LG CNS credits Claude Code with half-cost rebuild in new Anthropic case study

A 20-year-old construction management system became a viable rebuilding project with Claude Code, LG CNS says in a new Anthropic case study. Published October 30, the account describes a seven-month migration costing roughly half a conventional rebuild. The system went live in June 2026; the new disclosure is how LG CNS says it delivered it.

The customer, an unnamed Korean construction company, relied on its old project-management system for work spanning order wins through final settlement. It faced a fixed deadline and a tightly limited budget. LG CNS estimated that conventional delivery would require at least twice the available time and money, with costs running to several billion Korean won.

LG CNS reports converting 2,888 of 2,913 APIs—the interfaces through which software components communicate—for a 99.1% completion rate. It also moved 1,340 screens from the proprietary MiPlatform frontend to React. Backend and frontend conversion ran simultaneously on round-the-clock parallel tracks, rather than waiting for one side to finish before starting the other.

Work began in November 2025 and the replacement went live in June 2026. The timing, conversion figures and savings are customer-reported results published by the tool’s supplier, not independently verified findings. The cost comparison uses an estimated conventional rebuild as its baseline.

The ROI is not simply that we did it faster. The ROI is that we made a previously difficult-to-start project possible.

Hyosup Bae, Director and Head of Build Center, LG CNS

LG CNS says it evaluated multiple models, treating public benchmarks as reference points rather than the deciding factor. Its priority was whether a model finished an assigned task. Evaluations examined planning, task sequencing, tool use and whether a model found another route after encountering an error. The account does not name the competing models or provide comparative scores.

Bae credits Claude with interpreting goals beyond explicit instructions, particularly during architecture design. But model selection was only part of the setup. The Build Center constructed an engineering workflow around Claude Code that connects legacy-system analysis, business-logic extraction, code generation, testing, verification and quality management. Typical project teams contain five to seven people, including architecture, engineering and operations roles.

• Set the direction: Engineers give Claude the project’s objective and intent, then ask it to propose an approach.
• Divide the work: The chosen approach becomes small, clearly scoped tasks defined in YAML, a structured text format.
• Pass context forward: Each stage saves its output as local files that become inputs for the next stage.

The team uses Claude Opus 4.6 for core work, including architecture, legacy analysis and complex code generation. File handoffs began as a workaround for declining quality as context accumulated. LG CNS still treats task decomposition and context chaining as central, arguing that consistent outputs and organized context matter more at enterprise scale than context-window size alone.

Bae argues that trying to prevent every possible failure with an overly elaborate control system can constrain the model and increase token consumption. LG CNS instead describes lightweight layers for validation, limiting task context and checking quality. Its approach keeps those checks while leaving room for the model to reason through a problem.

Access also shaped deployment. LG CNS uses Claude through Amazon Bedrock, allowing it to operate within customers’ existing AWS security configurations. Bae says that changes adoption discussions for regulated clients: the question becomes how to use a model already inside their security boundary.

Claude Code is now the standard development approach across the 200-engineer Build Center and underpins its Build Factory modernization service. LG CNS plans further investment in automated quality measurement against reference datasets and more repeatable validation. Bae describes senior engineers’ work as increasingly focused on defining tasks, choosing boundaries and sequencing handoffs—not simply producing the implementation details.