AI agents in recruiting: what 36 primary sources actually measured

Every figure here traces to a document I opened, or to a named press release from a publisher that blocks automated readers but serves human ones, Gartner and the Bureau of Labor Statistics being the usual two. Figures are classed three ways. Primary means the original survey, study, or government dataset. Vendor-self means a vendor measuring its own product or customer, which is real information about a claim and never independent evidence. Secondhand numbers I could not chase back to a document got dropped, not footnoted.

That comes to 59 figures from 36 distinct primary documents, published between 2023 and 2026, most of them from 2025 and 2026. Where a survey did not state its population or sample, I wrote unknown rather than guessing, and I flag it in the sentence. One thing would change the conclusions here more than any other: a single independent measurement of time-to-hire with and without an AI agent. It does not exist yet. Keep that in mind through everything that follows.

Between 15% and 26% of the large organizations already running agents, and moving fast. KPMG's AI Quarterly Pulse, which surveys US leaders at companies with $1B+ revenue, put agents-for-recruiting use among its agent-active organizations at 15% in late 2024 and 26% in Q1 2025, with 60% planning to use them within twelve months. That is the fastest-moving number in this entire dataset.

The broader numbers are older and softer. Gartner's January 2024 survey of 179 HR leaders found 38% piloting or implementing generative AI, and 41% of organizations using it for recruiting work like job descriptions and skills data. PwC's April 2025 survey of 308 US executives found 79% saying AI agents are already being adopted somewhere in their company, a number that mostly tells you how loose the word adopted is.

Two academic studies puncture the tidy picture. A February 2025 survey of 410 German HR managers found over 20% of HR departments using no AI at all, while roughly 45% of the professionals in them used tools like ChatGPT informally, with or without permission. And in a US interview study run September to December 2025, 18 of 22 recruiting professionals described using generative AI chatbots in their workflow. The official adoption number and the actual one are different numbers, and the actual one is higher. For how this compares outside HR, the sibling benchmark on how many companies have deployed agents in production covers the whole economy.

Here is the pair of numbers that should be pinned above every AI recruiting pitch. In October 2025, 88% of HR leaders told Gartner their organizations have not realized significant business value from AI tools. The same year, 95% of CHROs had adopted AI and 82% of HR leaders planned to deploy agentic AI within twelve months. Nearly everyone bought it. Nearly nobody can point at the value yet.

BCG's 2024 CHRO survey reads the opposite way: 92% of firms reported seeing benefits from AI in recruitment, with more than 10% claiming productivity gains above 30%. Both surveys are real. What differs is the population, CHROs versus HR leaders broadly, and the bar: seeing benefits is a lower hurdle than significant business value. When two credible sources disagree this hard, the definition is doing the work, and you should ask any vendor which definition their deck is using.

The money is moving anyway. Gartner's November 2025 survey of 469 CEOs found 88% planning to increase AI investment, and its October 2025 CFO survey shows HR budget growth collapsing from 2.4% to a planned 0.7% for 2026, with AI efficiency named as a reason. HR is being asked to spend more on AI and run on less, at the same time, before the value has shown up. Gartner has also measured the sharp end of that bet: among organizations piloting autonomous business capability, about 80% report workforce reductions, and those cuts do not reliably convert into ROI. The pattern matches what the data shows on AI ROI across every function, not just recruiting.

Ask how much faster AI makes hiring and the published answers are specific, impressive, and entirely self-reported. IBM's Blue Pearl case study claims screening time cut 97%, from about 90 minutes per candidate to 3. IBM's Careerforce Pro study claims 85% faster time-to-hire. PwC's Chipotle write-up claims up to 75%. Accenture's HSBC engagement claims 35%, alongside $28.5M in savings and an 18% cut in operating expense. Salesforce's Adecco story claims 28% faster end-to-end hiring and 93% less campaign prep time, and its Capita story claims recruitment compressed from as much as three months to 24 hours.

Five vendors, claims spread across a 69-point range, on the same basic question, and not one discloses a baseline or a method. The spread is the finding: if the technology delivered a stable speed gain, vendors measuring honestly would not land at 28% and 97% for the same job. Somewhere in those numbers is deployment scope, cherry-picked cohorts, and definitions of time-to-hire that do not match. No independent referee exists, because no independent study of AI agent time-to-hire has been published. I looked.

Strip out everyone with something to sell and a different picture emerges: the pipeline design matters more than the pitch.

The closest thing to an independent speed study is a December 2025 experiment that ran 64 real applicants for a Python backend role through a multi-agent screening system and an experienced human recruiter. The AI spent 1.70 hours per qualified candidate found; the human spent 3.33. Real, but 64 applicants and one role, which is a pilot, not a benchmark.

Accuracy is where the independent record gets loud. A 2026 controlled study of 1,000 resume-job pairs found keyword-based screening wrongly rejected 55% of qualified candidates. A semantic pipeline in the same experiment wrongly rejected 8%. Both get sold as AI screening. One throws away more than half of the people you wanted to meet, and nothing on the pricing page tells you which one you are buying. That is an architecture decision, which is exactly the argument for treating screening logic as something you own rather than something you rent blind.

Then the audits. A 2024 study of embedding-model resume retrieval across 500+ resumes and nine occupations found the system favored White-associated names in 85.1% of retrieval cases, and disadvantaged Black male candidates in up to 100% of the settings tested. A FAccT 2026 study of 3 million real applicants screened by one vendor's algorithm found 4% were recommended for rejection from every one of ten jobs they applied to, more than chance predicts. The authors call it algorithmic monoculture: when everyone rents the same screen, being unreadable to one model means being unemployable across all of them. And a study of 332,044 real job postings found the female callback rate ranged from 1.39% to 87.33% on identical postings depending only on which LLM did the recommending. Whatever an AI screen measures, it is not a stable property of the candidates.

