Why AI Is Coming for the Educated First
March 23, 2026
Last month, a software engineer in San Francisco described his job to a reporter. "I'm basically a proxy to Claude Code," he said. "I relay my manager's instructions to the AI, review what comes back, and hit submit."
He still has a job. He still gets a paycheck. He still shows up to an office in SoMa and sits in front of a screen for eight hours. But the thing that made him a software engineer — the thinking, the problem-solving, the writing of code — isn't his job anymore. The AI does that. He's a go-between. A human API.
He's one of the lucky ones.
Here is what has already happened. Not what might happen. Not what forecasters predict. What is happening right now, in 2026, while most people are still debating whether AI will "eventually" affect their careers.
Microsoft's Chief Executive Officer (CEO) estimated that 20 to 30 percent of the company's code is now written by Artificial Intelligence (AI). In the same period, over 40% of Microsoft's layoffs targeted software engineers. Those layoffs had multiple causes — post-pandemic over-hiring, strategic reorganization. But the direction is unmistakable: the machines are writing more of the code, and fewer humans are needed to supervise them.
A Stanford Digital Economy study found that employment for software developers aged 22 to 25 declined nearly 20% from its 2022 peak by mid-2025. The entry-level pipeline is drying up. Junior developers aren't being replaced by AI. They're simply never being hired.
But software engineering is supposed to be the safe career. The one guidance counselors recommend. The one that "AI can't touch." If the engineers who build AI are being displaced by it, what does that tell you about everyone else?
It tells you plenty. Look around.
Talk to a freelance translator. Not about the future — about last Tuesday. Their income has been cut in half. Not because translation work disappeared — there's more multilingual content than ever. But AI does the first pass now, and the human's job has been reduced to "post-editing" machine output. They sit at a desk cleaning up AI translations that are almost right but subtly wrong — fixing idioms the model flattened, restoring nuance the algorithm couldn't hear. They describe spending more time fixing AI output than it would take to translate from scratch. They're paid half as much for work that is, in many ways, harder. And they know — they can feel it — that "post-editor" is a transitional job title. The AI gets a little better every quarter. The gap between "almost right" and "right" is narrowing. When it closes, the human in the loop becomes the human out of work.
This is not a story about translators. This is the template. Law firms are replacing entire research teams with software subscriptions. Content writing jobs are projected to decline 50% by 2030. In every case, the pattern is identical: AI doesn't walk in and fire everyone on a Monday morning. The team of twelve becomes a team of six, because each person is now "augmented." The job posting that would have gone up doesn't. The contractor whose contract would have been renewed isn't.
The jobs aren't being eliminated. They're not being created.
Anthropic — the company that builds Claude, one of the most capable AI systems on Earth — published its own research in March 2026. Take their numbers with the obvious grain of salt: a company that sells AI has an interest in making AI look transformative. But the data is granular enough to be useful, and no one else is publishing anything this specific.
They measured two things: what AI could theoretically do in each profession, and what it is actually doing right now. The gap between those numbers is the most important data point in the economy.
For business and finance professionals, AI can theoretically handle 94.3% of tasks. It is currently handling about 28%. For software engineers: 94.3% theoretical, 35.8% actual — meaning two-thirds of what a developer does today could already be automated but hasn't been yet. For legal professionals: 89% theoretical, and climbing. Across every knowledge profession they measured, the same pattern: AI can already do most of the job. It just hasn't been asked to yet.
That gap — between what AI can do and what it is doing — is not a comfort. It is a warning. It is the wave that hasn't hit yet. It is 2007, and the mortgage-backed securities are still rated triple-A. Everything looks fine because the collapse hasn't arrived. But the math is already done.
Ninety percent of companies haven't even deployed generative AI in production yet. When they do — and they will, because their competitors will force them to — that 28% becomes 50%. Then 70%. Then the theoretical ceiling.
Who gets hit first — and why it inverts everything you think you know — that part's for the book.