One capability AI might bring to MDM is name synonyms.  One of the problems computers have is that "Rich" and "Richard" are different strings, but they are variations of the same name.  This leads to a lot of false negatives in matching data from informal sources to data from formal sources.  Same goes for Will/Bill for William, Bob/Rob for Robert, and Tom/Thom for Thomas.  Where things get crazy is non obvious names, such as Liz/Beth/Bess for Elizabeth, or Jack for John (ref: President Kennedy).  An LLM, especially one which is very familiar with names, could make that part much easier.

Phone numbers and addresses conform to specific rules, and can be formatted with simple algorithms or by referencing a central database (like USPS CASS for addresses).  Even the road name synonyms in the US are just reference data, no AI needed really.

MDM matching is very rules driven, and not every enterprise will have the same rules or tolerance for false positives or false negatives.  I can't see an LLM taking over the matching and data survivorship, they're just not built for that.  After working with LangChain and LangGraph the last few days, I can see how you might be able to orchestrate tools and agents to replace a traditional MDM or business rules engine.  But you're not really using any AI/Gen AI at that point, you're just using a slightly more flexible rules engine.

If a vendor came to me and said they have a totally AI powered MDM system, I would be extremely skeptical and somewhat nervous about how well it would work.