Amazon

Alexa's early architecture required explicit intent programming for every question, making it brittle and incapable of natural multi-turn conversation. Over 50% of queries that the original Alexa couldn't handle were simply abandoned by users.

Amazon rebuilt Alexa's core language understanding using large language models and neural retrieval systems. The new architecture enables multi-turn conversation, follows up on previous context, and handles novel queries without intent pre-programming.

Significantly improved response quality on open-ended queries. Reduced the failure rate on unscripted questions. Alexa now handles hundreds of millions of daily requests with a far higher rate of successful interactions.