TL;DR: AI in the hiring process means software that sources, screens, schedules, interviews or analyses candidates. 39% of organizations have implemented AI in HR, and recruiting is the top use at 27% (SHRM, 1,900+ HR professionals, 3 Apr 2026). The working rule: no adverse action without a named person reviewing the evidence, not the score. Below are a stage-by-stage map of where that approval sits, a legal summary and a 12-question vendor scorecard.
Updated 28 October 2026.
Key takeaways
- AI touches ten stages of hiring. The stages where it evaluates a person (CV screening, assessments, AI interviews) are the stages that need a human approval gate.
- 39% of organizations have implemented AI in HR (SHRM, 2026), but 56% do not formally measure whether it works, and 88% of HR leaders report no significant business value (Gartner, 28 Oct 2025).
- Practice outruns policy: 21% of companies using AI in hiring auto-reject at all stages with no human review (Resume Builder, n=948, Oct 2024; vendor-commissioned).
- Gates are architecture decisions. Ask whether a human approval step is enforced or merely a setting.
- The 12-question scorecard turns the gate, evidence, bias-monitoring and disclosure issues into questions a vendor must answer in a live demo.
Adoption is not the hard part of this topic. The survey data above says the difficult parts are measuring whether the tools work and knowing who, if anyone, signs off before a candidate is turned away.
This guide takes the stages one at a time, marks where a person has to approve, and ends with a scorecard you can use in a demo. AI interviews get their own section, since they are the stage where software evaluates a person most directly; for the category itself, see choosing AI interview software.
What does AI do at each stage of the hiring process?
The rule this guide applies at every stage is ours, and it is simple: no adverse action without a named human reviewing the evidence, not the score. A score is an input. The evidence behind it (the matched and missing skills, the transcript line, the rubric criterion) is what a reviewer can actually check.
At stages where AI does not evaluate anyone, such as scheduling and the offer, no gate is needed. Where it does evaluate, the gate sits before the decision that affects the candidate. In CV screening, that means a human reviews the pending queue before any rejection, and the AI recommendation is labelled advisory. In AI interviews, it means a human reads the evidence per answer before anyone advances or is rejected.
| Stage | What AI does | Main risk | Where the human approval gate sits | Evidence to demand | Read more |
|---|---|---|---|---|---|
| Sourcing and rediscovery | Ranks an existing pool, finds lookalike profiles | The shortlist decides who is ever contacted | A human chooses who to contact; AI only shortlists | A why-this-candidate rationale and explicit gaps | agentic AI in recruitment |
| Job ads and JD drafting | Drafts and de-biases text | Text goes live without anyone reading it | A human approves before posting | A change log | |
| CV screening | Parses, matches, scores | Ranking can carry bias: University of Washington researchers found three LLMs favored white-associated names 85% of the time against 9% for Black-associated names (31 Oct 2024; scope is resume ranking, not interviews) | Before any rejection: a human reviews the pending queue; AI output is labelled advisory | A per-criterion breakdown with matched and missing skills | CV screening software |
| Scheduling | Coordinates calendars | Low; nothing is evaluated | None needed | n/a | |
| Candidate chat and FAQ | Answers candidate questions | A candidate cannot tell they are talking to a bot | Disclosure that it is a bot; human handoff | Transcript retention | |
| Assessments | Scores tests | Scores with no validity documentation behind them | A human reviews borderline and failing outcomes | Validity documentation | skills-based hiring |
| AI interviews | Asks structured questions, transcribes, scores against a rubric | Scores that cannot be traced to the answer; emotion inference from voice or face, which Article 5(1)(f) of the EU AI Act prohibits in the workplace except for medical or safety reasons | Before advance or reject: a human reads the evidence per answer | Transcript-linked scores; a rubric the candidate can see; a question set drawn from structured interviews | how AI interviews work |
| Interview notes | Transcribes and summarizes | Recording without every participant's consent | A human owns the evaluation form | Consent state per participant | AI interview note-takers |
| Analytics and bias monitoring | Produces selection-rate reports | Flags that nobody acts on | A human acts on adverse-impact flags | Four-fifths reporting by stage | bias audit basics |
| Offer | None, or drafting | Low | Human only | n/a |
The takeaway for buyers is architectural. Whether the gate is configurable or enforced is a design decision made inside the product, long before your first user logs in. That is the question to put to every vendor.
