Not having lists means they can collect more training data.
Build a database of all the attempted interactions. Cluster them by task. Sort by most used (or most monetizable) that the system can’t support today. Bam! You’ve got a rough futures capabilities roadmap.
It’s more complicated of course but here you literally have a large customer base telling you what it wants, but your product can’t yet do, regularly.
Unfortunately there's also drift in behavior that comes from retraining. "OK Google, play NPR news headlines" get different results some days than others. Sometimes I get the latest hourly news, sometimes I get a robot voice reading headlines to me. Sometimes asking to dial someone calls them, sometimes it returns search results. Yadda.
Build a database of all the attempted interactions. Cluster them by task. Sort by most used (or most monetizable) that the system can’t support today. Bam! You’ve got a rough futures capabilities roadmap.
It’s more complicated of course but here you literally have a large customer base telling you what it wants, but your product can’t yet do, regularly.