Do you understand the way your software splits (or not) the client requests, i.e. can you say "if a section A of the output CSV looks like X, then I can break it up by (move this here, that there) to make that section look the same as section B" and/or "if it looks like Y, I leave it alone" ?
If so, then writing a preprocessing script is the way to proceed, and actually this is the reason why I use a Mac, a good chunk of my time involves writing UNIX shell script to preprocess, then Excel workbooks where I write formulae and then array formulae to process, the various text/log/CSV files that come my way as (data).
The coding of the scripts (bash calling the usual grep, awk, sed &c with jot loops aplenty) is to ensure consistentcy in the something
.csv which I can then open up in Excel and make arrays of, then write Excel formulae in sheets 2 and 3 (information), then plot (knowledge) ready for use in PowerPoint (wisdom!). FML.
TL;DL No matter how awful the output, if it's internally logical or semi-consistent then it's all fairly straightforward via moderate effort with UNIX scripting and Excel-fu capabilities, and once you've written the preprocessor you're sorted for future reports from that source.
If you're only on Windows, my sympathies and I hope this is similarly tractable, if perhaps more painful, in PowersHell