Three years ago, OpenAI predicted that professions unaffected by AI would be crushed at four times the speed.

On February 27th, U.S. fintech company Block announced a 40% layoffs, about 4,000 people, as part of a complete transformation into an AI company. The AI concept dramatically caused its stock price to surge over 20%. This case, from a company not considered a major player in Silicon Valley, reveals the potential economic ripple effects triggered by rapid AI development.

Behind this, there is a number that has been rewritten four times in the past three years.

In March 2023, OpenAI stated: About 19% of U.S. workers will see over 50% of their tasks affected by AI, a process that will take ten years.

By January 2026, Cognizant said: This proportion has already reached 30%, and it’s only three years since ChatGPT was released.

In the same month, Stanford Digital Economy Lab analyzed 285 million job ads and found: In high AI-exposure industries, entry-level job postings decreased by 18%-40%, while demand for senior employees is rising.

If you are still thinking of this revolution as a binary question—“Will AI take human jobs?”—you are already behind. What is truly happening is not job disappearance, but a rupture in the labor market structure: entry points are closing, middle layers are collapsing, and a tiny few “AI drivers” at the top are harvesting everything.

Even more frightening, according to Citrini Research’s projection for 2028, this tearing apart has only just begun.

The March 2023 “Fool’s Errand” and the Winter of 2026

Rewind to March 2023, when ChatGPT first exploded globally. OpenAI researchers, in collaboration with several universities, published a milestone paper, “GPTs are GPTs” (Generative Pre-trained Transformers are General Purpose Technologies).

At that time, OpenAI’s team used a task exposure-based scoring model. They concluded that: About 80% of U.S. workers would have at least 10% of their tasks affected by GPT, and about 19% would see over 50% of their tasks impacted.

Interestingly, they discovered a “high-wage paradox”: unlike past decades where automation (like robotic arms) first eliminated blue-collar jobs, in the GPT era, higher-paid cognitive jobs are more exposed. Programming and writing skills are strongly positively correlated with AI exposure, while science and critical thinking are considered “safe zones.”

At that point, researchers clearly noted a limitation: they did not include multimodal abilities like vision. They hadn’t even considered tool-use capabilities.

In the 2023 framework, AI was still confined to screens, only capable of processing text and code. Their upper limit prediction was that this restructuring might take up to ten years (until 2032) to fully unfold.

By early 2026, global IT services giant Cognizant released an updated report on their 2023 research, titled “New Jobs, New World 2026.”

The report’s opening statement indicated: “What we initially predicted would take ten years (until 2032) is now already happening six years earlier.”

Data shows that today, 93% of jobs in the U.S. are affected by AI to some extent.

Cognizant introduced a metric called “Velocity Score,” which measures how quickly your profession is being eaten by AI.

As shown in the chart, the average annual AI exposure growth rate for all professions was 2%, now it has surged to 9%, accelerating by 4.5 times. This means that jobs previously considered “AI-proof” in 2023 are now being incorporated at four times the previous speed.

Specifically, the proportion of jobs with task exposure over 50% has skyrocketed from 0% in 2023 to 30% (the original forecast for 2032 was only 15%), and jobs with at least 25% exposure increased by 17%, reaching 69%.

Cognizant estimates that, in the U.S. alone, this shifts about $4.5 trillion worth of human labor costs to AI, roughly 15% of U.S. GDP.

Where does this acceleration come from?

The report used a detailed classification, depicting different layers of exposure:

• E0 (No exposure) — 32% of tasks

• E1 (Direct exposure) — 10% of tasks, where GPT alone can save half the time

• E2 (LLM + tools) — 17% of tasks, requiring supporting software

• E3 (With image capability) — 17% of tasks, feasible with visual abilities

• Full automation — 10% of tasks (the biggest jump from 2023 to 2026, from 1% to 10%)

From this classification, we see that the biggest changes from E1 to E3—adding multimodal (eyes and ears), advanced reasoning (brain), and agentic AI (hands and feet)—are where the most impact occurs. Pure ChatGPT has limited influence (10%), but once agents can use specialized tools, impact expands to 27%, and with visual processing, directly affecting 44% of jobs.

