AI marketing companies discover that language models make up answers — and try to route around it

A technology startup is testing a system where AI language models no longer generate answers directly, but instead identify what the user is actually asking for and hand it to a specialized agent designed to answer that specific question. This shifts the risk from 'the AI will hallucinate' to 'the routing decision is correct,' which is a different and potentially more manageable problem.

For three years, companies have been trying to use retrieval-augmented generation — which is fancy for 'let the AI search documents and then answer based on what it found' — to stop AI from making things up. It never really worked. The AI still hallucinates. This paper is saying: stop asking the AI to answer at all. Make it a router. Hand the actual task to something built specifically to do it. If this works at scale, it means the next generation of AI-powered search engines and chatbots won't be trying to solve the hallucination problem directly, they'll be sidestepping it entirely. That's a structural shift in how companies will build these products.