A support platform receives three kinds of messages: simple FAQ lookups, billing disputes that need account data, and technical troubleshooting. A single large prompt with every instruction and a top-tier model handles all of them, but it is slow and costly for the simple FAQ traffic, and quality on billing disputes suffers because the prompt is bloated. The team wants to improve cost, latency and quality without letting the model decide its own control flow. Which design should the architect choose?