Nobody Realizes What's About to Happen: Why AI Agents Hollow Out White-Collar Work

Amazon just laid off roughly 16,000 people. At Davos, CEO Andy Jassy stood in front of the world's business elite and said the line every white-collar worker needs to hear without the PR varnish: as Amazon rolls out generative AI agents, it will need fewer people doing some of the jobs done today. Not might. Will. This is not a cash-crisis story. Amazon raised 2025 capital-expenditure guidance to about $125 billion—the highest among the mega-cap tech companies—with more expected in 2026. They are investing in AI instead of headcount.

If you think this is an Amazon-only restructuring, look at what CEOs say when they stop performing for PR. Ford's Jim Farley said AI will replace literally half of all white-collar workers in the United States. Salesforce's Marc Benioff cut support staff from about 9,000 to 5,000. Asked why, he did not soften it: he needs less heads. Two months earlier he had been publicly arguing AI would not cause huge mass white-collar layoffs. Then he fired about 4,000 people. Klarna shrank from roughly 5,500 employees to 3,400—about 40%. Its AI assistant now handles about 2.3 million customer conversations a month, work that used to take around 700 human agents.

The sharpest warning comes from Dario Amodei, CEO of Anthropic, the company behind Claude and Claude Code. His prediction: about 50% of entry-level white-collar jobs could disappear within one to five years—not ten, not twenty—and that shock could push unemployment toward 10% to 20%. He repeated the same call at Davos 2026. When the person building the tools tells you half of the bottom rung is going away, treat it as data.

The barbell almost nobody is pricing

There is a shape to what is coming that most career plans still ignore. Call it the barbell. Heavy weights on both ends. A thin bar in the middle. The top roughly 20% thrives. The bottom roughly 20% actually gets lifted. The middle 60%—the connector—gets crushed in the transition. Economists have talked about job polarization for years. AI is accelerating it on a clock measured in years, not generations.

Farzad has tracked technology disruption since 2012—Tesla, SpaceX, AI, robotics—every day. The pattern that keeps repeating: people who adapt early compound an edge; people who dismiss the shift get left behind. This cycle is larger.

Top end: capital plus agents

At the heavy top end sit capital owners—people with businesses, equity, and assets. AI is a force multiplier for them. Legal work, marketing, customer service, software, accounting, research: work that once meant hiring fifty people can now run with five plus agents. The top 1% now holds about 32% of U.S. wealth, the highest share since the Federal Reserve started tracking in 1989. Elon Musk alone gained about $187 billion in 2025. The top 10% gained about $5 trillion in a single quarter of 2025. That is not mostly harder hours. That is capital compounding when the market prices AI-native businesses at a premium, then those same owners buy more compute and more leverage.

A founder with a small team and good tools can compete with a firm that needed 200 employees five years ago. Gains accrue to whoever deploys the tools—usually whoever can afford to. Early adopters build intuitions that compound. The person who has been orchestrating agents for two years is not in the same labor market as the person who just opened ChatGPT to rewrite an email.

Agents, stripped of marketing language, are digital workers: models that decide and operate a computer—documents, bookings, code, video edits, research. Jobs already exposed to AI skills command about a 56% wage premium, up from about 25% the prior year. The gap is widening, not closing.

Bottom end: cheaper basics and scarce hands

The other heavy end surprises people. AI and robotics also help the global poor, because their binding constraint is often delivery cost, not whether clean water, medical knowledge, or construction skill exist. Human labor does not scale cheaply into every village. When robotic labor trends toward a few dollars an hour, desalination, housing, and healthcare delivery start to pencil at scale. Smartphone diagnostics put formerly scarce expertise in a pocket. AI-assisted construction reaches regions that cannot attract skilled builders.