For example, a plumber cannot be replaced by AI alone, but if AI can “understand leak locations + infer possible causes + generate repair plans + automatically order parts,” his work is restructured. He still screws in the fittings, but diagnostics and reports are no longer needed from him.

This explosion of combined capabilities leads to several unimaginable consequences in 2023.

First, management is no longer safe. Once upon a time, CEOs and executives believed coordination, budgeting, and decision-making were human-only skills. But in 2026, agents can autonomously schedule, reallocate budgets based on spending patterns, and track project progress. Cognizant’s data shows CEO AI exposure rising from 25% to over 60%.

Second, the defenses of blue-collar and physical world jobs are being breached. Construction workers, mechanics, and plumbers were once considered low-risk from AI. But with multimodal and AR wearables, AI can analyze on-site photos to diagnose leaks or read blueprints. AI exposure in construction rose from 4% to 12%, and transportation from 6% to 25%. A plumber won’t lose his job, but his work will be directed directly by AI headsets.

Ranking jobs by the percentage of tasks AI can perform, Cognizant identified the six most affected professions.

Top of the list is financial managers, with 84% of their work potentially handled by AI. Core tasks like financial planning, budgeting, and risk assessment can all be assisted by AI.

Computer and math-related roles follow, affected at 67%. Business and financial operations, office and administrative support are between 60% and 68%. Legal jobs at 63%, management (including executives) at 60%.

In recent months, changes in software development have been especially notable. Boris Cherny, CTO of Anthropic, revealed in January that nearly 100% of their code is written by their AI products Claude Code and Opus 4.5.

“Personally, I haven’t written a line of code with my own hands for over two months, not even small edits,” Cherny said. “Yesterday, I submitted 22 pull requests, the day before 27, all generated by Claude.”

Of course, they found that 34 professions have no task exposure at all. These are purely manual, on-site, craft jobs: bricklayers, butchers, dishwashers, stonemasons, tire repairers…

These changes may mean that labor market polarization will intensify.

Highly skilled workers become more productive with AI, while low-skilled workers remain stuck in low-wage, manual jobs that can’t be automated. The most dangerous are the middle-tier white-collar jobs that can be automated but haven’t been fully automated yet.

And this is exactly what is happening in the current job market.

Big data doesn’t lie: entry points are closing, middle layers are collapsing

Predictions seem urgent, but what has actually happened in the labor market over the past few years?

Looking at big online job ad data compiled by Lightcast, PwC, Indeed, Stanford, and others from 2023 to 2026, many predictions are confirmed.

The reports predicted that high-wage jobs generally show higher exposure, positively correlated with skills like programming and writing, and negatively with science and critical thinking skills. These were validated in the data.

The overall trend is also correct: the more knowledge-intensive, text-heavy, rule-based jobs, the faster AI penetrates; the more physical, on-site, interpersonal jobs, the lower the exposure.

What has been overtaken is speed. The 2023 report predicted these changes would unfold over ten years, but significant structural shifts appeared in just three. More importantly, the report emphasized that our exposure measurement does not distinguish between labor augmentation and labor replacement—technology feasibility does not equal adoption. But in reality, corporate adoption has been much faster than academic expectations.

Delving deeper, we see a picture called “Hybrid Transformation.” This gentle academic term masks its essence: an ongoing class restructuring.

First, in this transformation, the biggest beneficiaries are AI users. By late 2025 to early 2026, pure “AI-skilled jobs” still account for only about 4.2% of total hiring. But their growth rate is terrifying—mentions of generative AI roles increased more than threefold compared to 2023.

Moreover, from early 2023, hiring has differentiated: while overall hiring decreases, mentions of AI in job ads have been rising steadily.

