AI-assisted migration: Critical lessons from the modernization frontlines

· Source: Thoughtworks Insights · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Advanced, medium

Summary

Thoughtworks and AWS ProServe partnered to implement an AI-assisted framework for a leading sports data and technology company, addressing delays in modernizing a critical platform. The challenge involved manually extracting business logic from a legacy Java codebase for over 80 sports. The solution comprised four key components: "Golden Rules" to prevent AI hallucination and flag ambiguities with green, amber, or red confidence markers and source traceability; a phased approach separating expensive code analysis (30-60 minutes per module) from fast specification generation (5-15 minutes); a shared context layer utilizing steering files and Architecture Decision Records (ADRs); and direct code generation from structured specifications. This framework reduced a 10-sport migration from 2-3 years to 3-4 weeks, or 1-2 days with parallel execution, and cut new sport onboarding from 10-15 weeks to less than a day.

Key takeaway

For Directors of AI/ML overseeing large-scale platform modernization, recognize that AI can drastically cut migration timelines, but only when integrated within a disciplined framework. You should prioritize establishing clear "Golden Rules" for AI, ensuring source traceability, and building a shared context layer to prevent errors and maintain consistency. This approach transforms undocumented tribal knowledge into reviewable assets, accelerating migration from years to weeks while preserving accuracy and directing expert human review effectively.

Key insights

AI-assisted migration frameworks, guided by human expertise and structured context, significantly accelerate complex platform modernization.

Principles

Method

Implement Golden Rules for AI, use a phased approach for analysis and spec generation, maintain a shared context layer with steering files and ADRs, and generate new code directly from structured specs.

In practice

Topics

Best for: AI Engineer, MLOps Engineer, Director of AI/ML

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Thoughtworks Insights.