Tech employers are advertising more jobs than they were a year ago, 18% more in the US, and the growth is led by AI roles. The people applying do not feel it, and the reason is more specific than a bad mood: the names for what employers want are changing faster than the systems that read resumes.
The skills were renamed, not lost
In job postings, the phrase "machine learning algorithms" is declining while "agentic AI" is up 805%. Much of the work behind the two phrases overlaps. A person who has built, tested and fixed systems that make predictions has done most of what an agent project needs, but their resume does not contain the new words.
Job seekers have noticed. Nine in ten tech workers think screening software misses qualified people, and most now rewrite their resumes for it: 75% of women and 60% of men. That makes the problem worse. When everyone pastes in the same fashionable phrases, a phrase stops telling the employer anything, and the person who describes their real work plainly looks weaker than the person who copied the posting. We already reported that more than half of AI postings do not match their own job titles, so the posting and the resume are both drifting away from the work itself.
Last time it was credentials, this time it is vocabulary
The 2021 problem could in principle be fixed once, by removing requirements such as a degree that did not predict success. A vocabulary problem cannot be fixed once. If "qualified" is defined by a list of terms, the list has to be rewritten whenever the field renames itself, and in AI the names keep changing. Any employer that treats its keyword list as finished is falling behind a little every month.
The exposed employers are not the tech companies, which have people who know the new words. They are the ones now hiring fastest outside the usual hubs: insurance postings for tech roles are up 69% in six months, and manufacturers and consultancies are growing too. These companies often rely on job-board filters and vendor defaults.
What a better screen looks like
Take a regional insurer with 400 applicants for one agent-building role. Reading each resume for two minutes is about 13 hours of work, so in practice a keyword filter does the first cut. Instead, the recruiter could give a language model a plain description of the work ("a system that reads claims and decides what to do with them, and someone who keeps it running") and ask it, for each resume, two things: what has this person actually built, and which part of this job does it cover? The recruiter then reads the best 40 in under two hours and gives the finalists a short practical task. A person who wrote "machine learning automation" is now in the pile, and a person who only wrote "agentic AI" has to show they did something with it.
The model's reading is a first pass, not a verdict. Its job is to widen the pile, and a person still decides who is hired.
The same thing will happen in any field where the tools get new names every year, which by now is most of them: the bookkeeper who "reconciled accounts" and the one who "supervises automated matching" may do nearly the same work.
If you are the one applying, write each project as one plain sentence: what you built, for whom, and what changed. Then put the current name for it in brackets. A person or a model reading it will understand the work, and a keyword filter will still find the word.