In rich countries the same end shows up as physical work getting more valuable. Nvidia's Jensen Huang noted at Davos that plumbers, electricians, construction workers, and steel workers are in massive demand—because AI infrastructure needs hands. Construction and trades wages are up over 20% since 2020, outpacing most white-collar tracks. Top earners in some trades clear six figures without a degree. Power-line installers already show medians near $92,000 with tops around $126,000. Ford's CEO has pointed to shortages on the order of 600,000 factory workers and 500,000 construction workers. Estimates still call for large additional electrician, laborer, and supervisor headcount by 2030—before the full AI buildout. Gen Z is noticing: 77% say it matters that a future job is hard to automate, and 42% are already in or chasing skilled trades. When office tasks automate, the shortage of people who can wire a data center or fix a main water line gets worse, not better.

The thin middle: judgment on sale

The pain concentrates in the middle for the next few years: cognitive work that trades time and judgment for money—accountants, lawyers, marketers, analysts, project managers, customer-service reps, paralegals, recruiters, administrators. People who spent years in school building skills they thought would protect them. Eventually physical robots with strong models will pressure warehouse, factory, landscaping, and many hands-on roles too. What makes this wave different from the printing press or electricity is substitution of judgment itself. A single tool can already automate 30% to 40% of a knowledge worker's tasks. That does not delete every job overnight. It means companies need fewer people. Entry-level roles—the old stepping stones—go first.

Workers aged 22 to 25 in AI-exposed occupations have seen about a 13% employment decline since 2022 while other groups held steadier. College-graduate unemployment has hit multi-year highs around 5.8%. Microsoft has estimated millions of white-collar roles under pressure—management analysts, customer-service reps, sales engineers among them. Research into late-2026 org design points to roughly one in five organizations using AI to flatten hierarchies and cut more than half of middle-management seats. The U.S. saw about 1.2 million layoffs announced in 2025, up about 58% from 2024—outside the pandemic, the worst stretch since the Great Recession. IMF managing director Kristalina Georgieva called it a tsunami hitting the labor market and asked where the guardrails were. In advanced economies she has warned that about 60% of jobs could be transformed or eliminated; globally closer to 40%.

Prior transitions gave decades. Agriculture to industry took generations. Manufacturing to services took decades. Amodei and the CEOs deploying these systems are talking two to five years. There is a deeper break almost nobody prices: career-ladder destruction. You cannot become a senior analyst if junior seats disappear. You cannot become a senior lawyer if associate classes shrink to a trickle. The learn-on-the-job model is cracking.

Klarna is the cautionary equilibrium, not the happy ending. After cutting deep and leaning on AI, service quality slipped and some humans came back. The CEO admitted they cut too far. Even after that partial reversal, the company still runs with far fewer people than before. The common path: deploy AI, cut headcount, find quality gaps, hire a fraction back, settle at a leaner steady state. Most firms will accept slightly worse quality for dramatically lower cost.

Move toward an end—soon

Some of this is doom. Some of it is opportunity. Both readings can be half-right. History often creates more jobs than it destroys, and MIT's David Autor has argued AI could restore middle-skill work by extending human expertise rather than erasing it. Lowest-skill workers inside occupations sometimes gain the most from tools. Those arguments assume time—time for new industries, retraining, education reform, and social adjustment. Farzad's read after fourteen years of watching these curves: we do not have that kind of time this round. Capabilities that felt futuristic two years ago are table stakes now. Code that took a senior developer a week lands in hours. Analysis that needed a team runs through one person with the right stack. When this many CEOs stop hedging, they are usually looking at internal data that makes the outcome feel inevitable.

If you are reading this, you are probably not in the top 1% and you are not fighting for basic survival like the global poor. You are likely in that middle 60%. Decide which end of the barbell you are moving toward, and decide soon. Toward the top means ownership and deployment: equity, systems that produce value when you are offline, and real skill at orchestrating agents with tools like Claude Code, Gemini, and open models—not just chat for emails. Toward the durable bottom means skills that still need a body in a room: trades, craft, care, presence. The electrician powering the data center currently has more job security than the analyst inside it. The plumber who keeps billions of dollars of chips from frying has more security than the project manager scheduling the stand-up.

Staying in the thin middle—hoping it blows over, assuming your job is special, believing the model cannot really do what you do—is the riskiest seat. The barbell is forming. The middle is compressing. The CEOs are finally saying the quiet part out loud. Spot it early. Spot it late and the bar has already snapped.

Check the video here.