The stronger the AI, the more people feel exhausted, and "anxiety" becomes the norm for companies and employees.

Written by: Xu Chao

Source: Wall Street Insights

AI programming tools promise to free engineers, but in reality, they have sparked a new wave of efficiency anxiety.

As AI coding agents like Anthropic’s Claude Code and OpenAI’s Codex continue to improve, tech companies are caught in a top-down “productivity obsession.” Executives are personally coding, employees are being asked to increase their interactions with AI, and overtime hours are not decreasing but increasing. AI, which should be a labor-saving tool, has instead become a new source of workplace pressure.

Survey data reveal a clear perception gap: a study by consulting firm Section shows that over 40% of C-level executives believe AI tools save them at least 8 hours per week, while 67% of non-management employees say AI helps them save less than two hours or not at all. A continuous study at the University of California, Berkeley, involving a 200-person organization, found that even after delegating much work to AI, actual working hours are still lengthening.

This anxiety is structurally rooted. When CTOs code at 5 a.m., and CEOs measure team effort by billable amounts, the industry’s concept of “efficiency” has been redefined—and the cost of this redefinition is borne by ordinary employees.

Executives coding themselves, spreading efficiency anxiety from top to bottom

The term “Vibe coding” initially carried a sense of relaxed anticipation. In February 2025, former OpenAI researcher Andrej Karpathy brought this concept into the public eye, describing a new programming mode where engineers only need to chat with AI to complete development—“completely immersed in the vibe.”

But a year later, the vibe had already shifted.

Alex Balazs, CTO of Intuit, describes his recent routine: his wife comes downstairs at 8 a.m., and he’s already been working for hours. “She asked how long I’d been up, and I said I got up at 5 a.m. to write code.” More precisely, he’s guiding AI to write code for him, which has allowed him to reconnect with low-level code he hadn’t touched in years.

This behavior among executives is now cascading downward, increasing pressure. OpenAI President Greg Brockman recently posted on X, saying, “Every moment your agent isn’t running feels like a waste of opportunity.” This statement sharply triggers the already prevalent workaholic culture in tech.

Alex Salazar, co-founder and CEO of AI startup Arcade.dev, is more direct. He regularly checks the company’s Claude Code bills—linked directly to how often engineers use the tools—and criticizes employees who “don’t spend enough.” “I tell them, ‘You need to hustle more,’” he said. After the first such “faith meeting,” the company’s AI coding bills skyrocketed tenfold, which he views as a sign of progress.

Quantified management of employees, “AI fatigue” quietly spreading

In this environment, how employees are evaluated is also subtly changing.

DocuSketch, a software company focused on property restoration, reports that its VP of Product, Andrew Wirick, now tracks how many times engineers interact with AI coding tools daily, assuming higher numbers mean greater productivity. Claude Code also generates weekly reports for each engineer, listing all patterns of ineffective AI interactions and offering suggestions for improvement.

Wirick admits he’s developed a kind of “addiction”: “I feel like I need to do more interactions every day, even thinking about how to do more before bed.” He attributes this to an “epiphany” he had in November last year when trying out Anthropic’s latest model, Opus 4.5. He handed a typical feature prototype task to the model, and after 20 minutes, saw it decompose and implement the task autonomously—“it felt like my brain was rebooted.”

This all-accelerating mindset is eroding the boundaries between work and life. Berkeley’s research shows that even when AI takes over many tasks, people’s working hours have not shortened. Some engineers are openly admitting to experiencing “AI fatigue”—constantly worried about missing the next breakthrough, which always seems just one prompt away.

The widening perception gap between executives and employees

Executives’ enthusiasm largely stems from the novelty of creating with AI firsthand. Salazar admits that building prototypes himself with AI feels more “productive” than handling approvals and decisions. Recently, he even responded directly to a major financial client’s request by building a demo app from scratch.

At Intuit, product managers and designers are now encouraged to use “vibe coding” to build prototypes within QuickBooks. Balazs says, “At least now, product managers can bring a concrete example to engineers and say, ‘I want something like this.’”

However, Section’s survey shows a significant perception gap.

There’s a huge disconnect between how executives perceive AI’s benefits and how frontline employees experience them. Salazar believes this partly results from employees bearing higher transformation costs when adapting to new tools: “They’re implicitly asked to find time to explore and experiment, but their daily work expectations haven’t been adjusted to free up that time.”

Job security concerns are also real. Salazar admits he planned to switch to a third-party web service provider, but now the marketing team can update the company’s website using AI tools, so the outsourcing expense was cut.

“Task expansion” and false prosperity—the other side of the efficiency myth

Berkeley researchers call this phenomenon “task expansion”: when non-technical colleagues start generating code with AI, engineers have to spend time cleaning up these semi-finished products, increasing their workload. Balazs admits this is reshaping the once-clear division of roles, leading to more “hybrid” roles and complicating collaboration.

The deeper issue is whether this wave of construction is creating valuable things or merely producing more stuff.

Analysts warn that if this AI-driven productivity obsession isn’t restrained, it could lead to a proliferation of “busyware”—sites with trivial updates, custom dashboards for a single user, half-finished prototypes abandoned by marketing—ultimately all handed over to engineers. While each seems justified in the moment, most will end up as discarded code.

Balazs states that, measured by code production and delivery speed, the productivity of company engineers has increased by about 30%. But in this increasingly “one-time” coding future, the real efficiency gain may lie in answering a different question: what should never have been built in the first place.