Sony Bank cuts its development cycle by 30% by applying AI to its core system
AI entered the process, not just the code
On September 14, Sony Bank and Fujitsu announced the at-scale application of generative AI in the development of the bank's core system. The initiative started in September 2025 and, according to the companies, reached in July 2026 a 30% reduction in the period between basic design and integration testing, plus 40% fewer work hours.
The workflow uses Claude and Claude Code models through Amazon Bedrock and connects design documents, source code and test assets. The application wasn't limited to a code generation demo: it moved through design, implementation, unit tests and integration in real development.
The gain came from continuity between steps
The companies also report up to 90% fewer hours in the impact assessment of the basic design, up to 40% fewer in creating detailed design documents, a 99% code generation rate in the implementation step and up to 90% fewer hours running integration tests.
These numbers describe the Sony Bank case, not a market average. The context itself helps explain the result: the banking solution has a cloud-native, microservices-based architecture, the assets were organized for reuse and specialists followed each step. A company that just adds a copilot to the code editor doesn't automatically reproduce the same effect.
A critical system requires human judgment
The announcement says AI was used in developing a core system in operation, but kept human judgment, approval and quality assurance. That combination is decisive in environments where a specification error can reappear in the code, in the test and in the system's behavior.
The case suggests a practical decision: before buying a tool, map where rework starts between one step and the next. If the document isn't reliable, if the test doesn't cover the risk or if nobody approves the output, more speed may just spread defects.
The test starts with the baseline
A team can pick a bounded flow, record cycle time, rework, defects found after delivery and review hours, and only then introduce AI with authorized data and approval points. The pilot should move step by step, with the option to stop using it when quality drops.
The bank's lesson isn't that all programming can be automated. It's that productivity shows up when the knowledge from one step stays available to the next, with clear responsibility for the final decision.
Sources
Sony Bank e Fujitsu, 14 Sep 2026. https://global.fujitsu/en-global/pr/news/2026/09/14-01