Code is getting cheaper. Understanding what to build is not.

AI-assisted tools turn weeks of implementation into hours. But organizations are producing more software, faster, that doesn't match how the business actually works. Specification stays stuck in backlogs, spreadsheets, slides, and scattered notes that can't be validated or versioned. The gap between IT and business is widening, not closing.

This session shows a practical LLM-based workflow tested in production - presented from a developer's perspective, which makes it different. Both sides get a shared toolset and language that actually works.

The real skill isn't generating code - it's deciding what NOT to build. Before any code generation: understand how the business creates value, what the cost structure looks like, how work gets done in reality. Build the business case. Estimate ROI. Validate change hypotheses against actual operations. Prioritization becomes obvious: low-effort, high-impact changes first. Prove value quickly, learn what works, then tackle harder problems.

The key is making business knowledge source-controlled and queryable. When context lives in Git alongside code, both humans and AI agents can ask clarifying questions without hunting down business analysts. You'll see what good specifications look like: format, semantics, structure that works for both humans and machines - surviving code regeneration while remaining stable as implementation changes.