DORA調査:AI導入から効果的なSDLC活用へ、緊張関係をどう調整するか Balancing AI tensions: Moving from AI adoption to effective SDLC use

Google's DevOps Research and Assessment (DORA) team has published a new insight piece focused on the tensions organizations encounter when embedding AI into the software development lifecycle (SDLC), arguing that the real challenge is no longer adoption but effective use.

For more than a decade DORA has shaped how the industry thinks about software delivery performance through its Four Keys metrics — deployment frequency, lead time for changes, change failure rate, and time to restore service — and through the annual Accelerate State of DevOps Report. This latest piece extends that lineage into the AI era, where coding assistants and increasingly autonomous agents are reshaping daily developer workflows. The framing suggests that simply tracking AI tool adoption rates tells leaders very little about whether those tools are actually improving outcomes.

The core argument centers on a set of trade-offs, or tensions, that surface as AI moves deeper into engineering work. These include speed of generation versus output quality, developer autonomy versus governance and policy, trust in AI suggestions versus the need for human verification, and individual productivity gains versus team-level learning, code health, and long-term maintainability. DORA appears to position these not as binary choices but as ongoing balances each organization must calibrate based on its risk profile, domain, and engineering culture.

The context for this discussion is well established. Tools such as GitHub Copilot, Cursor, Anthropic's Claude Code, and Google's own Gemini Code Assist have spread rapidly across professional development teams. Yet research from DORA itself, as well as from McKinsey, GitHub, and various academic groups, has repeatedly shown that raw productivity metrics like lines generated or pull request throughput can mask downstream costs: heavier review burdens, subtler bugs, security regressions, and erosion of shared understanding within teams. The link between adoption percentages and business value has proven less direct than early vendor narratives implied.

DORA's pivot toward effective use rather than adoption echoes a long-standing DevOps principle: tools matter less than the processes, culture, and feedback loops that surround them. The same playbook that distinguished elite performers in earlier State of DevOps reports — small batch sizes, fast feedback, psychological safety, strong testing and observability — likely remains the foundation on which AI productivity gains either compound or evaporate.

Looking forward, as AI agents take on more autonomous coding, review, and operational tasks, the measurement frameworks themselves may need to evolve. Questions about attribution (who or what authored a change), accountability when AI-generated code fails in production, and how to evaluate team capability when much of the typing is done by a model are likely to become central. DORA's choice to frame the current moment around tensions rather than prescriptions suggests that definitive best practices are still emerging, and that organizations willing to experiment thoughtfully — while measuring outcomes, not just usage — will be best positioned to benefit.