Written by: Xu Chao

Source: Wall Street Journal

AI programming tools promised to free engineers, but the reality has spawned a new wave of efficiency anxiety.

As AI programming agents like Claude Code from Anthropic and Codex from OpenAI continue to advance, tech companies are caught in a top-down "productivity obsession." Executives are personally getting involved in coding, employees are being asked to increase their interactions with AI, while overtime hours are not decreasing but increasing. AI was supposed to be a labor-saving tool, but it has become a new source of stress in many workplaces.

Survey data reveals a significant cognitive disconnect: a survey by consulting firm Section shows that over 40% of C-level executives believe AI tools save them at least 8 hours a week, while 67% of non-management employees said AI helps them save less than two hours, or is even unhelpful. A ongoing study from the University of California, Berkeley on a 200-person organization found that even when employees have delegated a large amount of work to AI, actual working hours are still increasing.

The spread of this anxiety has structural reasons. When chief technology officers code with AI at 5 AM and CEOs measure team effort by billing amounts, the entire industry's imagination of "efficiency" has been redefined— and the cost of this redefinition is borne by ordinary employees.

Executives enter the coding arena, efficiency anxiety spreads from the top down

The term "vibe coding" originally carried a sense of lazy expectation. Former OpenAI researcher Andrej Karpathy brought this concept to the public eye in February 2025, describing a new programming model where engineers only need to chat with AI to complete development— "fully immersed in the atmosphere."

However, a year later, the atmosphere has already shifted.

Intuit's chief technology officer Alex Balazs described his recent routine: his wife came downstairs at 8 AM to find he had been working for hours. "She asked how long I had been up, and I said I started coding at 5 AM." To be precise, he was guiding the AI agent to write code for him, saying that it allowed him to reconnect with foundational code he hadn't touched for years.

Such executive behavior is passing down pressure. OpenAI President Greg Brockman recently posted on X, stating, "Every moment your agent is not running feels like a wasted opportunity." This statement precisely triggers the already prevalent workaholic culture in the tech industry.

Alex Salazar, co-founder and CEO of AI startup Arcade.dev, is even more direct. He regularly reviews the company's Claude Code bills—where the billing amount is directly tied to how frequently engineers use the tool—and publicly criticizes employees who "aren't spending enough": "I would say, 'You aren't working hard enough.'" He noted that after the first such "faith meeting," the company's AI programming tool bills skyrocketed tenfold, which he views as a sign of progress.

Employees face quantitative management, "AI fatigue" spreads quietly

In this atmosphere, the evaluation methods for employees have been quietly changing.

DocuSketch, a software company focused on property restoration, has its VP of Products Andrew Wirick stating that the company now tracks how many times engineers interact with AI programming tools daily, assuming that a higher number means stronger team productivity. Claude Code also generates weekly reports for each engineer, listing all patterns of ineffective interactions with AI and providing suggestions for improvement.

Wirick himself admitted to experiencing a sense of "addiction." "It feels like I must interact a few more times every day, thinking about how to do a few more before bed." He attributes this state to the "epiphany experience" when he tried Anthropic’s latest model Opus 4.5 last November—when he delegated typically engineer-assigned feature prototype tasks to the model, which independently broke down and implemented the task 20 minutes later, "it felt like my brain was rebooted."

This mindset of acceleration is eroding the boundaries between work and life. The Berkeley study found that even when large numbers of tasks have been taken over by AI, people’s working hours have not shortened. Some engineers have also begun to publicly admit they are experiencing "AI fatigue"—constantly worrying about missing the next breakthrough, which seems to always just require one more prompt.

The cognitive gap between executives and employees is widening

The enthusiasm of executives largely comes from the freshness of creating something themselves. Salazar admits that personally building prototypes with AI feels more "visibly productive" than handling authorization and decision-making in the usual manner. He recently even responded directly to a service request from a major financial client, building a demo application from scratch.

At Intuit, product managers and designers are now also encouraged to use "vibe coding" to build feature prototypes in QuickBooks, with Balazs stating, "At least for now, product managers can hold something concrete and say to engineers, 'I want something like this.'"

However, survey data from Section Consulting shows that this cognitive disconnect is quite pronounced.

There is a huge gap between executives’ experiences of the AI dividend and the experiences of front-line employees. Salazar believes this partly stems from the higher transformational costs employees bear as they adapt to new tools: "They are implicitly required to find time to explore and experiment, but the expectations of daily work have not been adjusted accordingly to free up this space."

Concerns about job security are also very real. Salazar confessed that he originally planned to change third-party web service providers, but now the marketing team is able to update the company website using AI tools, eliminating the need for that outsourcing expense.

"Task expansion" and false prosperity: the other side of the efficiency myth

Berkeley researchers have named this phenomenon "task expansion": When non-technical colleagues begin using AI to generate code, engineers must spend time cleaning up these half-finished products, thus increasing their workload. Intuit's Balazs admits this is reshaping once-clear job boundaries, leading to a more "blended" role for many positions, which complicates existing collaboration relationships.

A deeper issue lies in whether this construction boom is genuinely creating valuable things or just generating more items.

Analysts point out that if this AI-driven productivity obsession is not constrained, it could lead to the emergence of a large amount of "busyware" (wasted software)—small changes to websites that no one cares about, customized dashboards for a single user, and prototype projects abandoned midway by marketing heads, all eventually dumped on engineers to implement. Each of these, at the moment, seems to have its rationale, but most will ultimately end up in the garbage bin of obsolete code.

Intuit's Balazs states that by measuring code production and delivery speed, the productivity of the company's engineers has increased by approximately 30%. However, in this increasingly "disposable" future of code, the true efficiency dividends may lie in answering another question: what things should not be built at all.