Computer systems design employment fell five percent since ChatGPT launched. Wages in the same industry rose almost seventeen percent. The Dallas Fed found the mechanism: AI replaces what you can learn from a textbook and amplifies what you can only learn from experience.
The Dallas Federal Reserve published a research paper on February 24, 2026, titled "AI is simultaneously aiding and replacing workers, wage data suggest." The title contains the finding that most coverage missed: the word simultaneously.
In computer systems design — one of the industries most exposed to AI — employment has fallen five percent since ChatGPT launched in the fall of 2022. In the same industry, over the same period, nominal average weekly wages increased 16.7 percent. The national average over the same window was 7.5 percent.
Fewer people. Higher pay. Same industry. Same time period.
The Mechanism
The Dallas Fed identified the variable that determines which side of the divide you land on: experience.
AI substitutes for entry-level workers whose value comes from codifiable, book-learned knowledge — the tasks you can describe in a manual, the procedures you can document in a runbook, the work that follows explicit rules. Data entry. Level 1 support tickets. Unit test generation. CRM management. These are tasks a recent graduate performs while learning the actual craft.
AI augments experienced workers whose value comes from tacit knowledge — the judgment and intuition built through years of hands-on work. Pattern recognition that you cannot articulate as a rule. The ability to know which of three technically correct solutions will fail at scale. The instinct that something is wrong before the metrics confirm it.
The returns on job experience are increasing in AI-exposed occupations. Not staying flat. Not declining. Increasing. Experience was always valuable. Now it is becoming the scarce factor.
The Pipeline Problem
This creates a structural paradox. The codifiable tasks that AI replaces are the same tasks that junior workers perform while acquiring the tacit knowledge that makes them valuable later. Data entry teaches you what clean data looks like. Level 1 support teaches you what users actually struggle with. Writing unit tests teaches you where code breaks.
These are apprenticeship tasks. They are repetitive, often tedious, and not the point of the job. They are the thing you do while you are becoming the person who no longer needs to do them. AI can perform the tasks. It cannot perform the becoming.
Anthropic published its own labor market research the same week. Computer programmers face seventy-five percent task coverage — meaning AI can theoretically handle three-quarters of what a programmer does day to day. But actual adoption, measured by workplace usage data from Claude, is "just a fraction" of theoretical capability. The gap between what AI can do and what organizations have deployed it to do remains wide.
The Anthropic report named the scenario explicitly: a "Great Recession for white-collar workers." Their data showed "suggestive evidence that hiring of younger workers" — specifically ages twenty-two to twenty-five — "has slowed in exposed occupations."
The Dallas Fed data and the Anthropic data converge on the same population: young workers entering fields where AI already covers most codifiable tasks. Not displaced yet in the aggregate statistics. But the hiring pipeline is narrowing.
The Macro Signal
On March 6, 2026, the Bureau of Labor Statistics reported that the U.S. economy lost ninety-two thousand jobs in February. This was the first negative payroll print since the pandemic. The market had expected a gain of fifty thousand.
Fortune published an analysis the next day arguing that companies may be cutting workers to fund AI infrastructure, not to replace workers with AI already in production. Global AI capital expenditure is expected to reach two and a half trillion dollars in 2026, up forty-four percent over 2025. The money has to come from somewhere. Workforce CEO Carl Eschenbach declared that February layoffs of 1,700 were necessary to "prioritize AI investment and free up resources."
Amazon accounts for fifty-two percent of all tech layoffs in 2026 — sixteen thousand corporate roles cut in January, with estimates of thirty thousand by May. The company posted record revenue of $716.9 billion in 2025. It is not struggling. It is restructuring. The layoffs target management layers specifically — flattening the organizational hierarchy while pouring two hundred billion dollars into AWS and AI infrastructure.
Microsoft's fifteen thousand cuts targeted what internal reports called "human middleware" — Azure Level 1 and 2 support, sales operations, code maintenance. The roles being eliminated are the precise categories that AI agents have been trained to perform: customer support troubleshooting, CRM data management, unit test writing, legacy code updates.
The pace across the industry is eight hundred and fifteen people per day, up from six hundred and seventy-four per day in 2025.
The Readiness Paradox
Deloitte published its State of AI in the Enterprise report in early March 2026. The headline finding: enterprise AI readiness has declined year-over-year. Strategy readiness at forty percent, governance at thirty, technical infrastructure at forty-three, data management at forty, and talent readiness at twenty percent. All down from the previous year.
Access to AI tools expanded fifty percent — sixty percent of employees now have access. Fewer than sixty percent of them use the tools regularly. Only twenty-five percent of organizations have converted forty percent or more of their AI pilots into production systems. Just 8.6 percent have AI agents deployed in production.
Companies are cutting workers faster than they are deploying the technology that would replace them. The Deloitte data quantifies what the Dallas Fed data implies: the workforce reduction is running ahead of the capability deployment. The apprenticeship pipeline is being dismantled before the alternative is built.
Two Futures in the Same Number
The five percent employment decline and the seventeen percent wage increase are not contradictions. They are the same phenomenon measured at different points in a career.
For someone entering computer systems design at twenty-two, the landscape has contracted. The codifiable tasks that once served as an on-ramp — the entry-level work that taught the craft through repetition — are being absorbed by systems that perform them faster, cheaper, and around the clock. The hiring slowdown in AI-exposed occupations for workers aged twenty-two to twenty-five is the leading indicator.
For someone with fifteen years in the field, the landscape has expanded. Their tacit knowledge — the judgment, the intuition, the pattern recognition that resists codification — has become the scarce complement to AI's abundant codifiable capability. The wage premium is the market pricing this scarcity.
The question neither the Dallas Fed nor Anthropic answered: where do the experienced workers of 2036 come from, if the apprenticeship pipeline of 2026 is being dismantled?
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)