Rakuten doubled bug-fixing speed with AI

Japanese technology giant Rakuten has announced significant results from integrating artificial intelligence into software development processes. The company integrated Codex — an intelligent code writing and analysis agent from OpenAI — into its engineering pipeline, and the first measurable outcomes of this step look more than convincing. The average time to resolve incidents has been cut exactly in half, and complex fullstack products that previously took months to create now go to production in just weeks.

Rakuten is one of Asia's largest technology ecosystems, combining e-commerce, financial services, streaming, telecommunications and dozens of other directions. Such a sprawling infrastructure implies a vast codebase, constant updates, inter-service integrations, and the high cost of any error. In conditions where a failure in the payment system or in the logic of the recommendation engine directly affects the revenue of millions of transactions, the speed of incident detection and resolution becomes not just an operational KPI but a strategic priority. This is why the choice of Codex as a tool does not look accidental: the company was looking for a solution capable not only of generating code but of deeply integrating into real engineering processes.

Codex from OpenAI is not just a chatbot capable of writing functions on demand. It is an agentic system capable of performing multi-step tasks: analyzing repositories, finding vulnerable code sections, proposing and even automatically applying fixes, and conducting change reviews within CI/CD pipelines. This depth of integration is precisely what distinguishes Rakuten's approach from superficial use of generative tools. Instead of simply offering suggestions to the developer, Codex becomes a full-fledged participant in the engineering process: it reviews pull requests, identifies potential regressions, and frees up team time for tasks requiring human judgment and architectural thinking.

A 50% reduction in MTTR — a figure that at first glance might seem like marketing exaggeration, but in the context of Rakuten's scale acquires quite concrete economic meaning. Each minute of an incident in a high-load e-commerce system is potentially thousands of incomplete transactions, missed conversions, and blows to user trust. If previously an average team spent, say, two hours on diagnosis, localization, and patching, now that same cycle fits into one hour. Multiply this by the frequency of incidents in an ecosystem of such scale, and the extent of savings becomes obvious. Automation of CI/CD review adds another layer: fewer bottlenecks when rolling out changes, fewer human errors in routine checks, more space for iterative development.

The implications of this case extend far beyond one company. Rakuten effectively demonstrates that AI agents in development are no longer experimental territory but a mature tool with measurable returns. For large technology organizations worldwide, this means the necessity to reconsider the very model of engineering work: the role of the developer shifts from code writing to its oversight, architectural decisions, and management of automated agents. For startups and mid-size companies, the Rakuten case opens the possibility to compete with giants, compressing product time-to-market. For the industry as a whole, this is a signal: companies not investing in such automation today risk finding themselves in a structurally disadvantaged position within the next two to three years.

Rakuten's experience with Codex is compelling evidence that the era of "AI as assistant" gives way to the era of "AI as co-author of the engineering process." When a major corporation with a multi-billion infrastructure publicly fixes a twofold acceleration of bug resolution, this ceases to be a story about experiments and becomes a story about a new industry standard. The question now is not whether to implement such tools, but how quickly the rest of the market players are ready to do so.