Instead there would be misinterpretations, misrouting, and deaths of despair when a model misjudges a situation. And it sounds like the model would be the first line of defense before routing, which is the WORST possible implementation of generative AI in a system like this.

Ideally, models could be used here to provide additional information to a human operator (advanced transcription/translation for cases where callers are impaired in speaking, do not speak the same language as the operator, or have a bad connection or heavy accents-- for instance).

Instead, the first line of decision-making in potentially the highest possible stakes for certain callers is delegated to a text generator with no mind, adding a delay to the process.

Hats off to the consultancy and the soulless salesperson who nailed this contract.