Big data, for example, could reveal the dominance of certain strategies and ignorance of others, or what kinds of cards have high "ragequit %"s, or trends like the perceived value of Silver and Gold. New cards could be designed to work well with overlooked existing cards, which really adds value to the game. The most popular cards (cards that get bought relatively high compared to their win %, or low rage quit %s) could be revisited to improve their chances of appearing in full random.
There's also immense value in doing simulations for cards that are likely to work well with a Big Money strategy (e.g. Patrol as DXV said) or are otherwise at risk of monolithic strategies, or compare far too favourably with similar cards.
Why not get super feedback and super playtesting? (that's not a question for anyone here)
We're already doing all of this in the Dominion community with existing cards. (And you won't be getting big data out of playtesting games.) Geronimoo does
interesting simulations, while markus attempts to answer
interesting questions about usage of different cards, culling from the stats of thousands of games. Donald X. is aware of all of those discussions, and takes part in them from time to time on Discord. If these "big data" observations aren't influencing him in his card creation, that's mostly because we either haven't generated enough interesting insights or those insights don't indicate anything of relevance to new cards that don't exist yet. I would suspect that the insights have influenced him, though much more subtly than him thinking, "let me see how many times people play Villages they buy on average and set Acting Troupe to give that many Villagers."