AI doesn't need perfect data. It needs findable data
Ask where your customer list lives and the honest answer is usually three places. The CRM primarily. A spreadsheet somebody maintains in Finance. Some stuff lives in a shared Dropbox folder. None of that is unusual and none of it is a crisis.
It does become a crisis the moment you try to automate something. An automation has to know which data is real. Nobody can tell, so the project stalls in week one. It gets labeled a data problem when it's a data findability problem.
Not one system for everything, one authoritative place per kind of record. Everyone knows which.
Consistency beats cleanliness in AI. A messy field with a predictable name is workable. A clean field called something different in each system is not.
When the CRM and the spreadsheet disagree, there has to be an answer, and it has to be written down.
Most first projects need a year of history. See if you can export it (many businesses discover they can't). That's best discovered now.
A spreadsheet showing what the record is, where it lives, who maintains it. Can we export 12 months, which copy wins?
☐ Every record type has one named source of truth
☐ You could export 12 months from each system this week
☐ Somebody owns each list, by name
☐ Naming is consistent enough to join two systems on a customer
☐ You know which copy wins in a conflict