This seems like a fundamentally hard problem given the breadth of all the originating data sources and possibility of user entry producing non-canonical case. That said, if given the choice between a fully-normalized, lowercase dataset or the original-case data, I think I'd choose the latter (and maybe provide both fields in the response, normalized and original). Better still would be some glossary lookup of proper nouns that PDL handles.

When did thing start being lower case? Our previous data delivery was not all lowercase, so this most recent delivery has broken things for us.

We've seen this same issue where a simple title case function doesn't capture the true casing of the company name for example.