In 2008, Masayoshi Son talked Steve Jobs into bringing the iPhone to Japan as a SoftBank exclusive. The Japanese IT industry’s smartphone transition started there. On March 6, 2026, at DeNA × AI Day 2026, when Chairwoman Namba Tomoko presented the results of one year of going all in on AI, that scene came to mind. Looking at the scale of the numbers and how deeply the organization had been taken apart, this wasn’t a story about bringing in a tool.
That’s the presentation. What about the inside? I asked Developer A, who works at DeNA, directly. The numbers on stage and the temperature on the floor were different.
#Half the headcount
The line about cutting headcount in half first landed on the floor at a company-wide presentation in June 2024. Dropped without context, it flooded the internal chat with people asking whether that meant layoffs. The explanation came afterward. The current work would be done by half the people, and the rest would go find new AI businesses, over a span of about two years. In practice, it meant each person had to use AI to do the work of two or more.
Then in February 2025, the company declared it was going all in on AI. The plan was to keep the existing businesses, which about 3,000 people had been running, going with half that headcount, and to reorganize the other half into 10-person squads that would churn out new ventures.
On the new-venture side, the strategy is to compete on domain depth at the application layer rather than build foundation models in-house, and DeNA set up a subsidiary dedicated to agent businesses, DeNA AI Link, to execute it. Namba warned that mediocre expertise would be wiped out in one stroke. With OpenAI and Google raising general-purpose capability, she said, the only combination that survives is proprietary data plus domain knowledge.
#The numbers on stage
A year later, DeNA announced that on some projects AI now handled 95% of the work and people 5%, and presented that as a case of 20x productivity. Since Claude Opus 4.5 came out in November 2025, hands-on coding has dropped sharply. Hours on legal review fell 90%, QA hours fell 50%, and the cost of screening Pococha live streams fell 60%.
Developer A put the share AI handles on his team somewhere between 40% and 50%. The 95% in the keynote is a case from a specific project, and by his account the teams around him look much like his. I couldn’t confirm whether the keynote’s 95% and his 50% were measured with the same ruler.
The point is that DeNA didn’t layer AI onto existing work. It redesigned the workflow with AI as the starting assumption. DeNA published more than 100 use cases that came up from the floor under the name 100 Drills. The term comes from baseball practice, where a fielder takes 100 ground balls in a row, and it means drilling the real thing until it’s in your body.
The keynote sorted AI use into three stages. The first is prompt engineering, the stage of asking AI good questions. One-off requests like code review and contract summaries belong here. A year ago this alone was enough, the keynote said, and most companies are still here.
The second is context engineering, where you hand over project documents, the codebase, and internal rules together to raise the accuracy of the answers. The third is environment engineering. Once agents fetch information on their own, the core job becomes a human deciding the scope and permissions, how far they can read and which actions they’re allowed. DeNA said it had gotten this far.
The IT department registered an agent, Lemon-kun, as an employee. It can read the wiki and the calendar, but its permissions are designed so that important actions need a person’s approval, and Namba said she talks with this agent on Slack every day as part of her work. The approach onboards AI as a member of the organization, not a tool.
#Different teams, different speeds
As Developer A sees it, even on the same floor, utilization splits by what a team does. Teams that handle formulaic shared work, like security issues and infrastructure migrations, have high utilization, and teams designing new features have low utilization. Those teams are spending their time first building good context to hand to the AI. They’re living the second stage from the keynote, context engineering, firsthand.
Speed splits by team too. Among the teams Developer A has seen, some finished a three-month plan in two weeks, and others are pressed to fit into three months. The development cycle that used to be uniform is being reshaped at a different speed for each team.
#A ruler for skill level
If the keynote’s three stages describe the technology, people are measured separately. Since August 2025, DeNA has measured each employee’s level of AI use on a five-step scale called DARS (DeNA AI Readiness Score). It splits non-developers from developers. Non-developers are asked whether they can refine AI output and put it to use, and developers whether they can design and operate agents. The same level demands different skills.
The developer criteria go like this.
- Level 1: Can generate code and gather information through conversation with AI.
- Level 2: Can integrate AI into an editor and refine prompts.
- Level 3: Can replace parts of a workflow with agents and design prompt chains.
