What AI Has Actually Changed About Hiring
Hiring Technology
AI did not make hiring faster. So far it has mostly made the pile bigger.
The clearest number on this comes from LinkedIn, which now sees around eleven thousand job applications submitted every minute. That is up roughly 45% in a year. The number of jobs did not go up 45%.
Something had to absorb that, and what absorbed it was the time of the people on the other side. CNBC quoted recruiters in late 2025 describing the experience as drinking through a fire hose, with individual roles taking three to five hundred applications inside a weekend. Greenhouse put the increase in recruiter workload at 26% in a single quarter.
| Reported figure | |
|---|---|
| Applications submitted on LinkedIn, per minute | ~11,000 |
| Year on year change in application volume | +45% |
| Applicants a popular role can collect over a weekend | 300 – 500+ |
| Change in recruiter workload in one quarter | +26% |
| Job seekers who now apply in bulk | 38% |
The arms race nobody chose to enter
Follow the sequence, because every step in it is individually sensible.
A candidate who is getting nothing back starts using AI to write applications, because writing forty by hand was already unpaid evening work and now it takes an afternoon. The employer, receiving four hundred applications instead of ninety, starts using AI to cut that down to something a person can read. The candidate, hearing back even less often than before, concludes that the answer is more volume, and sends two hundred.
Both parties are behaving rationally. The outcome is worse for both. The employer is now reading text that was not written by the candidate, and the candidate is being filtered by software that will never explain itself. The interesting thing about the last two years is not that AI arrived in hiring. It is that it arrived on both sides at once, and mostly cancelled itself out.
Around eleven thousand applications are now submitted every minute, up about 45% in a year. The number of jobs is unchanged.
What the technology is genuinely good at
None of that means the technology is useless here. It means it has been pointed at the wrong step. Three things it does well, and one it does badly:
Reading documents into structure. Turning an unstructured work history into fields, dates, certifications and equipment is exactly the kind of dull, high-volume task machines are better at than people. This part works.
Comparing a requirement against stated facts. A shift needs someone with a current forklift ticket, inside forty minutes' travel, available Sunday nights, at or above a certain rate. Checking that against a large set of profiles is arithmetic, and it is arithmetic no human does well at three in the morning when someone calls in sick.
Handling the conversation around the edges. Answering "where do I park", "what do I wear", "when am I paid", at nine at night in the applicant's own language. That is real work being done, and nobody misses doing it.
Judging who is worth talking to, invisibly. This is the one to be careful with. The moment software is deciding who a person gets to be considered for and cannot show its working, you have built something the candidate cannot argue with and you cannot audit either. Most of the distrust people now carry into hiring processes comes from exactly this, and we have written about why that distrust is rational.
Matching is a different shape from filtering
The distinction that matters is not human versus machine. It is whether the pile gets created at all.
Filtering starts with everyone and removes. It requires a mass of applications to exist first, which means someone had to write them, which means the fire hose. It rewards volume on the candidate side and produces exhaustion on the employer side.
Matching starts from the requirement. The worker has already said what they hold, where they will travel, which shifts they can work, and what they will accept, once. When a job needs filling, the system finds the people who fit and shows them the job with the reasons attached. No four hundred applications were written, so no four hundred applications need discarding.
The reasons matter as much as the mechanism. What a worker should see is the specific things about them that lined up: the certification, the distance, the shift pattern, the rate. Not a number, not a placing, not a grade. A person can check a list of reasons and correct it if something on it is wrong. Nobody can argue with a score, which is precisely why scores are the wrong output.
If you are hiring
Look at where your time goes. If most of it is spent removing people from a list, the volume is the problem and better removal tools will only ever tread water. Ask instead what you actually need to know before someone is worth a call, and whether that could be known in advance rather than collected twenty-five fields at a time.
If you are looking for work
The advice that used to work, which was apply to more things, is now the advice everyone received. Two hundred AI-written applications compete against each other and land nowhere. The thing that is actually scarce is verified, specific, current detail about you: the ticket with the real expiry date, the honest travel radius, the shifts you can genuinely do, the rate you will accept.
That is boring to fill in. It is also, increasingly, the only part of your application that a machine cannot generate for somebody else.