Something in the hiring market has stopped adding up. According to recent data1, drawn from thousands of employers and 691 million candidates, job openings across Europe were up 12% year on year as of July 2026 and applications were up 6%, while hires were down 1%.
Employer demand and candidate interest are returning, and AI now sits inside nearly every stage of the recruiting process. By every input we track, hiring should be getting easier, yet the output has barely moved. If AI is helping organisations process more candidates than ever, it is worth asking why fewer of them are being hired.
My explanation is that the industry has aimed a genuinely capable technology at the wrong objective. For years we have pointed AI at a matching problem, when the thing we are actually paid to improve is the quality of hiring decisions made under uncertainty, and a decade of throughput metrics has hidden the difference.
The arms race nobody meant to start
The pitch for hiring AI has always been straightforward. Applications arrive faster than any human can read them, so let the system score every candidate against the job description, sort the list and trim the bottom. The volume pressure behind that pitch is real, and the intent is entirely reasonable.
The difficulty is that both sides of the market adopted the same tools at the same time. While employers deploy AI to filter applications down, candidates deploy AI to push applications up; between 40% and 80% of job seekers now use AI to write or polish their CVs, according to SHRM2. Two systems are now optimising against each other, CVs converge on the same polished template, and the signal a recruiter relies on to separate a strong candidate from an average one degrades.
This is worth being precise about. A well-written CV has never proved that somebody can do a job; it contains evidence about their experience and capability that still needs to be assessed and, where it matters, validated later in the process. Generative AI widens the gap between presentation and underlying evidence, which means textual similarity between a CV and a job description can be useful information but cannot be the objective in itself. The technology performed exactly as instructed; the instruction was to reward polish at volume, and polish at volume is what we got.
Screening takes too much on trust
Screening is now the single most common use of AI in hiring; research3 shows 58% of organisations use it, more than any other application in talent acquisition. Most automated screening rests on three assumptions, and none of them holds.
The first is that the job description accurately captures what the role requires. In practice, job descriptions are written by several people, copied from earlier recruitment cycles, and rarely capture the qualities that determine whether someone succeeds in a particular team, with a particular manager, facing a particular set of challenges. Hiring managers themselves often refine their view of the role after meeting the first few candidates.
The second is that a CV accurately captures who a candidate is. It never did, and it was never meant to. A CV is evidence about a person, condensed into a few pages and increasingly assisted by AI, rather than validated knowledge of what they can do.
The third is that a model can measure the gap between those two artefacts cleanly and without bias. Models learn from human decisions, human decisions contain bias, and unless systems are designed, monitored and audited against that reality, the bias is carried forward at scale.
Feed three noisy inputs into a system built to draw a hard line, and error stops being the exception and becomes part of the design. That is a solvable design problem, and solving it starts with being honest about the noise.
Hiring was always a probability problem
Picture the job description as a point in space and every candidate as a point around it. If the description were a perfect proxy for the role and every CV a perfect proxy for the person, hiring would be trivial: pick the nearest point, send the offer, skip the interview, and time to hire falls to zero. Nobody wants that outcome, and no candidate would trust it, which tells you the perfect-match framing was never the real goal.
In reality, both points are blurry and both of them move. The requirement shifts as interviews reveal what the team actually needs, and the candidate ranked fourth on paper is regularly the obvious hire once you meet them. Any system that draws hard lines through that uncertainty is optimising against an illusion of certainty.
This is precisely the class of problem AI handles well when we ask it to. A strong model can identify candidates whose experience appears relevant and prioritise them for attention, which is extremely valuable when there are hundreds of applications. It cannot establish from the CV alone that somebody possesses every claimed skill, will perform well in the role or should be hired.
Those are different questions, they require additional evidence, and a well-designed system treats them as such: it exposes its uncertainty rather than hiding it, directs scarce human judgement to where it changes the outcome, and leaves the judgement itself with the human.
Ranking respects the noise
Screening assumes the data on both sides of the line is clean, draws a threshold and discards everyone beneath it, and once a candidate is cut that decision is rarely reversed. Every flaw in the underlying data becomes permanent at the moment the line is drawn. Small errors in the requirements, the candidate information or the model become consequential the instant they are converted into a hard decision.
Ranking uses the same information differently. It starts from the more honest premise that job descriptions and CVs are the noisy artefacts they are, surfaces the most likely fits towards the top and keeps every candidate available for consideration. A recruiter works down a list ordered by likely fit, inspects the underlying evidence, challenges the model where something looks wrong and stops when the quality clearly drops off, rather than trusting an arbitrary cutoff. That saves real time, and it protects against the most expensive mistake in recruiting, which is removing the right person before anyone has looked at them.
Ranking is not automatically correct, and it deserves the same scrutiny screening should have received. Position in a ranked list determines who receives attention, so these systems need rigorous evaluation of their own: whether qualified candidates consistently appear where recruiters will see them, how performance varies across different jobs and populations, and whether the ordering produces unjustified differences between groups.
The claim for ranking is narrower and more defensible than the claim screening made. It uses an uncertain prediction as information that helps a person decide where to look, rather than as a substitute for the decision.
Evidence, validation and genuine oversight
The same principle extends beyond the shortlist. AI can help a recruiter see which requirements matter, which candidate evidence relates to them and where the picture remains incomplete, so the interview tests the things that actually matter rather than treating the CV as established fact. The model contributes evidence to a decision; it does not manufacture certainty about it.
This is also where responsible AI becomes something you operate rather than a value you declare, and where I would push back on the industry’s favourite reassurance. Putting a person somewhere in the workflow is not sufficient on its own. Genuine oversight requires evaluation for fairness, continuous monitoring, evidence that can be audited, clear governance and a real ability for people to challenge or override model output.
Regulation across Europe and in jurisdictions such as New York City and Colorado is increasingly formalising exactly these obligations. Built properly, these controls reinforce the quality of the hiring process rather than sitting alongside it as a compliance exercise, and the human stays genuinely in control of the decision rather than nominally on top of it.
Measure the decision, not the throughput
For a decade this industry has celebrated efficiency: CVs reviewed per hour, cost per hire, time to fill. Those numbers are worth watching, and they have now been optimised to the point where they tell us very little. Openings and applications are climbing while hires stay flat, which is the clearest available evidence that processing speed has stopped translating into hiring outcomes.
The distinction that matters is between intermediate measures and the outcome itself. Speed, model accuracy, recruiter adoption and cost tell us how a system is performing; fairness and responsible AI measures provide essential guardrails. The outcome the technology exists to improve is whether organisations make good hires, and that is measurable with the same discipline we applied to speed: stage-to-stage pass-through rates for qualified candidates, offer acceptance, quality of hire at six and twelve months, and adverse impact monitored continuously rather than audited annually.
Organisations that instrument those numbers will know whether their AI is finding the right people. Those that do not will keep getting faster at the wrong thing.
Used well, AI makes hiring teams faster, and it does something more valuable at the same time: it makes them faster towards the right person. That is where the technology is genuinely capable today, and what responsible recruitment should look like.
References
1 https://www.icims.com/en-gb/company/corporate-newsroom/research/
2 https://www.shrm.org/topics-tools/news/hr-trends/recruitment-is-broken