Most AI proposals I read assume something they never state: that your data is fit to use.
It usually isn't. The same customer sits in three systems under three different IDs, and the person who reconciled them has left.
That was survivable when the data fed a monthly report someone read and quietly corrected. It isn't when it feeds something that answers in a second and sounds certain.
Fixing it is slow and unglamorous, so no supplier's proposal includes it. It ends up in nobody's scope and turns up in month two as a surprise.
When I built an assistant to size deals, the first rule I gave it was to refuse a server inventory that hadn't passed a data-quality check.
Ask a supplier what state your data has to be in for their numbers to hold. The answer tells you how much of the work is actually theirs.