Amazon and AWS transitioned their internal ticketing label system to a specialized decision model, moving away from the Claude Sonnet large language model. The switch lowered the cost of labeling internal tickets from $22,200 to $500 per million requests, which represents a 44x reduction in expense and a 4.5x increase in processing speed without affecting accuracy.
The new model cannot generate text and is limited to selecting labels from a known catalog. This specialized approach is designed for high-volume tasks with fixed options where latency is a priority, whereas generative LLMs remain necessary for tasks that require explanation text or frequent changes to the label catalog.
