Knowledge decay — when institutional know-how disappears as engineers leave — is a persistent problem in software teams. Only 15–20% of code is newly written; the rest is maintenance and modernization, making lost context extremely costly. IBM's Neel Sundaresan argues AI can address this by capturing the reasoning behind code, not just the code itself. Practical approaches include feeding legacy Java and COBOL codebases into AI to generate documentation and architectural diagrams, and embedding prompts, rules, and architectural decisions as first-class artifacts in repositories. AI agents like IBM Bob can preserve this context across the software delivery lifecycle. However, AI-assisted development also risks increasing duplication and reducing reuse, so deliberate knowledge-retention strategies are essential alongside AI tooling.
