In 2025, a fintech company with 150 engineers saw a stark contrast in productivity gains from AI coding assistants. On new projects, productivity surged by 40%. However, when engineers turned to the company’s legacy payment processing system—a complex, eight-year-old codebase with 2.1 million lines—the benefit dropped to just 7%. The disparity didn’t stem from the AI's technical skills but from the lack of context surrounding the legacy code. Historical decisions, edge cases, and architectural nuances were lost to time, making it impossible for the AI to provide accurate support. The context collapse problem arises when institutional knowledge isn’t documented in accessible ways. Engineers spent an average of 11.3 hours a week answering questions from colleagues about this invisible context, costing the company about $680,000 annually. AI tools also underperformed with legacy systems, showing a suggestion acceptance rate below 15%, which meant engineers wasted time evaluating AI outputs instead of coding. An attempt to migrate a checkout flow using AI resulted in numerous failures, costing over $140,000 in engineering time alone. To bridge the gap, companies that succeeded focused on three main strategies. First, they documented context in a way that AI could understand, like architecture decision records that explained not just what was built, but why. This effort required around 400 engineering hours and led to a significant increase in AI suggestion acceptance from 15% to 61% on the legacy system. Second, they restructured code incrementally to improve AI readability. Finally, they prioritized targeted documentation to capture historical decisions and constraints, which directly impacted productivity and reduced reliance on senior engineers for context clarification.