I’m going to throw my hand in for flat csv files, if only because you’ve got the greatest control over the code. Just make sure that you read in the rows and process them one at a time (reading the document to the end and splitting will eat up all of your memory – same with writing, stream it out).
Yes, the user will have to save-as CSV in excel before you can process it, but perhaps this limitation can be overcome by training and providing clear instructions on the page?
Finally, when you export to the customer, if you set the mime type to text/csv, Excel is usually mapped to that type so it appears to the user to be ‘an Excel file’.