The market rewards those who master new productivity tools with extremely high pay. PwC and Lightcast data are highly consistent: within the same profession, jobs requiring AI skills command a 15% to 30% salary premium, and in some core fields (like lawyers, financial analysts), the differential can reach 56%.

This is not “common prosperity” for all workers but a sharp wage polarization. Companies are willing to pay high salaries for those who can use AI to boost productivity tenfold, while freezing wages for those doing traditional repetitive mental work.

Second, there is the “implicit death” of entry-level white-collar jobs over these three years. AI hasn’t caused a catastrophic drop in total employment (current hiring remains within post-pandemic normal fluctuations), but a slaughter has occurred in the “rookie” sector.

Stanford Digital Economy Lab, combining ADP salary data and millions of resumes, shows that since ChatGPT’s explosion at the end of 2022, employment among 22-25-year-olds in high AI-exposure industries has shrunk significantly (about 6% decline, even 20% in software development), while older, more experienced workers continue to grow in employment.

A causal analysis based on 285 million US job ads estimates that after ChatGPT’s release, the number of job ads in high AI-replaceable professions decreased by about 12% relative to low-replaceability jobs. The effect is stronger for entry-level roles requiring no advanced education or experience, with declines of 18% and 20%, respectively. Administrative support roles saw declines approaching 40%.

This is called “Senior-Biased Technological Change.” Historically, large companies recruited many fresh graduates and junior staff for basic code review, data cleaning, financial reporting, and legal document drafting. Now, senior employees, aided by a few AI agents, can handle these dirty, repetitive tasks.

A study covering 62 million workers found that from Q1 2023 onward, companies adopting GenAI significantly reduced entry-level employment. Companies are not laying off workers but simply not hiring anymore.

Because mid-level employees, with AI, can do more work. Companies are even reluctant to fire junior staff, preferring to let older workers naturally attrit. This “boiling frog” style layoffs evade labor laws.

The “first step” of career entry-level positions is being taken away by AI.

Finally, task rewriting (Task Rewriting) is replacing job disappearance. A famous 2013 Oxford prediction estimated that “47% of jobs will be automated.” Why hasn’t this happened yet? Because jobs are shells containing countless “tasks.”

Indeed and Revelio Labs data show that job titles haven’t disappeared, but the “job responsibilities” in ads are being rewritten. In finance, clerical, and entry-level coding roles, tasks like “daily reconciliation” and “generating standard code” are declining sharply; replaced by demands for “complex management,” “AI system guidance,” “edge case handling,” and “quality verification.”

This confirms Cognizant’s insight: even if 39% of a job’s tasks are taken over by AI, the remaining 61% still require humans to integrate AI outputs into larger business contexts. The next one or two years will see a “human + AI” reconstruction, where pure executors are eliminated, leaving only judges and coordinators.

But even judges and coordinators won’t need many.

A senior judge + AI can do what ten entry-level workers used to do; companies only need about one-fifth of the original workforce. The essence of human-AI collaboration is to use a small elite + AI to replace most ordinary workers.

Toward 2028, the Agent Singularity and Global Intelligence Crisis

What if we extend the current “structural squeeze” in the job market and the evolution of Agent technology forward?

Before answering, let’s review what has happened in the past three years. In 2023, OpenAI said it would take ten years to change the job structure; by 2026, Cognizant said a huge change had already occurred; in 2023, fully automated tasks accounted for 1%, and in 2026, this was 10%; in 2023, entry-level jobs were still being recruited normally, but by 2026, hiring in high AI-exposure industries for entry-level positions had already decreased by 18%-40%.

If this acceleration continues, what will 2028 look like?

Citrini Research, in a deep projection titled “The 2028 Global Intelligence Crisis: A Future Financial Thought Experiment,” depicts a chilling post-singularity world.

In this scenario, set in June 2028.

