FWIW I know of a pharmacy network that's implementing AI. Not only does the AI make egregious errors like speaking in Spanish when the conversation is in English. It also has a limited context window for ground rules of 5000 tokens or less. So correcting all the various errors goes over the instructions limit. As a result if the implementation goes through it will be of poor quality. All I can think is that people with no technical expertise are making these decisions, it is easy for any techie to predict that AI of this level won't be capable.

▲

PhilipRoman 2 days ago | parent | next [-]

This is an underappreciated point when people talk about AI/chat integrations. If they put a claude-level model behind it, I'm sure it could resolve 99% of the situations faster and better than a human could, or at least realize its limitations when it can't. But of course that is too expensive so they often use extremely weak models that are literally worse than a good incremental full text search box.

▲

HPsquared 2 days ago | parent | next [-]

I wonder what the price level (per token) needs to be, to match a minimum wage employee.

	▲	themgt 2 days ago \| parent [-]
		100x higher than it is now?

▲

asdff 2 days ago | parent | prev [-]

Another aspect of the human version is that the human version is deterministic. Call center workers have a flow that is pretty much a series of if else statements with appropriate edge cases already carved out and places to break out and escalate to another provider. These flows have been iterated over for some time and refined to this point.

AI unfortunately is not deterministic. It would be nice if you could actually use it deterministically in this way, set a seed such that the logic flow are preserved across all sessions well into the future unless the logic flow instructions are changed. They just don't work like that though. Even if you have a good harness they can go off the rails pretty easily and need coaxing to get back to appropriate context. That is a nonstarter for the typical caller of a call center. Functionally speaking just an automated call where you touch type through options would be equivocal to the human operator and that paradigm has replaced a lot of call center work already, with the human operator serving to smooth things over if the person is old or can't comprehend the options for whatever reason and starts yelling incoherently into the automated line (my actual strategy for getting human operators on the line).

Everything becomes a nail though when you have a hammer though, even more so when you are selling hammers. So expect to see AI in plenty of places where it makes no sense to belong given the supremacy of existing tooling. There is also no advocate for a lot of existing tooling unlike AI that has sales teams active in pursuit of new customers integrating this into their product.

▲

munk-a 2 days ago | parent | prev [-]

This is probably (if the business here is large enough to support it) an excellent area for a tailored model instead of a general purpose LLM. A lot of the context around medication comprehension is going to be rather universal to pharmacies and other HCPs so ideally that information would be baked into the model itself with the actual user supplied context mostly being focused on business specific logic. The commodity LLMs are just an amalgam of a hammer welded to a screwdriver, a car, a compass and a toothbrush - more specialized models, with more specialized testing and offered by a company specialized in the field would possibly offer a decent customer experience here?

With something like healthcare though I'll remain eternally skeptical that anything other than a licensed expert will be palatable to the public.