Once a company’s AI runs on its own, improving the product, the funnel and the support queue around the clock, most of today’s roles won’t be competing with a tool. They’ll be competing with something that runs a thousand experiments an hour. What happens to the people is the question I keep coming back to.
I build with these models every day. I’ve watched work that took a team two weeks of planning, specs, handoffs and testing collapse into minutes of an agent’s time. When leadership can ask a system directly for what used to take a department months, the layers in between start to look like cost, not capability.
So I’ll say the uncomfortable part plainly: a lot of jobs, as people know them today, are going away. Not all of them, not all at once, and the data so far is quieter than the headlines. But the direction is clear, and waiting for it to be obvious in the unemployment rate means waiting too long.
The first effect isn’t mass layoffs. It’s that the entry-level door is narrowing in the jobs AI does best. Stanford’s payroll research found that workers aged 22 to 25 in the most AI-exposed jobs saw a 16% relative drop in employment, while older workers in the same jobs didn’t. The Dallas Fed traced it mostly to fewer people being hired in, not more people being let go.
Layoffs are catching up. Companies cited AI in about 116,000 announced U.S. job cuts through August 2026, roughly 22% of all cuts, up from about 55,000 and 5% in all of 2025.
Change in employment since late 2022, software developers at the same firms, by age.
Stanford HAI 2026 AI Index, from ADP payroll data, as reported in arXiv 2605.01160. Older-worker growth shown at the midpoint of the reported 6–12% range.
Both things are true at once. The aggregate numbers are still small, and a growing number of companies are telling investors, out loud, that AI is why they need fewer people. The second trend tends to lead the first.
History is the best argument for calm. About 60% of the jobs Americans held in 2018 didn’t exist as job titles in 1940. New work keeps appearing, and it usually does. But the same research found automation erased work about twice as fast after 1980 as before, and its authors are clear that there’s no law saying new work arrives one-for-one.
What’s different now is breadth. The old advice was: if a machine takes your desk job, move toward physical work or care. Physical work is automating in parallel. Waymo went from 50,000 paid rides a week in May 2024 to about 500,000 in early 2026, with a goal of a million by year end. Humanoid robots are in warehouse and factory pilots, and a single Chinese factory can now build thousands of them a year.
My assessment, informed by BLS projections and exposure research. Toggle the horizon to see physical automation catch up.
Horizontal: share of the job’s tasks current AI and robotics can do. Vertical: how much the job depends on physical presence, dexterity, trust or care. Dot size: rough U.S. employment. Projections for 2034 from BLS, July 2026.
What looks most protected: work where the value is a person being there and being trusted. Nursing, home health and personal care, skilled trades in messy one-off environments like old houses and field repair, teaching young children, therapy, and roles that carry legal accountability. BLS projects healthcare and social assistance to add about 2 million jobs by 2034, the most of any sector. It’s not glamorous, and much of it pays less than the desk jobs it would replace. That gap is the real problem.
When the engine runs, the org chart flattens from the middle. Agents take the execution work first, then the coordination work that existed to manage execution. What’s left is a thin layer of people who set direction, own the risk, and pay for the compute. The budget line that grows is tokens.
An illustration, not a forecast. Drag to see the org compress as agents take on more of the work.
Public companies feel the most pressure to follow that curve all the way down, because investors reward it. Some leaders will keep people on out of loyalty for a while. I don’t think that holds once it’s obvious there’s nothing productive for those roles to do. Kindness without work isn’t a plan for anyone.
The question isn’t whether companies get smaller. It’s whether the gains stay with the few people who own the engine, or reach the people it replaced.
Two composite people, not real individuals, in two of the most exposed jobs in tech. Step through their next two years, and what I’d tell each of them to do now.
None of these is new, and each has real costs. Several are already being proposed in Congress and by the AI labs themselves. Open each to see what it does, what it would cost, and what the evidence says.
These are my views on a contested question. People I respect disagree, especially on costs and on how fast any of this arrives.
Speed. The aggregate data still says “early.” Adoption inside most companies is uneven, and a large Danish study found AI reshaping tasks without yet changing employment or earnings. Tokens may get expensive enough, for frontier work running around the clock, that some human work stays cheaper for longer than I expect. And new work does tend to appear in places nobody predicted.
But I’d rather prepare for the version where the engine works, and be pleasantly wrong, than the other way around.