One of my specialties is implementing hosted call centers using Amazon Connect - the AWS version of the call center that Amazon uses internally.

The fully allocated cost of one call to a human agent is $3-5. That pays for a lot of inference.

▲

rhubarbtree 3 hours ago | parent [-]

pricing as an objection to ai is just cope. it might hurt the current status quo when they can no longer burn capital, but the long-term course will remain unchanged

▲

raw_anon_1111 2 hours ago | parent [-]

Prices will literally have to go up by 1500x+ in the above scenario to make it cheaper for a human agent. I assure you that AWS is big losing money running Nova Lite 2 (the model I use for speed and low latency) with their own models running on their own custom chips.

I calculated the full cost of a call to be $0.05 cents a minute when handled by AI and that includes charges for a 1800 number and the varios other AWS services it uses.

Nova does the processing of the text after a separate voice to text service (Lex)

▲

jimnotgym an hour ago | parent [-]

That's great, but it sounds like Amazon has a specific model trained for this service? And I guess that Amazon funded the development in a sustainable way. Is any of that true for Anthropic and OpenAI?

	▲	raw_anon_1111 38 minutes ago \| parent [-]
		No, its just one of the many models hosted on Bedrock. Bedrock has all of the popular models except for OpenAI which is exclusive to Microsoft. Nova Lite isn’t the best model. But what I need it for to turn random voice transcriptions into well defined JSON I can use deterministically downstream quickly, it’s perfect and fast.