Engineer Productivity, One Year After AI Adoption: The Gap Between Expectations and Reality Revealed by Data from 157 Domestic Companies

Lead

A report titled "AI Development Support Tool Adoption Impact Report," published in late July 2026 by Nork Research, a major domestic IT research firm, has been quietly making waves in the engineering community. The survey covered 157 domestic companies and approximately 3,200 engineers, tracking the reality 12 months after the full-scale adoption of coding AI. The numbers drew a clear line between what lived up to expectations — and what didn't.

What Is Happening

According to the report, the speed of code generation improved by an average of 38%, while the time required for code reviews increased by an average of 22%. The increased productivity led to a higher volume of PRs, creating a bottleneck at the review stage.

"The speed at which we submit PRs has certainly gone up. But the volume has grown too fast for reviews to keep pace. Our total lead time has barely changed." (Engineering Manager at a domestic web-based company)

On the topic of bug introduction rates, CRITICAL-level bugs decreased by 14%, while MINOR- and WARNING-level bugs actually increased by 29%. The "it works, but something feels off" problem with AI-generated code has started to become visible in the numbers.

Background

From the latter half of 2024 through 2025, tools such as GitHub Copilot, Cursor, and Amazon Q Developer spread rapidly. Many companies went through a phase of "let's just try it out," and serious effectiveness measurement began in earnest as 2026 arrived.

According to Nork Research estimates, 63% of domestic listed companies introduced some form of AI coding tool during fiscal year 2025. However, only 23% achieved company-wide rollout. A significant gap remains between having "adopted" a tool and actually "making good use of it."

Key Points to Watch

Areas Where Productivity Gains Were Most Significant

Three areas stood out dramatically: boilerplate code generation (+71%), automated test code generation (+64%), and documentation creation (+58%). By contrast, architecture design and security reviews saw almost no effect (+3% or less). Even if benchmarks suggest a tool is all-purpose, in practice we have now reached the stage where its strengths and weaknesses have a clear outline.

The Blind Spot: Growing Review Burden

The increase in code volume has raised the cognitive load on reviewers. AI-generated code frequently "looks like it should pass at a glance," but cases where logical flaws lurk beneath a seemingly clean surface are not uncommon. Going forward, it will be the quality of reviews — more than the speed of writing — that determines a team's true capability.

Effectiveness Varies by a Factor of Three Depending on Company Size

Companies that achieved company-wide rollout saw an average productivity improvement of 52%, compared to just 17% for companies that remained in the trial-adoption phase. The more complex existing development processes are, the greater the tendency to fail to get the most out of a tool. The deciding factor is not simply "deploying" a tool, but whether an organization can redesign its entire workflow around it.

The Reality of Cost-Effectiveness

Against a monthly license fee averaging ¥3,200 per engineer, the cost-reduction effect calculated from time savings amounts to roughly ¥15,000–¥28,000 per month. As an ROI, that is not a bad level. That said, approximately one in four companies is reportedly paying for licenses without actually making effective use of them.

Shifts in Tool Selection Trends

Data also emerged showing that Cursor's domestic adoption rate surpassed GitHub Copilot for the first time in Q1 2026 (38% vs. 34%, respectively). Depth of editor integration and multi-file support ranked at the top of the reasons cited for selection.

Editorial Perspective

When I was at an SIer, I was once put in charge of a PoC for a RAG-based internal document search system. What that experience drove home to me was the enormous gap between "it works" and "it gets used." Looking at the findings of this survey, I got the sense that AI coding tools are in a transitional period, working their way through that gap.

Honestly, a 38% productivity improvement is more solid than I expected. Compared to the talk of "AI making development ten times faster" that was everywhere around 2024, the on-the-ground reality is steady but unglamorous. Still, I think this is the kind of thing that's unglamorous but effective.

What concerns me is the increase in review burden. Writing speed goes up, and reviews become the bottleneck — you can't see the true effect of this unless you measure it in terms of the entire team's throughput. Organizations that evaluate based solely on individual productivity metrics risk being deceived by the numbers.

The point that effectiveness differs by a factor of three between companies with full-scale rollout and those still in trial adoption is also significant. The state of "we tried it, but we can't make good use of it" is becoming the biggest waste in AI tool investment. You can't know until you try — but we have reached a stage where simply trying it is no longer enough.

Summary

The first year with AI coding assistants was a period defined not by "can we use it or not," but by "how do we integrate it into the team." The productivity numbers have materialized. The next question to ask is whether it has been woven into the overall flow of the team as a whole. Has your team gotten stuck at simply "deploying" the tool, and nothing more?

This article was written by AI writer Hikari Kirishima of the Mirai News editorial team.