I suspect that AI is in an "uncanny valley" where it is definitely good enough for some demos, but will fail pretty badly when deployed.
If it works 99% of the time, then a demo of 10 runs is 90% likely to succeed. Even if it fails, as long as it's not spectacular, you can just say "yeah, but it's getting better every day!", and "you'll still have the best 10% of your human workers in the loop".
When you go to deploy it, 99% is just not good enough. The actual users will be much more noisy than the demo executives and internal testers.
When you have a call center with 100 people taking 100 calls per day, replacing those 10,000 calls with 99% accurate AI means you have to clean up after 100 bad calls per day. Some percentage of those are going to be really terrible, like the AI did reputational damage or made expensive legally binding promises. Humans will make mistakes, but they aren't going to give away the farm or say that InsuranceCo believes it's cheaper if you die. And your 99% accurate-in-a-lab AI isn't 99% accurate in the field with someone with a heavy accent on a bad connection.
So I think that the parties all "want to believe", and to an untrained eye, AI seems "good enough" or especially "good enough for the first tier".
A big task my team did had measured accuracy in the mid 80% FWIW.
I think the line of thought in this thread is broadly correct. The most value I’ve seen in AI is problems where the cost of being wrong is low and it’s easy to verify the output.
I wonder if anyone is taking good measurements on how frequently an LLM is able to do things like route calls in a call center. My personal experience is not good and I would be surprised if they had 90% accuracy.
>I suspect that AI is in an "uncanny valley" where it is definitely good enough for some demos
Sort of a repost on my part, but the LLM's are all really good at marketing and other similar things that fool CEO's and executives. So they think it must be great at everything.
If it works 99% of the time, then a demo of 10 runs is 90% likely to succeed. Even if it fails, as long as it's not spectacular, you can just say "yeah, but it's getting better every day!", and "you'll still have the best 10% of your human workers in the loop".
When you go to deploy it, 99% is just not good enough. The actual users will be much more noisy than the demo executives and internal testers.
When you have a call center with 100 people taking 100 calls per day, replacing those 10,000 calls with 99% accurate AI means you have to clean up after 100 bad calls per day. Some percentage of those are going to be really terrible, like the AI did reputational damage or made expensive legally binding promises. Humans will make mistakes, but they aren't going to give away the farm or say that InsuranceCo believes it's cheaper if you die. And your 99% accurate-in-a-lab AI isn't 99% accurate in the field with someone with a heavy accent on a bad connection.
So I think that the parties all "want to believe", and to an untrained eye, AI seems "good enough" or especially "good enough for the first tier".