Local speech to text has been easy to do for at least a decade and then you’re just firing off a text file to HQ to add keywords to a user file. These days an AI will likely parse the text to find recommendable products, ten years ago you’d have just had a gigantic list of all your partners’ brand names and desired key trigger phrases in a database and run the conversation text against the database and look for matches. Super easy to accomplish. Updating someone’s ad preferences 15-30 minutes after they talk about a product may as well be considered real time.