LLMs are stateless. But given an actual database of task-shaped items and some work, I could see the potential.

With a canonical source of truth, and set input/output expectations, the potential blast radius is quite small.