AI can write AL in seconds, and for a year we let it, on real Business Central work. Some of that was the best productivity change in a decade. Some of it cost us more time than writing the code by hand ever would have. This session is the honest field report: where spec-driven development and AI-assisted workflows genuinely accelerated us, where they quietly slowed us down, and the failure modes nobody warns you about.

You will see the real ones, not a tidy demo. Plausible-but-wrong AL that compiled, read well, and mishandled dimensions. Specifications that drifted from the code until the spec itself became a lie people trusted. Tests that were green and coverage that looked healthy because the AI wrote tests that mirrored its own mistake. And the 80/20 trap, where AI does the easy 80 percent in minutes and the last 20 percent takes longer than doing it by hand. We will also show where AI clearly won, including the 2026 AL agent tools closing the compile-fix loop.

You will leave knowing which guardrails survived contact with reality, where a human still has to look, and a short list of changes you can apply Monday morning.