Interesting. I did this for medical records, mid-aughts.
The phrase I used was "transform & translate at the last possible moment". Ideally in the reporting step (server side rendering the HTML).
Since all of our customers kept changing their minds, totally their prerogative, my strategy minimized cost of change and turnaround time. Prior products would have to "rerun the data" from scratch, for schema changes. We'd make the same query tweaks, but at the end instead of the beginning of the pipeline. (Hardest part was always nailing down the requirements, expectations. Then the code basically wrote itself.)
This strategy made people's heads explode. Old schoolers just couldn't abide by it. Despite our small team running circles around everyone else. Data must be normalized! All the same! We had to be cheating somehow. They did everything possible to destroy us, and eventually succeeded.
I end up just doing both - having one table store the blob, and another with the copied values (shopify_sales, shopify_sales_jsonb).
Databases are generally pretty good at handling large BLOBS/off-page heap values, but I prefer not to tempt the fates of buffer pool management.
But I suspect OP's point is probably that you do at least impose a schema on import rather than letting it just be featureless JSON blobs or EAV tuples, which makes the job of understanding that data later so much easier.