Field Report ①: Fully modernizing 20-year-old Java in about a week

AI case studies

The core system of a membership service (about 20,000 members, running for over 10 years). Here is the state it was in.

• Language: Java 1.7 (released in 2011, end-of-support)
• UI: JSF + PrimeFaces (a declining UI technology)
• Running on an authentication library that reached end-of-life (EOL)
• Passwords hashed with SHA-1 and no salt (dangerous by modern standards)
• No documentation. The core code was last updated 5 years ago.The original developers are no longer around

Estimated the traditional way, this is a project of "months just to decipher it, tens of millions of yen in total, measured in years." Many companies stall in front of that estimate, and the system ends up frozen in place.

In the first few days, AI took a full inventory of everything.

The first thing we did was not rewriting, butreading. We handed the entire repository and the production database to an AI agent (Claude Code), and within a few days the following came to light: the structure and real operating scale of about 670,000 rows of live data across 35 tables; design debt built up without ever being normalized; every instance of hard-coded IP addresses and plaintext passwords; and the true nature of the core decision logic that "no one knew the spec for" —in fact a simple, master-table-driven structure. The human assumption that it was "hard to decipher" turned out to be just that — an assumption.

Most of a traditional estimate is the cost not of implementation but of "deciphering." An AI agent that reads code hundreds of times faster than a human dismantles that cost first.

AI proposed the strategy; humans decided it.

AI also presented a plan to fully normalize the DB schema (the textbook ideal), but we did not adopt it, because data migration is the single biggest risk factor in a modernization project. What the humans chose was: "leave the DB schema untouched — not one table — and modernize only the application layer to the latest stack (Java 21 / Spring Boot 3)." Because there is no data migration, parallel validation is easy and we can roll back to the old system at any time.

There were three implementation highlights.

• Zero-downtime auth migration: Each member's SHA-1 password is automatically re-hashed to a secure scheme (bcrypt) the next time they log in. Security is upgraded gradually, without anyone noticing.
• Faithful recreation of the old screens: We built the new screens to the measured values of the production screens (colors, down to the pixel), driving the users' learning curve and any rise in support inquiries to zero.
• Full feature validation with real data: E2E validation with production-equivalent data before going live. AI's output is always verified against the real thing.

As a result,implementing and validating every member-facing feature plus 18 admin screens took about a week of actual work. Not a single piece of the 20 years of business logic or accumulated data was discarded. Now that development costs have dropped by an order of magnitude, escaping legacy is no longer only about "switching to SaaS and adapting your operations to it" —a full modernization that keeps your business logichas come back as a realistic option.

Are you facing the same situation?

"No one knows the spec." "The estimate is stuck at tens of millions of yen." That system — why not start byreadingit first?

Series: "Legacy to AI — Field Reports"

• Overview: 22 systems and 246,000 lines on a single PC — a measured record of 3.5 months of AI development
• Field Report ①: Fully modernizing 20-year-old Java in about a week (this article)
• Field Report ②: Adding a new entry point to a core system with zero lines of existing code changed
• Field Report ③: A month-scale architecture study, done with a 2-day prototype + real measurements

* The figures in this article are measured values as of the end of July 2026. Development outcomes vary depending on the state and requirements of the system.