DeNA Cut Workload 90% With AI. Nobody Left Early.

The 90% Reduction That Changed Nothing

Tomoko Namba, chairman of DeNA, said it plainly at a March event: when AI cuts workload, employees don’t go home early. They fill the gap themselves. Some tasks at the internet company shrank by 60 to 90 percent. Nobody noticed, because the hours stayed the same.

It is a quiet paradox that deserves more attention than it gets. Companies worldwide are spending billions on AI tools under the assumption that automation will buy back human time. DeNA’s experience suggests that assumption may be wrong—not because the technology failed, but because the workplace structure around it did not.

The numbers from DeNA’s internal rollout were striking. A single employee using AI-assisted code generation reported cutting a twelve-hour debugging session down to two. Another transformed a four-hour report into forty minutes of editing an AI-drafted outline. These were not marginal improvements. They were wholesale collapses of effort. And yet, at the end of each workday, people clocked out at roughly the same hour they always had.

Namba’s observation was notable precisely because it came from the top. CEOs and executives often praise AI for efficiency gains without acknowledging what happens to those gains once they materialize. By putting the paradox on record, Namba named something most workers already lived but rarely discussed openly.

The Anonymous Diary That Broke the Internet

In August, an article appeared on Hatena Anonymous Diary titled, in essence, “I used AI to streamline my work, and somehow my workload doubled.” The author described a straightforward experiment: taking tasks that once consumed eight hours and completing them in four using AI. The expectation was simple—go home at 4 p.m.

The reality was different. Management did not shorten the workday. Instead, the four remaining hours filled with new assignments. The employee who had shared AI techniques with colleagues became the de facto trainer, fielding questions and explanations. Pay did not change. The writer called themselves a “confusing Santa Claus”—someone who gave away efficiency and received nothing in return.

The post split its audience. Some readers called it a realistic portrait of Japanese corporate life. Others assumed it was satire. But the underlying pattern is documented, not imagined.

What made the post resonate was not just its specificity but its ordinariness. Commenters shared nearly identical stories from their own workplaces—a customer service representative who replaced three hours of manual note-taking with AI-generated summaries and then was reassigned to handle escalated complaints; a marketing coordinator whose pitch decks went from all-day projects to afternoon tasks, only to find three new campaigns added to her queue the following week; a data analyst who automated a weekly reporting process and was promptly given responsibility for two additional business units.

The recurring structure was unmistakable: efficiency created a vacuum, and organizational gravity filled it immediately. No one had to order the extra work. It arrived as a matter of course, as though capacity and expectation were permanently locked in place.

The Numbers Don’t Lie

A survey by the Personnel Comprehensive Research Institute found that generative AI reduced average task time by 16.7 percent—roughly 26 minutes per week. Only 25.4 percent of respondents actually left work earlier because of it. The rest used the saved time for more work. Sixty-one point two percent of freed time went back into tasks. Among those people, 75.4 percent said the new work was just more of the same daily business.

Motoki Tamura, a researcher at the institute, put it bluntly: the 40-hour workweek has not changed. Speed has. When a task drops from 100 units of effort to 50, management does not reduce the total to 50. It adds 50 more units—perhaps training others on AI, or handling the overflow that the speed created. The total reaches 100 again, or climbs to 150. The excess becomes overtime.

Tamura’s framing exposed the mechanism clearly. This is not a failure of workers to take advantage of efficiency gains. It is a structural feature of how modern organizations allocate time. When output is measured by volume—number of reports written, tickets resolved, lines of code reviewed—the natural reaction to a tool that expands capacity is to expand expectations, not to compress hours.

The survey also revealed a second-order effect worth tracking: early adopters of AI within companies bore disproportionate invisible labor. They became informal help desks, document-writers of best practices, and unintentional standard-bearers whose productivity gains set implicit benchmarks for their peers. Those benchmarks, in turn, raised the floor for everyone else—without anyone formally announcing that the floor had risen.

The Productivity Illusion

Labor productivity is calculated as value added divided by hours worked. AI shrinks the denominator. But it does not automatically grow the numerator. As Tamura noted, no company can simply raise prices because it now uses AI. The value delivered to customers remains the same. When every competitor adopts the same tool, quality rises across the board and no one gains a competitive edge.

In practice, the denominator may not shrink at all. If saved time is immediately reinvested in additional work within the same 40-hour frame, the hours worked are unchanged. The speed appears real to the individual, but the aggregate numbers tell a different story.

There is also a subtle compounding problem. As individual workers accelerate, teams that depend on sequential handoffs can experience bottlenecks elsewhere. A designer who finishes mockups in two hours using AI still waits for approval cycles that move at human speed. The gain stays trapped in pockets rather than flowing through the organization. Meanwhile, the worker who finished early is expected to absorb the next task before the previous approval lands.

This creates what might be called the acceleration trap: the faster you work, the more work arrives, and the less room you have to recover. Over time, this erodes the margin that makes sustainable performance possible. Workers do not slow down to protect their bandwidth—they accelerate further, hoping to stay ahead of the queue. The result is not more free time. It is more anxiety.

Why This Matters Beyond Japan

Japan’s work culture is often described as uniquely rigid. But the DeNA case reveals a structural problem, not a cultural one. Any organization—in Silicon Valley, Berlin, or Singapore—that rewards output measured in hours rather than outcomes will absorb AI efficiency gains the same way. The mechanism is universal: speed creates capacity, capacity creates expectations, expectations erase the time gain.

American companies have been especially eager to promise AI-driven workweek reductions. Some have floated four-day weeks powered by automation. DeNA’s data suggests those promises will hit the same wall unless leadership is willing to redefine what “work complete” means. The gap between rhetoric and reality is where employee trust erodes fastest. Promising shorter weeks and delivering busier ones is a recipe for quiet cynicism.

Second-order consequences extend beyond individual burnout. When efficiency gains are routinely captured by increased workload rather than reduced hours, investment in AI tools becomes harder to justify on human terms. Employees learn early that faster work means more work, and adoption slows—not because the technology is ineffective, but because the incentive structure is perverse. The most rational response to an efficiency tool that increases your load is to use it sparingly, to hoard the time savings rather than display them. This breeds a kind of strategic inefficiency that benefits no one.

What Would Actually Change Things

The fix is not harder to implement than it is to ignore. It requires two shifts. First, leadership must treat freed time as time off, not as capacity for more work. That means formally reducing workloads when AI cuts task time, not informally filling the gap. This might look like explicit policies—when a process drops from eight hours to two, the schedule adjusts accordingly. It might look like protected blocks of time that managers are instructed not to fill. Either way, the signal matters: time saved is time yours, not time the company reclaims.

Second, companies need to measure productivity by output quality and value created, not by hours logged or tasks completed. AI should be evaluated on whether it enables better decisions, faster innovation, or higher customer satisfaction—not on whether it lets managers pile on more assignments. This requires a willingness to tolerate ambiguity in performance metrics and to resist the comfort of counting things that are easy to count.

A third, less visible change is cultural. Organizations need to normalize the idea that working fewer hours is not laziness but a legitimate outcome of technological progress. This normalization has to come from the top. When executives leave at five on the dot after adopting AI, it sends a different message than when they stay until seven to “model commitment.” The former makes policy real; the latter makes it decorative.

DeNA’s employees already proved that the technology works. The question now is whether their companies will prove that they value the people using it.

The question for every company adopting AI is not whether it can cut task time. It is whether it has the courage to give that time back.