You don't need fully broken encryption to gain useful information. Knowing how much data is transferred, to which servers, and when (especially with details like how various endpoints will inadvertently chunk up HTTPS requests based on the details about the content or how interactive sessions will have certain back-and-forth transmit patterns) is sufficent to generate a traffic "fingerprint" which you can correlate to other users, to automated traces crawling those same servers, and otherwise get a very good sense of what a user is up to online even above and beyond just knowing which IP is being queried.

Toss that into any sort of "anomaly detection" or other such nonsense, and it's easy to create rare edge cases at an ISP level.

It's somewhat analogous to how you can sometimes "reverse" hashes like SHA256. E.g., suppose the thing you're hashing is an IPV4 address. There are only 4 billion of those, so a pre-image attack just iterating through all of them and checking the forward direction of the hash is extremely effective. TLS makes that a little more complicated since the content itself is actually hidden, but time and space side-channels give you a lot of stochastic information. You might not be able to deduce somebody's bank password, but you can probably figure out where in the bank's login flow they are and approximately what they did once they logged in.