 11,004 US adults between 12 and 18 December 2022 and found 66% saying they would not want to apply for a job with an employer that uses AI to help make hiring decisions, with opposition to AI making the final hiring decision running 71% to 7%.

For an employer that is a pipeline number. The risk you are managing is the applications you never receive, and they appear in no funnel report your applicant tracking system can produce. The date works both ways. That measurement predates generative AI as a consumer product and nobody has replicated it at that scale since, which makes it the best baseline available and the oldest number on this page. Treat it as a dated floor, not a current reading.

Where the AI hiring rules stand in 2026

Three of the jurisdictions that were supposed to tighten AI hiring rules in 2026 moved backwards instead. Colorado repealed and reenacted SB 24-205 through SB 26-189, signed on 14 May 2026, before the original law had ever taken effect. Illinois left Public Act 103-0804 in force from 1 January 2026 while its implementing rules, proposed on 15 May 2026, were withdrawn on 2 June 2026. And the EU moved its high-risk hiring obligations from 2 August 2026 to 2 December 2027 under Regulation (EU) 2026/1744, in force since 27 July 2026. The one obligation live throughout, New York City’s Local Law 144, produced 18 published audit reports across the 391 employers examined in a 2024 study. Over the same period, what kept moving forward sat outside the statutes: federal enforcement in 2023, then private litigation through the Mobley certification in May 2025. On this record, the machinery still advancing is the courtroom, not the rulebook.

The Boss-Slide Test in this section is your audit trail: who was rejected, by what, on which criteria, and whether you can export it.

When do the EU AI Act hiring obligations apply?

For most recruitment systems, the answer moved in July 2026: 2 December 2027, not 2 August 2026. Under Regulation (EU) 2024/1689, in force since 1 August 2024, AI systems intended for recruitment or selection, including systems that filter job applications and evaluate candidates, are classified as high-risk in Annex III. That classification has not changed. The deadline has. The Digital Omnibus, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, moving the Annex III high-risk obligations, hiring included, to 2 December 2027, with Annex I to 2 August 2028. The Article 50 transparency obligations were not moved and still apply from 2 August 2026. Whether the Regulation reaches a particular company, and which duties fall on the employer as opposed to the vendor supplying the system, are not questions this page settles. The dates above are the part that changed in July 2026, and a page still giving you the earlier date has not been updated since.

Which US states regulate AI in hiring in 2026?

Three jurisdictions are worth reading in 2026, and this page tracks those three.

Illinois has a law in force with no rules behind it. Public Act 103-0804 took effect on 1 January 2026, barring discriminatory use of AI and requiring notice across recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline and tenure. The Illinois Department of Human Rights published proposed implementing regulations on 15 May 2026 and withdrew them on 2 June 2026, citing the need for continued collaboration with other state agencies. The proposal never took effect, so the position today is a statute in force with nothing interpreting it.

Colorado’s pioneering SB 24-205 never took effect at all: its start date slid from February to June 2026, and then SB 26-189 on automated decision-making technology, signed by the governor on 14 May 2026, repealed and reenacted those provisions with new requirements before the 30 June 2026 effective date arrived. What replaced them is not nothing. SB 26-189 reenacted the provisions with new requirements, and their content and dates sit outside what this page verified on 21 August 2026, so a Colorado employer should read SB 26-189 itself and not any 2024 summary of the old act.

New York City is the one jurisdiction here whose obligation has produced a public compliance record. Local Law 144 has required bias audits of automated employment decision tools since July 2023, and a 2024 ACM FAccT study by researchers at Cornell and Data & Society, working with 155 investigators, examined 391 employers and found 18 published audit reports and 13 transparency notices, roughly 4.6% visible compliance.

What actually creates legal risk in 2026

Two cases carry more practical weight than the statutes above them. In 2023 the EEOC settled its first AI hiring case: iTutorGroup paid $365,000 under a consent decree announced on 11 September 2023, over 2020 conduct, after its recruiting software was programmed to automatically reject female applicants aged 55 or older and male applicants aged 60 or older, screening out more than 200 qualified applicants. Then on 16 May 2025, Judge Rita Lin of the Northern District of California granted preliminary certification of a nationwide ADEA collective action in Mobley v. Workday, covering applicants aged 40 and older denied employment recommendations through the platform since 24 September 2020.

The new thing is the defendant, which is the vendor. Your own exposure is not new: an employer’s liability under the ADEA and Title VII did not wait for AI and does not transfer to the software company when the software gets it wrong. For a twenty-person team that reduces to one rule. A person signs the advance-or-reject decision and the reasoning stays visible, which is what the sorting layer already required for its own reasons.

This is not legal advice, and the regulatory state described here was verified on 21 August 2026.

The last layer: paying the person you just hired

The fear at this end is specific, and anyone who has run payroll for a small team knows it. It is the payslip that does not land on Friday, a withholding deposited in the wrong jurisdiction and found by letter three quarters later with interest attached, or the discovery that you have been paying somebody in a state you never registered in because they moved and told the wrong person. Nobody arrives here curious about AI. They arrive because something broke, or because they can see it about to.

That shape is the honest bridge to the international version. One more state breaks a payroll the way one more country does: a registration you did not make, a filing calendar you did not know existed, a return nobody owns. The mechanism is familiar; the magnitude is not: Deel’s own global payroll compliance checklist, updated in August 2026, counts more than 30 countries that changed payroll, employment tax, or mandatory benefits rules between 2025 and 2026. One dated market figure belongs here, stated without overreach: the World Economic Forum’s Future of Jobs Survey, conducted in late 2024 across more than 1,000 employers in 55 economies, found two-thirds of employers planning to hire talent with specific AI skills. Those two datasets meet in one place.

