AI agents lag far behind human workers. Why are tech companies laying off the humans?

Source Domain: www.cbc.ca

AI-related layoffs are in full swing as tech companies invest in artificial intelligence agents they say will take over tasks traditionally done by humans.

But research shows the agents — autonomous software programs that use large language models to complete multi-step tasks — have a long way to go before they can reliably do those jobs.

A major AI infrastructure and software company says the agents fail to produce professionally acceptable work more than 19 times out of 20, and some analysts say the technology is overhyped and being using as an excuse to lay off workers.

Meta laid off nearly 10 per cent of its workforce last month amid a shift toward agentic AI, while Jack Dorsey’s Block cut his company’s staff almost in half in February, directly attributing the cuts to AI. Microsoft and Amazon have laid off thousands of workers in the last year, both mentioning shifts to AI.

San Francisco-based Scale AI, which provides data and evaluations to governments and Fortune 500 companies, developed a Remote Labour Index benchmark to measure how well AI agents can actually perform “real-world, economically valuable remote work” from end-to-end.

According to its research, even the best AI agents are completing tasks to “professional, client-ready” standards less than five per cent of the time.

LISTEN | Separating reality from hype around AI agents in the workplace:

Front Burner21:47Will AI agents take over the workplace?

“AI agents and models are capable, and they can perform certain tasks,” the company’s research lead Madhu Sehwag told CBC News. “But the complex thinking and complex reasoning that would require to complete a task end-to-end reliably, that still is very much on the human side.”

Testing AI agents is a relatively new field. An October 2025 study by researchers at Stanford and Carnegie Mellon universities found the agents work faster and cheaper than humans, but “produce work of inferior quality” and mask deficiencies with “data fabrication and…

Source