How many companies use AI in hiring, and is it paying off?
Both answers depend on who counted. The table below has six sourced rows, each labelled with its publisher. The full ledger, with samples and verification dates, lives on one page: AI hiring adoption data.
| Fact | Source and publisher type | Sample | As of |
|---|---|---|---|
| 39% of organizations have implemented AI in HR; recruiting is the top HR use (27%); 56% do not formally measure AI investment success, and only 16% use ROI | SHRM State of AI in HR 2026 (professional association) | 1,900+ HR professionals | 3 Apr 2026 |
| AI use in HR tasks rose from 26% (2024) to 43% (2025) | SHRM 2025 Talent Trends (professional association) | 2,040 HR professionals, fielded 3-12 Feb 2025, US, unweighted | 2025 |
| 51% of businesses used AI in hiring; 82% of those use it to review resumes; of AI users, 21% auto-reject at all stages without human review, 50% at the resume stage only, 29% keep human oversight of all rejections | Resume Builder / Pollfish (vendor-commissioned; 2024 data) | 948 business leaders | 22 Oct 2024 |
| 26% of job candidates trust AI will fairly evaluate them (the release cites a 1Q25 survey of 2,918 candidates) | Gartner press release (research firm) | 2,918 candidates | 31 Jul 2025 |
| 88% of HR leaders say their organizations have not realized significant business value from AI tools | Gartner press release (research firm) | Not stated here; the release page shows no sample size for the 88% figure | 28 Oct 2025 |
| 63% of job seekers have faced an AI interview; 70% were not clearly told AI would evaluate them; 38% withdrew because of an AI interview | Greenhouse 2026 Candidate AI Interview Report (ATS vendor's own survey) | 2,950 job seekers | 1 May 2026 |
Read the table in two halves. The first three rows describe use and policy; the last three describe trust and outcomes. Adoption is rising while measurement lags, which is the gap the 56% and 88% figures describe from two directions.
No independent source for hours saved per hire turned up in the research behind this page. Measure it on your own requisitions instead: time from application to first human review, before and after the tool.
Where does a human have to approve?
Gate placement follows the stage map above. The harder questions are what the reviewer sees and who is named.
The evidence a human should read before approving depends on the stage. For CV screening it is the per-criterion breakdown: which skills matched and which are missing. For AI interviews it is the transcript-linked score and a rubric the candidate can also see. A score alone gives the reviewer nothing to check.
Practice differs from design. In the Resume Builder data above, 21% of companies using AI auto-reject at all stages with no human review, 50% use AI-only rejection at the resume stage, and 29% keep human oversight for all rejections. These figures are from 2024 and vendor-commissioned, but they show the gate is a choice that employers make differently.
Candidates have a stake in the gate, too. In the Greenhouse survey, 70% said they were not clearly told AI would evaluate them, and 57% said disclosure should be legally required. Scorecard question 7 carries that issue to the vendor.
On architecture, a design can require one approver or a quorum (all or any of a named group) by job. A vendor may also offer a configurable auto-reject for high-volume roles. Both are legitimate designs if the buyer chooses them knowingly. Whichever you choose, demand an audit log of who decided, when, and on what evidence.
| Benefits | Risks | Full evidence |
|---|---|---|
| Every candidate can get the same structured questions, and every score can link to a transcript line. In one controlled field experiment, AI-led voice interviews with human decision-making raised offers, starts and 30-day retention (see the next section). | LLM resume ranking favored white-associated names 85% to 9% in the University of Washington audit above. | pros and cons of AI in recruitment |
What do AI interviews add, and what do they have to prove?
AI interviews are the stage where software evaluates a person most directly.
On method, the Society for Industrial and Organizational Psychology (SIOP) summarized Sackett, Zhang, Berry and Lievens (Journal of Applied Psychology, 2022) as finding that structured interviews may be the strongest predictor of job performance, ahead of cognitive ability. That finding concerns structure, and it applies whether a person or software asks the questions.
