Just have the LLM output verifiable scripts instead of manipulating the data directly itself and have version control for the data so the AI can undo changes. All pretty doable, though maybe tricky to get into old apps.