The last five times I've looked at something in case it was a legitimate user email it was AI promotion of someone just like in the article.
Their only way to escalate, apart from pure volume, is to take pains to intentionally emulate the signals of someone who's a legitimate user needing help or having a complaint. Logically, if you want to pursue the adversarial nature of this farther, the AIs will have to be trained to study up and mimic the dialogue trees of legitimate users needing support, only to introduce their promotion after I've done several exchanges of seemingly legitimate support work, in the guise of a friend and happy customer. All pretend, to get to the pitch. AI's already capable of this if directed adeptly enough. You could write a script for it by asking AI for a script to do exactly this social exploit.
By then I'll be locked in a room that's also a Faraday cage, poking products through a slot in the door—and mocking my captors with the em-dashes I used back when I was one of the people THEY learned em-dashes from.
One thing about it, it's a very modern sort of dystopia!