Josh Bersin spent an episode this month on a market he knows better than almost anyone, and titled it “The Messy World of AI-Powered Recruiting Where Nobody Is Happy.” That is a striking verdict on a category that has, by any adoption measure, already won. Recruiters have the tools. Candidates are being processed by them. And the people who signed the invoices are still waiting for the part where hiring visibly improves.
The numbers now say the same thing the practitioners do. Everyone bought in. Almost nobody can prove it worked.
How many companies actually got value from AI in hiring?
Research from Everest Group, commissioned by ManpowerGroup Talent Solutions, surveyed 80 C-suite, CHRO and senior talent acquisition leaders across the United States and the United Kingdom in healthcare, life sciences, manufacturing and technology. It found that more than 90% of organizations have deployed AI in talent acquisition, while fewer than 5% describe their outcomes as transformational on any key metric.
Look one level down and the shape of the gap becomes clear. Thirty-nine percent report a significant impact on operational efficiency — the clearest measurable gain anywhere in the study. Improvements in decision quality, workforce agility and strategic capacity stay stubbornly moderate. In plain terms: the work got faster, the hiring did not get better.
That is not a failure of the software. It is a description of what the software was pointed at.
Why does adoption keep outrunning results?
The researchers name a structural cause rather than a technical one: most organizations are layering AI onto workflows built for a pre-AI environment. Isolated tools, siloed data and unchanged hiring processes stop any individual gain from compounding into a result anyone outside the recruiting team can feel.
The barriers leaders cite line up with that reading — change management and adoption at 58%, governance and compliance at 55%, data readiness at 55%. Notice what is absent from the top of that list: model quality. Nobody's central complaint is that the AI is bad at reading a résumé.
There is a second-order effect worth sitting with, because it is new. Fifty-four percent of organizations say AI-assisted candidate behavior — generated résumés, generated applications, coached interview answers — is making it harder to judge whether a candidate can actually do the work. Automation arrived on both sides of the table. One side got tools to apply faster; the other got tools to screen faster; the signal in between got worse.
Is anyone measuring whether it worked?
Mostly, no. SHRM's State of AI in HR 2026, drawn from 1,908 HR professionals across 138 distinct HR tasks, found recruiting to be the single largest AI use case in HR at 27% of organizations. It also found that 56% of HR functions do not formally measure the success of their AI investments, and only 16% use ROI as a metric.
Set that against the same study's finding that 87% of HR professionals report improved efficiency, and the two facts stop contradicting each other. Efficiency is the thing you can feel without instrumenting anything. It is fewer hours on scheduling, fewer résumés read at 9 p.m. Real, and worth having. But a felt improvement is not a measured one, and a measured one is what survives a budget review.
This is the uncomfortable middle the whole category sits in. Adoption is near-universal, satisfaction is moderate, measurement is rare, and so the honest answer to “did it work?” is usually “we think so.”
What does the gap look like inside a technology operating company?
It looks like a scoreboard nobody outside recruiting believes. A CIO or VP of IT Operations has twelve open requisitions — embedded and firmware engineers, test and QA, network and SRE, field-service technicians, a data-center operator, someone who genuinely knows the ERP — and a recruiting function reporting that time-to-fill improved 18%. Those two facts do not touch. The roles are still open. The projects still slip. The improvement was real and invisible at the same time.
Meanwhile the market underneath got harder, not easier. Reuters reported that in the UK, job postings fell 11% between the start of 2026 and mid-July and sat 32% below pre-pandemic levels, while demand for AI skills hit a record high. Fewer openings, sharper competition for specific capability. Precision matters more than throughput, and throughput is exactly what most deployments optimized.
What should a talent leader measure instead?
Three things, and pick them before switching anything on rather than after.
One outcome metric tied to the business, not the funnel. Ramp time to productivity. Retention at twelve months. Revenue or uptime contribution at six. Time-to-fill is an input dressed as a result; it improves when you lower the bar, which is the opposite of what anyone wants.
Coverage, not activity. Not “how many candidates did we contact” but “what share of the qualified market for this role did we actually reach, and where did we lose them.” A shortlist of eight from a pool the software never fully saw is a guess with good formatting.
Where the process stops. Every handoff between tools is a place work waits for a human to notice. Instrument those seams and you learn whether you bought automation or bought a faster way to generate queues. Bersin's own research with AMS describes the recruiter's role shifting from processor to strategic orchestrator — which only works if there is a system underneath doing the processing, end to end, rather than nine tools handing each other homework.
Where UPPER stands
Our position has not moved, and this month's data is the clearest argument for it we have seen. The individual tools in recruiting work. Sourcing works. Screening works. Outreach works. What nobody automated is the process between them — and that gap is precisely where the 90%-deployed, 5%-transformational number comes from.
So UPPER runs the whole sequence as one system, and does it for the operating companies that never had a recruiting department to begin with. A job description goes in. Twenty minutes later a ranked shortlist comes out. The outreach then runs itself until candidates reply with CVs attached — and the team's own recruiters and hiring managers operate it directly, in the US and Canada, insourcing work that used to be handed to an agency and never measured.
The instrumentation is the point as much as the automation. Every channel's yield, every stage's fall-off, cost per qualified conversation, where the process stalled and why — recorded, because a talent leader who cannot answer “did it work?” with a number is going to be asked that question again next quarter, with less patience.
Adoption was the easy half, and the market has finished it. The half that decides whether any of this was worth the money is proving value — which means automating the seams and measuring the outcome, not the activity.