Tabular should definitely be helpful in many tasks. It's more structured than just spreadsheets, though. If there's good documentation for workflows with it, it's rather hard to find though. I personally suspect that it simply doesn't exist yet...

Yes, it will be very useful for AI. It implements what already exists in other languages and needs to exist in Wolfram Language as well. There are very optimized ways of reading tabular file formats such as Parquet, and representing the data efficiently in memory. I thought that Dataset would be the way to represent tabular data in Wolfram Language, but they probably missed out on the optimizations and had to start on something new to include them.

The post mentions that Dataset is more general than Tabular in that it supports hierachical data. However, the most efficient ways of working with tabular data in Python such as pandas and polars support nested data. The file format Parquet itself also supports nested data. So the reason for tabular not supporting nested data should not be efficiency, I don't think.