"Deflection" in the customer service world means when you prevent a customer from talking to a human. It's a critical metric for CSR orgs. Often you improve deflection rates by putting hurdles between a customer and a rep. If you're being generous those hurdles are helpful because they consist of "Did you try this? Have you read these related articles?" and can solve people's basic problems without taking up valuable human time.<p>I am highly skeptical of this article because it doesn't mention deflection at all. There are other GPT powered B2B2C companies (e.g. <a href="https://yellow.ai/" rel="nofollow">https://yellow.ai/</a>) which are entirely focused on deflection. It's not a huge logical leap to ask, if you can use a chatbot to turn an unskilled worker into a skilled worker, what exactly is the human adding? In many cases the answer is 'not much' or 'nothing.' You can completely replace a low skilled worker with a chatbot some economically viable percentage of the time.<p>Because of this the claim that "the less experienced and less skilled workers [will] benefit the most" is suspect.
The study is about an AI support system (a "recent version of GPT", sources link to GPT4, finetuned on internal company data) supporting customer support agents at a software company that provides business process software. It increses productivity (measured in number of chats resolved by employee) by 13.8%, the lower the performance before, the higher the impact. The performance increase is mostly based on being able to handle more chats in parallel and handling each chat quicker. For reference, an agent resolved about 2.12 chats/h before and this increaded by 0.47 chats/h (I was surprised that it's "only" 2.12 chats/h).<p>Interestingly but also not surprising this increase in productivity at the lower end is atypical for tech innovation which tended to favour the higher end positions before. However, this type of customer support seems like a good candidate that will get automated away completely. It seems like the rollout can be in phases, it already increases productitivty as a support system. First, the productivity will be uplifted to a certain level for all workers where the skill difference matters less and less and finally the work will be done completely by AI.
It will be interesting to see how this company is doing 5 years. It seems that thanks to the AI system novice employees are able to cash out on skilled employees knowledge. But this system might also hinder the novices changes on gaining knowledge and becoming skilled employees. And those skilled emploees are essentialto maintain this system to respond to evolving needs. So the effect on job market might be the opposite than researcher wish: recruits need to be smarter than before and the gap between low-skilled and high-skilled employees will keep widening.<p>So in that case introducing AI to this process would have the exactly same outcome as automation has had in manufacturing.
It would be interesting to see how a general-purpose AI like GPT-4 trained on internal corporate knowledgebase would perform.<p>Theoretically, such an AI agent would know your internal policies, biases, guidelines. It would also have more knowledge about the company than any individual worker, collating data from across divisions.<p>In large corps, the silofication of knowledge is a real problem. But an AI agent trained across the company's functions would not have this issue.
This is an interesting empirical study indeed. Direct link to the study: <a href="https://www.nber.org/papers/w31161" rel="nofollow">https://www.nber.org/papers/w31161</a>
“And what this system did was it took people with just two months of experience and had them performing at the level of people with six months of experience," Brynjolfsson says. "So it got them up the learning curve a lot faster — and that led to very positive benefits for the company."<p>Literally the same story as every industrialization ever. Buckle up, we get to live through another industrial revolution.
This wasn't even GPT-4.<p>Terrific news.<p>It would be super cool if the people who were currently fretting about paperclips could focus a little more on bias, security, data, privacy, openness, and all the stuff we <i>actually</i> are going to have to deal with over the next five years.
I’ve always thought that in general, societies wants tend to expand to the limits of productivity.<p>The real danger seems to be of leadership using AI as a cudgel to threaten employees into doing what they want.
Doesn't take complex AI to benefit new workers (or even middle experience ones). A company I talked to did some simple sentiment analysis on emails, and reduced the closing time for a real estate office by 40%. Which means they can do that many more deals per month etc.<p>Maybe the real change here, is people are beginning to consider using computer aids to do their job. Instead of insisting their experience is all that matters.
While I get that people are more concerned with productivity insofar as it pertains to their job, but I think the bigger economic concern is when competing companies begin using the same models for navigating the competitive landscape. I highly suspect that we will begin to see accidental trusts form as competitors sharing models begin to optimize around each other.
I'd be A LOT more interested to see what happens to a whole set of workers who dabble in lower-skilled jobs and are now being rendered redundant with tools such as Chat GPT.