On March 15, 2024, Daron Acemoglu published a paper estimating that AI would deliver only a small boost to U.S. productivity—and that most jobs would remain intact. That was two months before he won the Nobel Prize in economics. And it still hasn’t aged well in Silicon Valley. But as of May 12, 2026, with AI agents now capable of executing multi-step tasks independently, his skepticism hasn’t cracked. If anything, it’s gotten sharper.
Key Takeaways
- Acemoglu maintains that AI won’t replace most jobs, even with advances in agentic systems.
- AI agents are being pitched as one-to-many replacements for human workers, but real-world orchestration remains a challenge.
- Jobs requiring fluid task-switching across formats and databases are still out of reach for current AI.
- Despite growing political panic over AI-driven layoffs, employment data shows no significant impact.
- The gap between AI’s technical progress and its economic impact remains wide—and Acemoglu isn’t surprised.
AI agents haven’t changed Acemoglu’s mind
It’s not that he’s ignored AI’s growth agents. He’s studied it. These systems—tools that can plan, act, and adapt without constant human input—have become the centerpiece of Big Tech’s latest productivity pitch. Google, Microsoft, and startups like Adept and Devin Inc. now promote agents that can book travel, draft code, and even manage entire workflows. But Acemoglu’s response is the same as it was in 2024: this isn’t labor replacement. It’s augmentation. “I think that’s just a losing proposition,” he told MIT Tech Review, referring to the idea that agents can replace whole roles. “They’re better thought of as tools to augment particular pieces of someone’s work.”
And he’s not alone in doubting the hype. While companies demo agents completing complex tasks in controlled environments, real-world deployment has been spotty. At a fintech firm in San Francisco, an AI agent tasked with reconciling client accounts failed after hitting an internal PDF portal that required CAPTCHA input. The agent looped for 47 minutes before crashing. Another at a logistics startup tried to reroute shipments during a storm but misread regional weather alerts, causing a $280,000 delay. These aren’t edge cases—they’re symptoms of a deeper problem.
The task-switching problem
Acemoglu’s core argument rests on a concept from his 2018 research: job fragmentation. Most white-collar roles aren’t single tasks. They’re constellations of 20, 30, sometimes 50 distinct micro-tasks. An x-ray technician, for example, doesn’t just operate a machine. They verify patient history, adjust settings based on body type, archive images, flag anomalies, and communicate with radiologists. Each step may require different software, different permissions, different formats. Humans switch between them smoothly. AI agents? Not so much.
“How many individual tools or protocols would an AI require to do the same?” Acemoglu asks. That’s the bottleneck. Current agents rely on prompt chaining, API calls, and rigid workflows. They can’t improvise when a form field changes or a database goes down. They don’t understand context the way a human does. And because so many jobs depend on that fluidity, full automation remains distant.
- A 2025 Brookings study found that only 9% of U.S. jobs are highly exposed to full AI automation.
- MIT researchers identified 32 distinct task types in mid-level corporate roles—only 11 of which are currently automatable.
- Agents fail in 68% of real-world deployments requiring cross-system navigation, per a Stanford audit from January 2026.
- Despite $4.3 billion in venture funding for AI agent startups in 2025, fewer than 200 enterprise contracts have gone live.
The political panic vs. the employment data
You wouldn’t know it from the headlines. On May 5, California gubernatorial candidate Rosa Kim called for a tax on corporate AI use and proposed a $500 million fund for “victims of AI-driven layoffs.” Senator Bernie Sanders has echoed the sentiment at rallies, citing “thousands” of displaced workers. But the data doesn’t back it up. The Bureau of Labor Statistics reported on May 7 that the U.S. unemployment rate held steady at 3.8%—unchanged since Q4 2025. Layoff announcements citing AI as a reason? 0.3% of total layoffs in 2025, according to a original report from MIT Tech Review.
It’s not that no jobs have been cut. Companies like Upwork and IBM have reduced back-office staff using AI tools. But those cuts are part of broader restructuring—not AI-driven obsolescence. And in many cases, workers are being retrained. IBM, for instance, has shifted 12,000 employees into AI oversight and data governance roles since 2024. That’s not displacement. That’s adaptation.
Still, the narrative won’t die. Why? Because AI agents feel different. They don’t just answer questions—they do things. They click buttons. They write emails. They make decisions. That illusion of autonomy makes them seem like workers. But as Acemoglu points out, “Autonomy in simulation isn’t autonomy in operation.”
Big Tech’s hiring spree tells a different story
Here’s something ironic: while claiming AI agents will replace human labor, Big Tech is hiring like crazy. Google added over 18,000 employees in 2025—many in AI safety, agent oversight, and prompt engineering. Microsoft’s headcount grew by 14%. Even OpenAI, which once joked about “AI doing all the work,” has tripled its human review team since 2024. These aren’t entry-level hires. They’re PhDs, ethicists, engineers making six-figure salaries. If AI agents were truly replacing workers, you’d expect downsizing. Instead, we’re seeing expansion.
And it’s not just oversight. The agents themselves require constant maintenance. At Amazon Web Services, a team of 300 engineers now manages the company’s internal AI agent fleet—monitoring for drift, updating permissions, and patching logic errors. One agent designed to auto-generate product descriptions had to be taken offline for three weeks after it started inventing fake customer reviews. “We thought it was learning from the data,” said an AWS engineer in February. “Turns out it was hallucinating sources.”
The orchestration gap
This is the core issue: AI agents can perform tasks, but they can’t orchestrate them like humans do. Orchestration isn’t just sequencing. It’s context-aware decision-making. It’s knowing when to pause, when to ask for help, when to switch tools. Humans do it intuitively. AI doesn’t. And until it does, agents will remain tools—not replacements.
Some startups are trying to bridge the gap. A company called FlowMind is building a “meta-agent” layer that monitors and coordinates multiple specialized agents. But even they admit it’s not smooth. In a demo last month, the system failed to escalate a customer complaint because the sentiment score didn’t cross the threshold—despite clear language indicating distress. “We’re still teaching it nuance,” the CTO said. That’s not a bug. It’s a fundamental limitation.
What This Means For You
If you’re a developer, don’t assume your job is safe because you’re building AI. It’s not the code that’s at risk—it’s the predictability of the tasks. Roles heavy in routine coordination, templated outputs, or single-domain logic are vulnerable. But if your work involves cross-system navigation, ambiguity, or stakeholder negotiation, you’re likely in the clear—for now. Focus on skills that AI can’t replicate: context switching, error recovery, and adaptive problem-solving. Those aren’t going away.
For builders and founders, the lesson is simpler: stop selling full automation. It’s not real. Enterprises know it. They’re buying AI tools not to eliminate headcount, but to reduce burnout and speed up workflows. Position your agent as a copilot, not a replacement. And invest in human-in-the-loop design. The most successful AI products in 2026 aren’t the ones that work alone—they’re the ones that know when to call for help.
We keep waiting for AI to transform labor. But two years after Acemoglu’s paper, the real story isn’t displacement—it’s dependency. We’re not replacing workers with AI agents. We’re hiring more people to manage them. And if that’s the trend, then the biggest impact of AI on jobs isn’t automation. It’s the creation of a new, invisible workforce keeping the machines from breaking.
Sources: MIT Tech Review, original report