The cross-border side of the same data

Deel’s 2025 State of Global Hiring Report, published on 11 March 2026 from platform data covering more than 1 million workers at over 37,000 companies in 150+ countries, reports that “AI trainer” was the fastest-growing cross-border role on its platform, up 283% in 2025, with more than 70,000 workers training AI systems across 600+ organizations: 58% in the United States, 7.2% in India, 4.6% in the Philippines, 1.7% in Kenya.

Put next to the WEF figure, that is the claim worth keeping, and it is narrower than a prediction: on the evidence available for 2025, the fastest-growing cross-border role on Deel’s platform was not one AI displaced but one that exists to build AI. That is what these two datasets show about the market. Neither is a forecast about where you will hire, and nothing here says you should hire outside your own country.

Two readings belong with it. These are platform figures describing the people being paid through Deel, and its public reputation leans the same way: on the first page of its Capterra reviews, sampled in July 2026, 13 of the 15 visible entries came from workers receiving payments rather than the employers who buy the product, so the 4.9 out of 5 across 4,292 Capterra reviews, checked July 2026, tells you Deel is pleasant to get paid through and little about the buyer’s job. It is also first-party data from the company selling what the data implies you need, and that company is an affiliate partner of this site, disclosed at the top of this page. Read it as a market signal, not a mandate.

Where a US payroll tool stops

A domestic payroll tool is built around one country’s tax system. It does not become a UK payroll because somebody added a country field: the filing calendar, the statutory deductions, the employer registration and the year-end forms are different objects, not different values in the same object. The currency field is the same kind of trap. In Deel’s 2025 platform data, US dollars appeared in five of the ten most common country-currency payment combinations, which means the country somebody works in and the currency they are paid in are two separate fields that often disagree. A payroll system either models that pair or it quietly assumes they match. That is a product boundary, not an argument for any particular employment structure. For the domestic half of that boundary, the payroll runs that fire without you covers the US tools that do automate the run end to end.

The platform with an account behind it here is Deel, tested at 8 out of 10: my tested Deel verdict, 8 out of 10. Three limits belong in the same breath as the name. The $599 EOR figure in its published pricing, captured July 2026, is a platform fee only, with salary and local employer taxes on top, plus a currency spread and a refundable deposit that independent reviews report and I have not verified in my own account. Budget from a real quote; the $599 is where the bill starts. Several products are quote-only, including US PEO, US payroll, IT and the fuller HRIS, so you cannot self-serve a total. And the test behind that score was partial: a real account, the dashboard walked, a genuine junior full-stack role posted through the recruiting flow on 7 July 2026 with two talent partners engaged, but no completed cross-border EOR hire and no payroll run, which caps the score below a 9 under this site’s own rubric. The full point it loses on value is a separate matter: the highest fees in this category, and a total you cannot forecast from the page. If you are buying because a person is about to land somewhere your company does not file, see Deel’s plans and budget from a real quote instead of the headline.

This page will not pick your employment model. Whether the answer is your own entity, an employer of record, or a contractor relationship that honestly fits the work depends on where the person is and what you already have registered, and how the EOR and PEO models differ lays out that trade rather than settling it here. The counterweight matters as much: if everyone on your team is in the United States, this layer is not yours, and Gusto or Deel, decided by where your people are runs that comparison.

The Boss-Slide Test is easiest to pass here and hardest to fake. Total cost per person per country and filings made on time are numbers a finance lead reads without translation, where candidates per role has to be explained first, and mistakes here are billed as penalties and interest on a filing calendar, not only as a bad hire. The list price is not one of those numbers. Decision: adopt when a person lands in a jurisdiction you do not already file in. Not before.

So which layer deserves your budget this quarter?

Four decisions, in the order the work happens, and the four percentages share one denominator: organizations that already use AI to recruit, in SHRM’s January 2024 fieldwork. Writing, at 65% of them: adopt, but do not buy, because it is inside something you already pay for. Sorting, at 34% of them: buy the tracking system and the AI comes with it, and only if you hire often enough to lose track of people. Interviewing and assessment, at 7% and 3% of them: not at your size, and I have not tested it. Payroll and compliance: only when a person lands in a jurisdiction you do not already file in. When that trigger fires, settle who the legal employer is before you shortlist anything, because that answer decides the contract, the deductions and the liability; how the EOR and PEO models differ lays out that trade. Once the structure is settled and a platform is the right object to price, see what Deel charges and budget the total, not the platform fee.

On the sorting layer a spreadsheet is still a defensible answer, and on the judging layer a structured question set and two interviewers is, and nothing above is written to embarrass anyone running either. If you have 100 or more people, or several legal entities, this page is the wrong shape for you: the thresholds here are set for a team that hires in bursts.

Run the three questions against any hiring statistic before you repeat it, and use the EU date as the worked example: a fact that became true on 27 July 2026 and was false two weeks before that shows freshness is a reading criterion, not a maintenance detail. The same discipline covers the numbers this page deliberately does not print, such as the recruiting-trends figure fielded before 2020 that still circulates as current, or the round percentage saving on cost per hire that I have never been able to trace back to a primary source. And the Boss-Slide Test one last time: if a layer does not produce a number you would show your boss, it is not a budget line. It is a habit.

This is not legal advice; the regulatory state on this page was verified on 21 August 2026.