A 450-Person-Month Legacy Mystery Cracked in Two: How a 100-Year-Old Sake Wholesaler Used AI
- Kakuyasu's 30-year-old core system had become an unreadable black box
- Lost design docs and missing source code put decoding at 450 person-months
- Generative AI completed the analysis in effectively two months
- It's a proven escape route from Japan's '2025 cliff' legacy crisis
Japan's most stubborn corporate-IT disease—the unopenable core system—just got a documented generative-AI cure. Century-old alcohol wholesaler Kakuyasu ran a 30-year-old backbone system with scattered design docs and partially lost source code; conventional decoding was estimated at 450 person-months. With generative AI, the analysis took roughly two months, as presented at AWS Summit Japan.
That two-orders-of-magnitude gap reveals AI's killer aptitude for legacy archaeology: reading massive codebases and reconstructing specifications is exactly what large language models do best, replacing line-by-line detective work by scarce COBOL veterans.
This is a national issue—METI's '2025 cliff' warned that aging systems devour IT budgets as their keepers retire. Kakuyasu proves a realistic path: use AI to turn the black box white first, then modernize. Legacy modernization is one of generative AI's most immediately billable enterprise uses—watch the SIer order books.