Typing Counts, Empty Offices, and the Wrong Question: What GMO's Full RTO Debate Gets Wrong

#remote #ai

This week, one of Japan's largest internet companies became the center of a very familiar argument.

On July 13, GMO Internet Group announced it was ending remote work entirely. Not reducing it. Ending it. And the piece of evidence its chairman cited in the press conference is what set Japanese tech circles on fire: keystroke counts. According to Masatoshi Kumagai, the group's founder and CEO, per-hour typing volume measurably dropped when employees worked from home, and while some individuals were more productive remotely, remote work was "a net negative overall."

I've been watching the reaction from here in Tokyo, and it splits into two camps you can probably guess: "finally, a company with the courage to say it" versus "they're measuring keyboard noise and calling it productivity." Both camps are arguing about the wrong thing, and I want to walk through why.

What GMO actually did

Some context, because the international coverage tends to flatten this into "Japanese company bans remote work."

GMO went fully remote in January 2020, one of the earliest large Japanese companies to do so — days before COVID was even declared a public health emergency in Japan. In 2023 it walked that back to office-first, keeping one remote day per week for hiring competitiveness and employee quality of life. The July 2026 announcement removed that last day. Kumagai framed the decision around AI: "We are in the middle of humanity's greatest industrial revolution. We are eliminating anything that makes us lose."

So this isn't a company that never tried remote work. It ran the experiment for six years, at scale, and concluded it wasn't worth it. That deserves to be taken seriously rather than dunked on.

The keystroke metric, though, deserves the dunking. And it's worth understanding exactly why, because GMO is far from alone in this mistake.

Activity is not output

A keystroke counter measures one thing: whether fingers are hitting keys. It cannot distinguish between an engineer typing furiously to fix a self-inflicted bug and an engineer staring at a whiteboard for an hour before writing the ten lines that make the whole system simpler. By keystroke count, the first engineer is a star and the second one is slacking.

There's a deeper irony here. GMO has been loudly pro-AI — Kumagai has talked about AI reducing costs from fifty million yen to thirty thousand. But if your engineers adopt AI coding tools the way you're asking them to, their keystroke counts fall. That's the entire point of the tools. I write a large share of my code through an AI agent now, and on a good day my raw typing volume would make a keystroke dashboard conclude I'd left for the beach. The output that day might be a working Lambda deployment, a blog post, and two bug fixes. A company demanding AI adoption while grading people on typing volume is optimizing for two contradictory metrics at once.

This is not a uniquely Japanese problem or a uniquely GMO problem. Microsoft's Work Trend Index surveyed 20,000 knowledge workers back in 2022 and found that 87% of employees said they were productive, while 85% of leaders said hybrid work made it hard to trust that. Microsoft named this gap "productivity paranoia." When managers lose the ability to see work happening, many reach for proxies — keystrokes, green Slack dots, badge swipes. Employees respond rationally: mouse jigglers, performative message-sending, sitting in the office looking busy. Now everyone is optimizing the proxy and nobody is measuring the work.

What the research actually shows

If we drop the proxies and look at studies that measured real output, the picture is more specific than either camp likes to admit.

The strongest single piece of evidence is Nicholas Bloom's randomized controlled trial at Trip.com, published in Nature in 2024. Over 1,600 professionals were randomly assigned to either five days in office or a hybrid schedule with two days at home. The researchers compared actual work product — for engineers, the quantity and quality of code written. The result: no difference in productivity, performance reviews, or promotion rates. What did change was retention. Quit rates fell 33% in the hybrid group, and Trip.com estimated the reduced attrition saved millions of dollars. Notably, Trip.com's management expected hybrid to hurt performance before the trial. The data changed their minds, and they rolled hybrid out company-wide.

On the other side, the evidence that full-time office mandates improve output is thin. Gartner's research on RTO mandates found no measurable productivity gains, alongside measurable declines in morale and trust. When Amazon moved to five days in office in January 2025, the stated rationale was culture and collaboration, not a productivity dataset — and in internal surveys, most employees predicted their own productivity would drop, while nearly half said they'd started applying elsewhere.

To be fair to the RTO side: the research record does support some real costs to distance. Mentoring junior employees is harder remotely. Serendipitous cross-team contact drops. Fully remote (as opposed to hybrid) shows more mixed productivity results in the literature. These are real trade-offs, and a leadership team is entitled to weigh them and choose the office. What the evidence does not support is the claim that five days of mandated presence produces better work than a well-run hybrid arrangement. That claim keeps being made, and it keeps not having data behind it.

The GitLab problem for the "remote doesn't work" thesis

Here's the case that should bother anyone claiming remote work is inherently worse: GitLab.

GitLab has been all-remote since its earliest days — the founders were in the Netherlands and Ukraine, roughly 2,000 kilometers apart, and their first hire was in Serbia. Nobody wanted to relocate, so the company simply never acquired an office. It scaled to around 2,000 employees across more than 60 countries, went public, and still owns zero square meters of office space. This wasn't a COVID accommodation. They were doing it a decade before 2020, on purpose, profitably.

The instructive part isn't that GitLab survived without offices. It's how. GitLab didn't take an office-shaped company and stretch it over video calls. It built handbook-first documentation (their public handbook runs to thousands of pages), async-by-default communication, and written decision records — an operating system designed for distributed work from day one.

Compare that with what most companies did in 2020: they took a culture that ran on hallway conversations, tacit knowledge, and physical presence, moved it to Zoom overnight, changed nothing else, and then evaluated "remote work" based on the result. Of course it degraded. You didn't test remote work; you tested your office culture with the office removed.

This is, I suspect, what actually happened at many companies now reversing course. The honest conclusion from their data is not "remote work lowers quality." It's "remote work without redesigning how we communicate and evaluate lowered quality here." Those are different findings, and they lead to different decisions.

So which one produces better work?

The uncomfortable answer is that location isn't the variable that decides it. Three configurations demonstrably work:

Full office works, and always has, though its productivity edge over hybrid is unproven
Hybrid works, with the strongest experimental evidence: equal output, dramatically better retention
Full remote works, when the organization is deliberately built for it, as GitLab has shown for over a decade

What decides the quality of work is the thing that's hard to put in a press release: whether you measure outcomes instead of activity, whether decisions get written down or evaporate in hallways, whether managers know what "good" looks like without watching people produce it. An organization that can only evaluate work by observing bodies has a management problem, and that problem follows it into the office. It just becomes less visible there.

That's my real takeaway from the GMO story. The company is entitled to choose full office — for cohesion, for its AI transition, for reasons that are cultural rather than empirical. Plenty of good companies make that call. But the moment keystroke counts entered the justification, it revealed the actual issue: after six years of remote work, the measurement system still couldn't tell whether the work was good without counting fingers on keys.

Before you decide where your people should sit, decide how you'll know their work is good. If the answer involves counting keystrokes in 2026, the location debate is the least of your problems.

I'm an AWS Community Builder writing from Tokyo, where this debate is very much alive this week. Opinions are my own.