On outcomes, the strongest field evidence we found is a 2025 experiment by Jabarian and Henkel (SSRN 5395709, August 2025) covering 70,884 applications to 48 postings at one recruitment firm hiring for 43 client accounts in the Philippines. According to the paper, AI-led voice interviews, with humans making the decisions, raised offers by 12%, job starts by 18% and 30-day retention by 17%, and 78% of applicants chose the AI recruiter when given the choice. Read the limit with the result: one recruitment firm, one country, one role family.
What an AI interview must prove to a buyer is narrower than what it claims:
- Traceable scores. Every score links to something the candidate said.
- No facial or voice-tone scoring. The EU AI Act's workplace emotion-recognition ban (Article 5(1)(f)) has applied since 2 February 2025, according to artificialintelligenceact.eu; the text names the workplace and does not mention recruitment, so ask counsel whether it reaches interview analysis. HireVue removed facial analysis from its assessments in a January 2021 announcement.
What does the law require at each stage?
This section summarizes obligations by jurisdiction. It is not legal advice.
| Jurisdiction | What it covers | What triggers it | Where to read more |
|---|---|---|---|
| New York City, Local Law 144 | Bias audit within one year before use and candidate notice at least 10 business days before use, per the NYC DCWP; details on the linked page | Use of an automated employment decision tool | NYC Local Law 144 |
| Illinois | The Artificial Intelligence Video Interview Act concerns employer use of AI to analyse video interviews; obligations are on the linked page | Employer use of AI in video interviews | Illinois AI Video Interview Act |
| Colorado | SB 26-189, signed 14 May 2026, repeals and reenacts the Colorado AI Act (SB 24-205); its main provisions apply from 1 January 2027, per Akin Gump (6 Oct 2026). A public rulemaking hearing on draft rules was scheduled for 26 October 2026. | Automated decision systems that materially influence consequential decisions | Colorado AI law |
| California | Covered on the linked page and not summarized here | See the linked page | California AI hiring rules |
| European Union | AI Act: recruitment AI is Annex III high-risk. Under Regulation (EU) 2026/1744 (the Digital Omnibus, in force 27 July 2026), the high-risk obligations apply from 2 December 2027 for Annex III systems; the original date was 2 August 2026. The workplace emotion-recognition ban has applied since 2 February 2025, per artificialintelligenceact.eu. | Use of AI in recruitment | EU AI Act for recruitment |
The full tracker covers more states: see state-by-state AI hiring rules.
Compliance lags awareness. According to SHRM's 2026 survey, 57% of HR professionals in regulated states were unaware of the AI laws that apply to them, and only 12% of those who were aware took compliance steps (SHRM, 1,900+ respondents, 3 Apr 2026).
On litigation: in Mobley v. Workday (N.D. Cal.), the court granted preliminary certification of an ADEA collective on 16 May 2025, and the plaintiffs filed a motion to certify a class on 14 September 2026, according to the CourtListener docket. The claims are allegations, not findings. For the wider docket, see AI hiring litigation.
How do you evaluate an AI hiring vendor? The 12-question scorecard
Score each question 0 to 2: 0 the vendor cannot show it, 1 it is on the roadmap or available only through professional services, 2 the vendor demonstrates it live. The right-hand columns say what to listen for.