While employers were piloting, candidates deployed. Gartner's Q4 2024 survey of 3,290 job candidates found 39% had used AI themselves during the application process. The same survey series found, in Q1 2025, that only 26% of 2,918 candidates trust AI to evaluate them fairly, and in Q2 2025 that 6% of 3,000 admitted to interview fraud, meaning posing as someone else or having someone else pose as them.

So both sides of the hiring table now run AI at each other, and neither trusts the other's. The cost of that arms race shows up in completion data: in a field experiment on a real recruitment platform, run in part by employees of micro1, the platform being studied, 76% of roughly 26,000 candidates invited to an AI-driven video interview never finished it. The interview worked, in the narrow sense that candidates who completed it passed final human interviews at higher rates. But three quarters of the pipeline walked away before the machine finished asking its questions. If your funnel is thin, that trade should terrify you; a tool that improves precision while dropping 76% of candidates is a tool for employers drowning in applicants, not for the ones fighting for them.

The employment effects are past the prediction stage. In Gartner's Q4 2025 survey of heads of HR, 22% of the 110 surveyed reported at least one business leader who has stopped hiring for entry-level roles because of AI automation. In supply chain, 55% of 509 leaders expect agentic AI to reduce entry-level hiring needs. PwC's 2026 Jobs Barometer, built on more than a billion job ads, analyzed 2.4 million US entry-level postings and found the roles that survive are being seniorized: entry-level jobs requiring senior, human-intensive skills grew 35% since 2019 while the rest shrank 10%, and AI-exposed entry-level roles are seven times more likely to demand senior skills.

Recruiters themselves are not disappearing on paper. The Bureau of Labor Statistics projects 6% employment growth for HR specialists from 2024 to 2034, on a May 2024 median wage of $72,910, with about 81,800 openings a year. But the skill premium is repricing fast. McKinsey's analysis of over 11 million US postings found demand for AI fluency grew nearly sevenfold in two years, reaching occupations that employ about seven million workers. Gartner pegs the pay premium for AI-related talent at up to three to four times average worker pay, and, in the same analysis, predicts up to 30% of roles displaced by AI will be rehired by 2029, often at a higher cost than the savings that justified the cut. Salesforce's own survey of 200 HR executives, a vendor number worth reading as one, has HR leaders expecting to redeploy about a quarter of their workforce as agentic AI lands.

A benchmark should also name its bad data, because the recruiting-AI conversation runs on some of the most laundered statistics in the industry.

The claim that 83% of companies use AI for resume screening appears in dozens of vendor blogs and stat roundups. I could not trace it to any primary document, and neither could anyone who cites it, because the citations point at each other. Same story for the claim that around 90% of employers use automated filtering, usually attributed to the World Economic Forum: the attribution circulates secondhand and the underlying document never surfaces. Neither number appears anywhere in this post's dataset, and if a pitch deck leads with either, ask for the study and watch what happens.

Softer failures matter too. KPMG's deployment series, the best quarterly data that exists, prints 42%, then 26%, then 54% across three consecutive quarters, and its own releases disagree on whether Q1 2026 was 54% or 55%. Gartner's much-quoted line that one in four candidate profiles will be fake by 2028 is a forecast, not a measurement, and it gets repeated as if someone counted. And the most instructive case: a rigorous-looking field experiment showing AI-assisted pipelines lift final-interview pass rates by 20 points turns out to have two of its four authors employed by micro1, the recruiting platform being evaluated. The paper is formatted like independent research, standard errors and all. It is a vendor claim wearing an arXiv ID, which does not make it false, but does move it into the same class as the IBM case studies.

Just as telling is what nobody has measured at all. No independent study of time-to-hire change. No primary-source cost-per-hire comparison with and without AI. No independent measurement of quality-of-hire, retention or performance, after an AI-assisted hire; the only published claim is the 35% improvement in workplace cohesiveness in IBM's Blue Pearl case study, with no stated method. The recruiting-specific adoption series stops at two points. For the fastest-growing category in HR software, the independent evidence base is close to empty, and everyone selling into it knows.

Read as a buyer, the 59 numbers collapse into four decisions.

Adoption is no longer the question. A quarter of the large organizations already running agents point them at recruiting and sourcing, and 60% say twelve months. Waiting to see if the category is real stopped being a strategy somewhere in 2025.

Value is not automatic either. The 88% number says most teams bolt an agent onto an unmeasured process and then cannot say what changed. If you do not know your current time-to-hire, screening false-negative rate, or cost per hire, an agent gives you the same ignorance at higher speed. Measure the baseline first; that is the whole reason we run a process audit before any build, because the number you capture before the agent exists is the only honest denominator you will ever have.

Then there is the screening architecture, which decides more than the brand on the invoice. A 55% false-negative screen and an 8% one are sold with the same adjectives. The only way to know which you bought is to test it against your own hiring bar with your own historical candidates, which is precisely the job of agent evals: feed the system the candidates you know you should have interviewed and count how many it throws away, before it ever talks to a live applicant.

And if you run a staffing or recruiting agency, this cuts both ways: the tools are aimed at your margins and available to your workflow in the same quarter. The practical starting list is in which workflows a recruiting agency should automate first. The honest version of this paragraph also says what the data says: if your candidate pipeline is thin, the 76% interview drop-off number means the fashionable AI-interview stack would cost you most of your funnel, and you should not buy it, from us or anyone.