DORA 2025年の振り返り：AI時代の開発生産性研究の進化 DORA 2025: Year in review

Google's DORA (DevOps Research and Assessment) team has published its 2025 year-in-review, summarizing a year in which the long-running research program visibly pivoted toward AI-assisted software development as its central theme. After more than a decade of shaping how the industry talks about delivery performance, DORA's research center of gravity is shifting to match where engineering organizations are actually spending their attention.

The headline output of the year is the State of AI-assisted Software Development Report, the spiritual successor to the long-running Accelerate State of DevOps Report. Rather than measuring DevOps adoption broadly, this edition focuses on how AI coding assistants are reshaping developer workflows, code quality, delivery throughput, and developer experience. A recurring theme in the broader industry discourse has been that while AI tooling adoption is now near-universal among professional developers, measurable productivity gains vary widely between organizations. DORA is explicitly trying to explain that variance with longitudinal, survey-based evidence.

The second major contribution is the DORA AI Capabilities Model, a framework describing the organizational capabilities that appear to determine whether AI investments actually pay off. It points to factors such as a clear AI strategy, strong internal platforms, a reliable data ecosystem, user-centric product thinking, and fast feedback loops. If the classic DORA Four Keys (deployment frequency, lead time for changes, change failure rate, and time to restore service) are outcome metrics, the AI Capabilities Model is best understood as a model of leading indicators — the practices and structures that plausibly cause better outcomes when AI is introduced into the SDLC.

This work lands in a crowded landscape. GitHub Copilot, Cursor, Claude Code, Amazon Q Developer, and Google's own Gemini Code Assist are competing aggressively, while organizations such as GitHub, McKinsey, Atlassian, and various academic groups have published their own studies on AI developer productivity, sometimes with conflicting conclusions about acceptance rates, time savings, and quality impact. DORA's comparative advantage, arguably, is more than a decade of longitudinal data and a relatively vendor-neutral, academically grounded methodology, which makes its findings useful as a reference point even for teams that disagree with specific recommendations.

There is also a clear connection to the broader platform engineering conversation. Internal developer platforms, golden paths, and well-curated data are exactly the substrate that AI coding tools need to be effective; without them, AI assistants can amplify existing dysfunction rather than resolve it. The DORA framing — that capability maturity, not tool choice, is the real lever — echoes what platform engineering advocates have argued for several years.

Looking ahead, it seems likely that DORA will continue refining the AI Capabilities Model, possibly correlating specific capabilities with hard delivery and business outcomes in future reports. There may also be deeper integration with adjacent frameworks like SPACE and DevEx, given the overlap in measuring developer effectiveness. For engineering leaders, the practical takeaway from the 2025 review is that getting value from AI in software development is increasingly being framed not as a procurement question but as an organizational capability question — one that rewards investment in platforms, data, and feedback loops as much as in models themselves.