Between 2026 and 2027, markets are immersed in a bizarre frenzy. Due to large-scale deployment of AI Agents, the S&P 500 and Nasdaq soared, corporate profits hit new highs, and labor productivity reached levels unseen since the 1950s. Agents creating products don’t sleep, need health insurance, or get sick.

But economists quickly discovered a fatal problem: “Ghost GDP.” It refers to wealth that shines in national accounts but never circulates in the real economy.

Why? Because a GPU cluster in North Dakota completed the work of ten thousand white-collar workers in Manhattan, yet machines don’t buy coffee, pay rent, watch movies, or go on vacation. The consumer-driven market, which accounts for 70% of the U.S. economy, begins to wither.

If we project the current recruitment “structural squeeze” and the evolution of Agent technology forward, this term might turn from metaphor into reality.

Past innovations (cloud computing, internet) mostly involved capital expenditure (CapEx), creating vast upstream and downstream employment. But the introduction of Agents is a direct replacement of operational expenditure (OpEx).

In 2026, as agentic tools (like an advanced version of Claude Code) reach new capabilities, CIOs find they can replace hundreds of thousands of dollars in SaaS services with internal AI prototypes in weeks. Software companies (like ServiceNow) cut 15% of their staff to reinvest in more powerful AI tools to stay competitive.

This creates a feedback loop with no physical brakes: AI gets stronger → companies lay off workers → savings from layoffs buy more AI power → AI gets stronger → further layoffs.

The optimized white-collar workers lose income, leading to reduced consumption, which decreases company revenues. To maintain profit margins, companies adopt AI more aggressively and cut more jobs. Wealth concentrates at the few who control computing power at an unprecedented speed.

By 2027, the crisis spreads from the software industry to the entire “intermediary layer.” Over the past fifty years, society built a vast “friction-based monetization” empire—travel platforms, insurance renewals, real estate agents—because humans lacked time, patience, and information. We tolerated commissions and fees.

But in 2028, consumers are fully integrated with personal AI Agents. These Agents perform 24/7 online price comparisons, automatically cancel forgotten subscriptions, and instantly handle real estate transactions and legal reviews. Traditional subscription and intermediary economies collapse overnight. The so-called “business stickiness” is proven to be just a layer of emotional “friction” before the relentless optimization of machines.

The next 24 months

For centuries, economists have comforted the public with the mantra: “Technology destroys old jobs but always creates new ones.” ATM machines replaced some tellers but led to more bank branches; the internet eliminated the Yellow Pages but created e-commerce and food delivery.

But this time is different. Because the new jobs of the past all had to be done by humans. When AI evolves into “General Intelligence,” it not only can do old jobs but learns and executes new ones faster and cheaper than humans. AI does create new roles (like prompt engineers, AI safety auditors), but each new role makes dozens of traditional high-paying white-collar jobs redundant. Moreover, these new roles have very short lifespans, quickly replaced by the next generation of stronger, cheaper Agents.

All signs point to the same conclusion. AI won’t physically eliminate humans like Terminator, but it is reconstructing the labor value network of society in an extremely efficient, rational way.

But this is only the first step.

By 2028, the real problem is: if a society’s value creation is 99% by machines, but machines do not consume, buy houses, see doctors, or pay taxes, how does this society’s cycle keep going?

We might dismiss Citrini’s 2028 scenario as alarmist, but data from the past three years already proves that technological acceleration far exceeds human societal adaptation. In 2023, OpenAI said it would take ten years; in 2026, Cognizant said it had already happened. So, in 2028, will we see that GDP skyrockets while consumption withers?

Perhaps the answer isn’t in technology itself but in an older question: When the main driver of productivity is no longer human, how does wealth get distributed?

Neither Adam Smith nor Marx answered this, because in their eras, labor was always human.

The 4,000 layoffs at Block, the 20% stock surge on Wall Street—these already tell us which path capital has chosen.

The question is, what will we choose?

In 2026, we must answer this question.

Because our time may only have 24 months left.