How clean does our data need to be before we even think about AI?

If your data is not cleaned, organized and labeled properly, the final output will be poor. That said, it doesn't have to be perfect before you start. Some mess is managed as you go along.

But if your records are full of bits and pieces of data without labels, or full of duplicate data, AI is just going to be all over the place. The best approach is to start small. Clean up a slice of the data you care about most, and let that pilot teach you where the rest of the skeletons are hiding.