AI demos work because the data is clean
AI demos work because the data is clean. Your data isn't, and that's the actual reason the pilot underwhelmed.
Customer records in three systems that disagree. Product names spelled four ways. A spreadsheet somebody maintains by hand that half the company depends on and nobody has documented.
Feed that to any model and you'll get confident, well written, wrong output.
The instinct is to launch a data cleanup project. Don't. Those run for eight months and get cancelled.
Do the opposite: pick one process, and clean only the data that process touches. Quote follow-ups need customer name, last order, and the open quote. That's three fields, not your entire database. You can fix three fields in a week.
Then the next process reuses what you cleaned, and the one after that. The cleanup stops being a project and becomes a side effect of shipping something.
Messy data doesn't block you. Trying to fix all of it does.