“For each turn in a conversation, Sierra's system runs approximately 20 inference calls, including online "supervisor models" that must approve an agent's response before it is sent to the user.”