IMO, a key place for proxies is resource access where the client is not fully trusted (or for any reason it may fail to log correctly).

AI agents seem like a pretty clear case of this. Both because of recent issues, and that they tend to run on end user devices where strict compliance is harder to maintain.