I very much agree with this. For an integration where you have control over both ends of the pipeline, CSV is not optimal unless there's existing work to build on, and even then it's a legacy choice.

Parquet and Avro are widely supported in backend languages and also in data analysis. I don't think the article is talking about exported-like-a-jpeg, but instead exported-like-a-daily-report-run: the data scientist doing the exporting is probably using R or Pandas instead of Excel, and can reasonably be expected to read https://arrow.apache.org/docs/r/reference/read_parquet.html.