Over the past three decades, the IT division of one of the leading automotive manufacturers has developed a system to support sales business processes. This system runs on a mainframe, comprises around 15 million lines of COBOL code, and uses an Integrated Database Management System (IDMS) — a technology stack for which skilled engineers are scarce.
To modernize this vital system and ensure it continued to deliver value, the company set up a program to migrate it to a microservices architecture and modern technology stack, which involved rewriting the entire codebase.
The program’s goal is to move off the mainframe by the end of 2025 to avoid licensing issues. But rewriting 15 million lines of code is a significant undertaking, and with the program running behind schedule, the automotive giant turned to Thoughtworks for help.
Reverse engineering 15 million lines of code delays progress
Prior to Thoughtworks’ involvement, the company had identified the reverse engineering process as the main factor causing the delay. The average lead time to reverse engineer 10,000 lines of code was six weeks, using two FTEs and a subject matter expert (SME) for review.
Since COBOL engineers were hard to find, the R&D team turned to third-party developers to reverse engineer the code. However, only a handful of the team from that third-party had specific contextual knowledge of the manufacturer and the unit’s domain, and business analysts were struggling to extract useful information from the technical documentation others had produced. The documentation had to be interpreted by SMEs, but these individuals were spread too thinly across multiple teams.
Prior to engaging Thoughtworks, in an effort to accelerate the program, the company explored using generative AI (genAI) to reverse engineer the legacy codebase. However, the genAI outputs didn’t meet the business analysts’ requirements and the company’s technical SMEs found they were only 60% accurate.
Identifying the blockers — and determining how to overcome them
The Thoughtworks team began the engagement by running a value stream mapping session looking at the end-to-end reverse engineering process. This gave us insights into overall process efficiency, resource allocation and major bottlenecks, and it was clear that the primary bottleneck was the availability of SMEs.
As a result, we suggested that by reducing SME time spent reviewing inaccurate reports and providing further input and clarity on the as-is state, they could instead spend more time providing input for forward engineering design and the to-be state.This would increase the ability of business analysts and SMEs to work faster or support multiple workstreams, helping get the program back on track to meet its goal.
To test this hypothesis, we conducted a proof of concept (PoC) using Thoughtworks’ CodeConcise Legacy Assistant to automate document generation, and provide a genAI chatbot to answer questions. CodeConcise is a genAI-based accelerator that leverages a Large Language Model (LLM) and a knowledge graph derived from Abstract Syntax Trees (ASTs) to analyze and document legacy code. It helps organizations address the complex challenges of legacy modernization and accelerate enterprise modernization by quickly providing a greater understanding of existing systems and reducing SME dependency.
Importantly, we worked closely with the company’s business analysts to gain a complete understanding of the technical documentation they received for each module and identify the critical information they needed to complete their work.
CodeConcise dramatically accelerates reverse engineering
During the PoC, the Thoughtworks team conducted an experiment to enhance the CodeConcise code comprehension pipeline using abstract syntax trees (ASTs) and genAI. The aim was to generate documentation faster, while enhancing its clarity and quality to meet users’ needs.
The company estimated that this would reduce the time and cost of documentation production significantly; what would have taken two FTEs over four weeks could now be done automatically in a few hours.
We also tested the CodeConcise chatbot’s ability to handle questions that business analysts typically directed to SMEs for requirements clarification. Using ASTs, the tool is able to extract logical units of the code (packages, classes, functions) and visualize how they relate to each other using a Knowledge Graph. This then helps LLMs focus better on relevant, related parts of the code, independently of how they’ve been organized by developers.
Smarter, faster reverse engineering
CodeConcise slashed manual effort — freeing teams to focus on higher-value work.
66% time reduction
Cut reverse engineering time from 6 weeks to 2 per module — thanks to CodeConcise and RAG.
60,000 person-days saved
Massive efficiency gains across 15M+ lines of COBOL — driving real business impact.
Scaling modernization with confidence
RAG helps teams move faster with fewer blockers — and big savings.
With this improved database, the Retrieval-Augmented Generation (RAG) system can find better, more relevant information about the code. As a result, it provides clearer and more helpful answers to users. So far, it’s helped the company reduce reverse engineering time by two-thirds — from six weeks to around two weeks for each 10,000-line module. For reverse engineering all 15 million lines of code, this could save the company 60,000 person-days.
Driving into the future with genAI
By accelerating reverse engineering and freeing SMEs to support parallel workstreams, the PoC has shown how the company can get its mainframe modernization program back on track.
The success of the PoC has given the manufacturer confidence in using genAI to accelerate the program, knowing that the outputs from CodeConcise will help business analysts rather than hinder them. After the successful PoC, we transitioned it to a production-ready solution and deployed it, supporting the revival of their modernization program.
While this implementation focused on low-level implementation details to support reverse engineering of the COBOL codebase, the Thoughtworks team has also explored other areas that could be addressed by CodeConcise. For example, the tool could help lead developers better understand legacy system design, or help business architects extract information from codebases with minimal SME input.
We’re delighted with the outcome of this implementation, and we look forward to using CodeConcise to help our client meet its modernization program goals. We’re also excited to explore opportunities to enhance the tool further, particularly in modernizing in-car software written in C and C++, to support the company’s ambitious modernization initiatives.
