Org mode handles both simple plain ASCII tables (created with `C-c |`)
and Emacs `table.el` tables. The latter are basically the same as
Pandoc's markdown grid tables.

It seems Pandoc's Org reader treats `table.el` tables as Org ASCII
tables, since on conversion to `docx` the output gets mangled when
dealing with rows that contain multi-line paragraphs. The resulting
table arranges every new line of text into---often several---new rows.

On the other hand, Pandoc's markdown documents with grid tables get
what could be called proper treatment, i.e. 1 paragraph (including
multi-lines) == 1 row. This produces somewhat *cleaner*
tables. Division lines are displayed correctly too.

Pandoc's Org reader should distinguish between plain ASCII tables and
`table.el` tables and treat the latter as the markdown reader treats
grid tables.