Nobody knows how AI will affect jobs over the next ten years. Not the people saying it will create more jobs than it destroys. Not the ones saying most knowledge work is already finished. Not the CEOs writing the memos that announce the cuts. They’re guessing too, and they have a quarterly number to hit while they guess.
Both camps are selling something. “AI is coming for your job” sells clicks. “AI will create millions of new jobs” sells conference tickets. Confidence is easier to package than uncertainty, and it travels better.
Also notice what “stated reason” actually means when companies cite AI. Through September 2026, US employers announced 573,195 job cuts. Artificial intelligence was the stated reason in 120,136 of them — about 21% of the total, and the most-cited reason of the year, according to Challenger, Gray & Christmas. In September alone, AI was the fifth most-cited reason: 3,961 cuts, about 9% of the month.
That’s the whole story in one line, and almost nobody reports it that way.
Stated reasons are what a company writes in a press release, not a measurement of whether software did the work. Some of it is real substitution. Some of it is a company that needed to shed headcount and picked the word that made the stock move. That’s the same incentive I wrote about in Why Profitable Companies Still Lay People Off, and it’s why that piece is about incentives rather than villains.
What we can actually observe
AI is genuinely good at some narrow, well-defined tasks. Summarizing. Drafting a first pass. Pattern-matching across a large pile of documents. Completing code in patterns it has already seen a thousand times.
IBM’s AskHR assistant reportedly resolves 94% of routine HR queries. The remaining 6% — the ones needing judgment, or a person who can be blamed when it’s wrong — still go to a human.
It remains weak at accountability, at novel judgment, at anything physical, and at anything where being wrong is expensive. Microsoft’s chief people officer said in July that AI was changing how work gets done, then cut about 4,800 roles and said the eliminated jobs were “not being replaced by AI.” Both halves of that sentence are true.
The cleanest evidence of how unresolved this is comes from METR, which runs randomized trials on real work. In 2025 they gave 16 experienced open-source developers 246 real tasks from repositories they maintain, randomly assigned to AI or no AI. With AI, the developers took 19% longer. They believed they’d been 20% faster. Outside experts had predicted 39% faster.
Then METR ran it again with newer tools and more developers. The slowdown disappeared.
Same researchers, same method, better models, opposite result. If a careful controlled trial can’t tell you which way the sign points, an opinion piece certainly can’t. And that first result is a snapshot of early-2025 tools — evidence of how fast the question moves, not of where it lands.
Companies follow the math, not the prophecy
None of which means AI won’t take work. It means companies adopt it where the economics work — cheaper, or faster for the same output — and skip it where they don’t. That’s the whole decision rule, and it isn’t ideological.
Ford is a useful example. After leaning on automated quality systems that weren’t catching enough, the company brought back 350 veteran engineers over three years. Not to redo the work by hand. To train younger staff and reprogram the tools that were failing. The tools stayed. The judgment came back.
That’s a more common shape than either camp admits. The task gets cheaper, the job changes, and the people who stay useful are the ones who can tell when the output is wrong.
The history is worth one paragraph, not a speech
Every automation wave gets called unprecedented. ATMs handled the cash and deposit tasks bank tellers used to do. Between 1988 and 2004, the average urban branch needed 13 tellers instead of 20 — and urban branches grew 43%, so teller jobs didn’t fall. The job changed, the headcount held. James Bessen documented it.
Don’t overread that. Demand elasticity wears out — Bessen says so himself, noting that power looms automated 98% of the labor in weaving and weaving jobs rose, until eventually they didn’t. “This time is different” has been said before, sometimes right about the details and wrong about the timeline.
What the history actually supports is weaker than either side wants: most past waves changed the shape of work more often than they deleted whole occupations overnight.
Two failure modes
Panic. Treating every headline as a personal countdown. Nothing improves your odds, and it costs you the years you’d have spent building the thing that does.
Complacency. Assuming your current skills are safe because “they’ll still need people.” Companies don’t keep roles because the work is meaningful. They keep them because the work is worth more than the salary.
Both are predictions dressed up as strategies. Neither is a plan.
What actually helps
Keep working. Show up, do the job well, build a record. Output and reliability still matter this year. Being good at the work still helps more than watching the headlines.
Keep learning — narrowly. Not “learn AI.” Use the tools that touch your actual work. Notice where they’re good and where they quietly fail. That judgment transfers. A certificate doesn’t.
Build a cushion. Same advice as the layoffs piece. Keep three to six months of essential expenses in cash. Income gets interrupted for a hundred reasons — a bad manager, a broken industry, a bad knee. AI is one of them. It isn’t the most likely one.
Own something outside your paycheck. An index fund, a side project, a network. Not because of AI. Depending on one employer has always been the real concentration risk.
What “learning” means here
Not a six-month bootcamp in machine learning, unless that’s already your field.
Yes to: using current tools for the parts of your job they genuinely help with, reading how people in your role are experimenting, and updating your view of what’s automatable versus what still requires you.
The goal isn’t to become an AI expert. It’s to avoid being the last person in the room to notice that a task you spent six years on is now cheap.
The honest close
We don’t know the endpoint. That’s the real answer to “will AI take my job,” and it should be the first thing anyone tells you.
We do know two things that compound either way: work you can point to, and a working understanding of the tools in your field.
Nobody needs a forecast to act on today.
