Thoughtworks has successfully reduced legacy-to-modern migration timelines from years to weeks across four engagements, including a Java sports platform, 15 million lines of automotive COBOL, a black-box Windows binary pilot, and a .NET4 B2B retail platform. The approach was informed by an automotive client's pre-engagement experiment, where unstructured generative AI achieved roughly 60 percent accuracy, leading Thoughtworks to implement guardrails to prevent confident hallucinations.
In the largest engagement, a sports data technology company modernized a Java platform serving over 80 sports with AWS ProServe. Thoughtworks developed a four-component extraction framework that included "Golden Rules" to flag ambiguity and attach a confidence marker and file:line citation to each extracted fact. A phased pipeline separated expensive codebase analysis from cheap spec generation, allowing reformatting without re-analysis. Shared steering files and architecture decision records preserved context across sessions, enabling each sport to inherit prior structural choices.
Operational numbers from the engagement show a ten-sport migration program that previously required two to three years was compressed to three or four weeks of total effort, or one to two days executed in parallel. Per-sport onboarding dropped from ten to fifteen weeks to under one day because each sport inherits shared templates, significantly reducing the budgeted time for setup, extraction, and SME review. The resulting specs were structured enough to feed directly into code generation, revealing undocumented business logic without a human-written equivalent.
The pattern held for more challenging substrates. For an automotive manufacturer facing a 2025 mainframe retirement deadline, Thoughtworks applied an AST plus RAG pipeline to 15 million lines of COBOL, visualizing code as a knowledge graph before retrieval. With the structured pipeline, reverse engineering time for a 10,000-line module fell 66 percent, from six weeks to approximately two weeks, representing a potential saving of 60,000 person-days for the full codebase. The figures come from a PoC that has since moved into phase 2 production.
In a separate two-week pilot, a five-person team used Gemini 2.5 Pro against a black-box Windows stack spanning one of 24 business domains—650 tables, 1,200 stored procedures, 350 screens, and 45 compiled DLLs, with ASP frontends and no source—decompiling binaries and validating a functional blueprint with domain experts. A fourth engagement replaced open-ended Copilot prompts with version-controlled YAML/JSON instruction files encoding deterministic migration rules in a .NET4 B2B retail platform. At traditional velocity—about two controllers per sprint per developer—the client estimated full migration of 25-plus APIs would take roughly ten years; deterministic instructions are intended to collapse that timeline.
The challenge in every case is the SME bottleneck. The 60 percent figure reflects the client's unassisted baseline, not Thoughtworks' output; the framework's actual job is to triage ambiguity so humans resolve only flagged amber and red items rather than verifying every extraction. In the black-box pilot, decompilation and ASM analysis were still required to separate system functions from business logic, and every hypothesis required cross-source validation to close coverage gaps. The sports data client's senior product manager emphasized that SMEs cannot spend time investigating every AI-extracted fact, confirming that the guardrails exist to focus attention, not remove it.
Treat the AI as a noisy compiler that requires deterministic guardrails—source traceability to file:line, confidence tiering that forces human escalation, immutable shared context, and strict separation of expensive analysis from cheap regeneration—not as an expert that replaces them.
Written and edited by AI agents · Methodology