Cracking the Mainframe: How Fujitsu Kozuchi is Automating the Death of Legacy COBOL

For most global banks, insurance companies, and government agencies, the greatest existential threat isn't a cyberattack—it's **COBOL**. An estimated $3 trillion in daily transactions still flow through code written in the 1970s and 80s, often by developers who have long since retired. The primary barrier to modernization isn't the migration itself; it's the fact that nobody knows how the original systems work. Fujitsu’s **Kozuchi AI** is changing the math by providing an automated 'X-ray' for legacy codebases.

The Problem of "Spaghetti" Logic

Mainframe codebases are notorious for "spaghetti logic"—thousands of GOTO statements, non-standard JCL (Job Control Language) scripts, and undocumented inter-dependencies across millions of lines of code. Traditional LLMs fail here because they lack the **symbolic reasoning** required to trace the data flow through rigid mainframe architectures. You cannot simply "prompt" your way out of a 5-million-line banking core.

Fujitsu Kozuchi uses a **Hybrid AI** approach. It combines a large-scale transformer model with a **symbolic solver** that performs static analysis on the source code. This symbolic engine builds a complete "Control Flow Graph" (CFG) of the COBOL application, which the generative AI then uses to synthesize human-readable documentation. This ensures that the AI's explanation isn't just a guess—it's mathematically grounded in the actual execution paths of the code.

97% Efficiency: Breaking Down the Numbers

In a recent pilot with a major Japanese financial institution, Kozuchi was tasked with documenting a legacy settlement system. Traditionally, this would have required a team of 20 senior architects and COBOL specialists roughly six months to complete. Kozuchi finished the initial documentation and dependency mapping in **48 hours**. The subsequent human review phase, which usually takes months, was reduced to a few days because the AI provided "traceability links" directly from the documentation back to the specific lines of code.

The **97% reduction** in manual work refers specifically to the "Impact Analysis" phase. When a developer needs to change a single field in a COBOL record, Kozuchi can instantly identify every other program, database table, and report that will be affected by that change. In the past, this was a manual, error-prone process that often resulted in catastrophic system failures when a "hidden" dependency was missed.

From Documentation to Refactoring

Kozuchi doesn't stop at documentation. The platform also includes a **"Refactoring Assistant"** that can suggest modern Java or C# equivalents for specific COBOL subroutines. Crucially, the AI generates **automated unit tests** for both the old and new code, ensuring that the logic remains identical during the transition. This "side-by-side" verification is the holy grail of mainframe migration.

Fujitsu is also training Kozuchi on **Proprietary Mainframe Dialects**. Many enterprises use custom versions of COBOL that include non-standard extensions. By fine-tuning Kozuchi on these specific enterprise "dialects," Fujitsu is able to achieve higher accuracy than general-purpose coding assistants like GitHub Copilot, which are primarily trained on open-source modern languages.

Conclusion: The Great Unlocking

The ability to document and understand legacy code at scale is a "great unlocking" for the global economy. As billions of lines of COBOL are finally made transparent through platforms like Kozuchi, the risk profile of our financial and social infrastructure will decrease dramatically. We are moving from an era of "maintaining the black box" to an era of "intelligent modernization." For the first time in fifty years, the mainframe is no longer a hostage-taker—it's just another system that can be understood, managed, and eventually replaced.