- Level 4: Can operate standalone agents as team members and design AI apps and agents.
- Level 5: Can build AI platforms and orchestration agents and lead AI development across the organization.
Organizational levels are measured separately from individual ones. They have quantitative criteria of their own, which makes them trackable per organization, and the goal in the official announcement is for every organization to reach organizational level 2 within fiscal 2025, except for a few with unavoidable circumstances such as collaboration constraints.
As far as Developer A knows, most employees sit at level 2 and nobody has reached level 4 or 5 yet. What the company wants is level 4. Running an agent as a team member, the way the IT department runs Lemon-kun, is close to what level 4 describes, and there’s still a distance to an individual pulling that off. DARS isn’t tied directly to performance reviews, but it is a recommended element in the grade definitions and goes into half-year goal setting. The way Developer A feels it, the message that it will be part of reviews someday is coming through loud and clear.
#Tighten the cost, loosen the tools
The flow of cost shows the stages of the year. Right after the declaration, headquarters paid for all AI usage. Under a use-as-much-as-you-want policy the whole company used the Vertex AI API, and with billing per token, some individuals passed ₩1,000,000 a month.
From April 2026 the cost moved from headquarters to each service organization. What an individual could spend shrank, and in exchange the company picked tools people could use without limit. They’re Devin, an autonomous coding agent that takes a task and plans, codes, and opens the PR on its own, and CodeRabbit, a code review tool. Devin is the tool DeNA announced in March 2026 it had rolled out company-wide to more than 2,000 employees. The period of broadly testing what’s possible is over, and the company has entered the stage of tightening the cost and loosening the tools.
Since Devin came in, GitHub PRs pile up 200 comments each. Every conversation with Devin stays on record, a process that was invisible when people worked locally. It means the work has become transparent, and it also means who did how much of what is visible to everyone.
On Developer A’s team, the order of development changed too. It used to be build a mockup, get it reviewed, then start development. Now the default is to build it with AI first and fix it from there, because fixing something already built is faster than polishing a mockup.
#Nobody reported having people to spare
By the numbers, it’s a success. But one thing didn’t go to plan. The efficiency came, and the people didn’t move as much as expected.
As working hours shrank, employees filled the gap with work they’d wanted to do but couldn’t. When time opens up, people reach for what they’d been putting off first. So nobody reported having people to spare, and the original goal of moving talent into new businesses went slower than expected.
Namba’s answer was blunt. She said a certain amount of rough leadership was needed. Rather than wait for the floor to move, she would flip the order, moving people first from the top down and filling the vacated seats with AI. She also said manager evaluations would reflect how many people they’d sent to new businesses.
The half-the-headcount message reached hiring too. Developer A’s team hasn’t hired a new graduate in two years. Once that line was out, aggressive hiring was off the table, and when they do hire, it’s mostly experienced engineers who can contribute right away. Developer A asked where Japan’s developers were supposed to go if a company this size stopped hiring new graduates. As AI raises productivity, the room for juniors to grow is narrowing.
Even calling it the floor, what this post saw is one developer’s team. Still, granting the whole temperature gap between the stage and the floor, it’s a fact that the order of building has flipped, and nobody around Developer A works the way they did a year ago. That doesn’t mean one side is wrong. It’s the difference between watching the same year from above and from inside. From either side, the point of no return is behind them.
Even after the iPhone arrived in Japan in 2008, the mobile transition took years. Tools change in a day, but it takes time for an organization to reshape itself around them. The keynote presented the 95% as a case from a specific project, and Developer A’s 50% shows where the teams outside it stand. The questions left are how fast the room freed up by efficiency gets filled, and who the fruits go to.
#References
- Full text of Namba Tomoko’s AI Day 2026 keynote (Full Swing by DeNA): the official transcript, including the passage where the 95% and the 20x are the same project’s case.
- Full text of Namba Tomoko’s AI Day 2026 keynote (Engineer type): the complete keynote, including the three stages of AI use.
- DeNA AI 100 Drills (PDF): 100 use cases that came up from the floor.
- DeNA AI Readiness Score official announcement: the five-step DARS criteria and the definitions of individual and organizational levels.
- Interview on the background of DARS (Full Swing by DeNA): the reasons for adopting the metric, from the people who built it.