Fujitsu Launches AI-Powered Kozuchi Service: Auto-Generates Design Docs, Cuts Work Time by 97%

Fujitsu Kozuchi: AI Code Analysis Service for Legacy System Modernization

Fujitsu launched an AI-powered code analysis SaaS that automatically generates design documents from source code, achieving 97% work time reduction, 95% documentation completeness improvement, and 60% readability improvement compared to generic AI tools.

Why Legacy Systems?

Japan has one of the highest legacy system densities among developed economies. Financial institutions, manufacturers, and government agencies still run core systems in COBOL and early Java, often without documentation and with 'don't touch it' maintenance strategies.

Beyond Generic AI

Kozuchi is specifically optimized for enterprise legacy code: deep semantic understanding of COBOL, million-line-scale project handling, standardized UML documentation output, and decades of Fujitsu enterprise IT expertise built in. This represents enterprise AI tools evolving from 'coding assistance' to 'system modernization' — a larger and harder problem.

The 'Immortality' of COBOL

Why COBOL matters in 2026: approximately 800 billion lines of COBOL code remain in production globally, processing 95% of ATM transactions, 80% of real-time banking transactions, and countless government systems. The average COBOL developer age exceeds 55, creating a knowledge crisis. Rewriting costs are staggering — Commonwealth Bank of Australia spent $750M and 5 years migrating from COBOL to Java, and Japanese banking systems are typically 2-5x larger.

This 'too big and critical to touch, but too dangerous to leave alone' dilemma makes AI code analysis tools commercially invaluable — providing a middle path of AI-assisted understanding followed by modular modernization rather than full rewrite.

Comparison with Generic LLMs

Fujitsu's benchmarks against direct GPT-5/Claude usage reveal key differences: scale (GPT-5's 1M token context cannot handle 500K+ line projects; Kozuchi uses hierarchical analysis), domain accuracy (generic LLMs misinterpret mainframe-specific concepts like JCL and CICS; Kozuchi is specifically fine-tuned), and output standardization (generic LLMs produce free-text; Kozuchi generates UML class diagrams, sequence diagrams, and ER diagrams meeting enterprise standards).

Future Roadmap: From Understanding to Transformation

Fujitsu's announced 2026 roadmap extends Kozuchi beyond documentation into active code transformation:

Phase 1 (Current): Code Understanding. Automatic design document generation from source code analysis — the current release.

Phase 2 (Mid-2026): Code Refactoring Recommendations. AI-generated refactoring plans with impact analysis, dependency mapping, and risk assessment for each proposed change.

Phase 3 (Late 2026): Assisted Code Rewriting. AI-generated replacement code in modern languages (Java, Python, Go) with automated test generation to verify behavioral equivalence with the original COBOL implementation.

This phased approach is critical — attempting automated rewriting without first achieving deep understanding (Phase 1) would produce unreliable results. The fact that Fujitsu is taking this deliberate, staged approach suggests they understand the complexity of the challenge and are building foundational capabilities rather than rushing to market with incomplete solutions.