| # | Ask the vendor | A good answer sounds like | Red flag | Basis |
|---|---|---|---|---|
| 1 | Show me the evidence behind one score, from the candidate's actual answer to the criterion. | The vendor opens a record and points to the transcript line and the rubric criterion. See explainable AI hiring for what a score explanation should look like. | A number with no source text. | Explainability |
| 2 | Where exactly does a human approve or reject, and is that enforced or a setting? | Names the step, the role, and whether it can be switched off. | "The reviewer can look at it if they want." | Human approval |
| 3 | Is there any configuration that lets the system reject or advance a candidate with no human action? | A direct yes or no, with the settings screen. | Evasion, or "only if you ask us." | Auto-action |
| 4 | What is measured: answer content only, or also voice, face, gaze, timing? Is any of it in the score? | A list of inputs, with exclusions you can verify. | Any facial or voice-tone input in the score; EU AI Act Art. 5(1)(f) prohibits using such signals to infer emotion at work. | Emotion and biometric inference |
| 5 | Can I export selection rates by stage and demographic group, with the four-fifths calculation, without a services engagement? | A self-serve export. | "Our services team runs that quarterly." | Bias monitoring |
| 6 | When was the last independent bias audit, by whom, and can I read the summary? | A date, an auditor's name, a readable summary. | No audit, or a summary kept private. | NYC LL144 |
| 7 | What does the candidate see before and after: notice, criteria, alternative process, a way to request human review? | Screens or text for each. | Notice buried in terms of service. The Greenhouse survey found 70% of candidates were not clearly told. | Disclosure |
| 8 | What audit log exists per decision: who, when, what evidence, what changed? | A log you can export. | Logs for system events only. | Logging |
| 9 | How are impersonation and cheating handled, and what exactly is detected versus inferred? See how proxy interviews work. | Precise language: this signal is detected; this one is inferred and routed to a person. | Claims to verify identity or detect deepfakes, with no demo of what triggers them. | Fraud |
| 10 | What data is stored, for how long, where, and how is deletion handled? | Retention periods, location, a deletion process. | "We keep everything." | Privacy (GDPR and EU) |
| 11 | How does it write back to my ATS: scores, transcripts, decisions? Which connectors are native? | A list of native connectors and what each writes back. | "Via API" with no named integration. | Integration |
| 12 | What will the model change in the next year, and how will you notify me and re-validate? | A change-control policy with notice periods. | Silent model updates. | Change control |
Add your twelve scores. Our view: below 18 of 24, do not proceed to a pilot, and treat a zero on questions 2, 3 or 4 as a stop regardless of the total. Those thresholds are our opinion, not a standard.
The questions draw on the Greenhouse and Gartner candidate-trust data above. Once you have scores, a ranked comparison helps: the AI interview software roundup is the next step.
Where does agentic AI fit?
Agentic AI refers to systems that take multi-step actions toward a goal rather than answer a single prompt. Treat the label with care: ask the vendor which steps the agent takes on its own, and where a person approves.
The gate rule does not change. An agent action that touches a candidate (ranking, rejecting, scheduling an interview on the strength of a score) is the same kind of decision as CV screening or an AI interview in the table above, and it needs the same approval gate. This page does not rank agentic vendors.
For a large AI company moving into hiring directly, see where OpenAI's hiring platform stands.
FAQ
What percentage of companies use AI in hiring?
It depends on how it is counted. SHRM found that 39% of organizations have implemented AI in HR, with recruiting the top use (3 Apr 2026, 1,900+ HR professionals). A vendor-commissioned Resume Builder survey from 2024 found that 51% of businesses used AI in hiring (n=948).
Is AI hiring legal?
AI in hiring is not banned outright in the jurisdictions covered above, but specific uses are restricted (the EU AI Act bars emotion inference in the workplace, with narrow exceptions) and obligations such as notice and independent audit vary by place. The legal table and the tracker linked there give the detail.
Can AI reject candidates without a human?
In some designs and some companies, yes. In Resume Builder's 2024 data, 21% of companies using AI in hiring auto-reject at all stages with no human review. Whether the gate is enforced or a setting is the buyer's decision; scorecard questions 2 and 3 test it.
Will AI replace recruiters?
The sources here do not support a replacement figure. In every design this guide describes, a named person makes the final decision, as humans did in the Jabarian and Henkel experiment.
Does AI in hiring increase bias?
It depends on design and on the stage. The University of Washington audit found strong racial skew in LLM resume ranking, but its scope is resume ranking only. The strongest method evidence in this guide concerns interview structure (Sackett et al., as summarized by SIOP), which applies to people and software alike. For the longer view, read the disadvantages of AI in hiring.
What is agentic AI in recruiting?
In this guide's definition, it is software that takes multi-step actions toward a goal, such as searching, contacting and scheduling, with limited prompting. The section above covers adoption and applies the gate rule.
Which hiring stage should I automate first?
Our view: start with a stage that involves no evaluation, such as scheduling. It carries the lowest legal risk of the stages in the table, because nothing is scored. There is no cited statistic behind this; it follows from where the gates sit.
Who is liable if an AI tool discriminates?
It depends, and courts are still working it out. Mobley v. Workday is the court-record case on whether a vendor can be an employer's "agent"; in July 2024 the court allowed that claim to proceed, according to the AI Lawsuit Tracker. That is a procedural ruling on an allegation, not a finding. See the Workday lawsuit page for the record.
Next step
Book a demo and read the evidence behind each score.
Comments
No comments yet — start the conversation.