AI Isn’t Eliminating Jobs Like We Thought

Well, as we continue to see the growth of artificial intelligence, there are numerous questions being asked right now about how it's going to be impacting both our personal and our professional lives. So new research takes a look at the professional side of that question. Peter Capelli is a professor of management here at the Wharton School and was part of the research, and he joins us right now. Peter, great to have you back with us. Thanks for a couple of moments. Thank you, Dan. I guess it's probably not a surprise because it's AI and so many people are talking about it that the backstory is probably connected

to just how AI is going to probably impact the professional world moving forward. Yeah, I think that's right. And I guess I should confess that I am lousy at making predictions, so I don't really even try. But I think what is important, and we can say a lot about, is what is actually happening now as opposed to what are we guessing will happen in the future. And I think, you know, the best way to think about predicting the future is it's like the past, only a little more so, and that is whatever direction you think we're

heading in now probably going to play out for a little bit further. Okay, so then how is AI impacting the professional world right now? Yeah, so we started looking at this 10 years ago, and we sort of forget, but machine learning burst on the scene. I mean, it's been around in engineering and statistics for forever, but as a practical use in the mid-2010s or so. And data science, particularly in areas of human resources, got really kicked off when we're starting to talk about people analytics stuff, right? So we were looking at it then around machine learning, and this particular paper follows from

one that we did in 2019. We were looking at machine learning now, and we were laying out basically what seemed to be the conflicts between machine learning-like approaches and the way we manage people in particular. And the way we manage people, frankly, has a lot to do with legal issues, which all have a lot to do with fairness questions. And machine learning's got no time for fairness questions, right? And so you could imagine all kinds of conflicts that bump up. So we got asked to look at what, if we could summarize, what do we see about how AI issues now are affecting people in particular, right? So this is for the annual review of

psychology and organization behavior. Neither of those are my field, so I was a little surprised they asked me to do it, but that's what we started to do. So maybe I could just begin with a quick little analogy, which I think does help us understand a little bit about the future, and that is to look back on the introduction of search engines, like the dominant Google search engine now. And many of our listeners are probably not old enough to remember this, but before search engines, if you wanted to know something that you didn't know off the top of your head, and you didn't know who to call who might know it, and our brother-in -laws only know so much, right?

So you'd have to go to the library, and you have to find a library, and you have to find a pretty big one. And then you had to go in, and they had books of abstracts, and you had to open up the book, and you had to look at the keyword you were interested in. So it might be machine learning. And then you have to see in there, it will show you all the current articles that talked about machine learning. You then had to go up into the stacks of the library and find the physical copies, and then look at them all, and then figure out afterwards what you thought the answer was. So this question, even on a simple level,

would take days to answer, anything complicated, just to gather some of the information. And then with search engines, you could do it with a couple of keystrokes now, right? Or not even that now. And that was an incredible change. Now, here's the thing about it. If you think about that change and look back on it, I never heard anybody talk about how many jobs that was going to eliminate. What you do here now, looking back, is how many companies it created, how many new industries, like Amazon delivery, all that stuff, were created by these search engines and the ability

to find stuff quickly online. For some odd reason, when we're thinking about artificial intelligence now, particularly the ability of large language models to do things, we're fixated on it cutting jobs. And I think that's just not, if we're looking back on it, going to be the story. And we already know it's not cutting many jobs, and not nearly the way in which the pundits, who were mainly AI people themselves, were predicting it was going to. So we're not seeing any of that. We are seeing other stuff, and maybe we could talk about that. Well, okay. So, and that's obviously a big question, but I think it brings to the forefront part

of the discussion, which is how AI and the human are going to work and coexist in the workplace moving forward. And I think the question is, what does the C-suite and the leaders of companies expect to play out here over the next, and again, I know we're not, I'm not doing a prediction, but I think you have to, to a degree, look a little bit forward as to what we might see play out, because that connection between human and AI and how they're going to work together is a very important one. Yeah. So, you know, why did we get fixated on the idea that AI was going to cut headcount? You know, back to the search engines, nobody

talked about it then. I think this is really driven by investor expectations and the hope that it would cut headcount because that could save companies so much more money and improve the evaluations, etc. And the AI pundits, the people who build it and those who follow it, were pushing that argument hard. And I think the problem is CEOs and boards heard that argument and really ran with it. And they, they had this idea that AI was like, I don't want to be too pejorative here, but a little like pixie dust that you could just spread around and it would then poof, take away jobs and away

you go. Right. And we got a lot of really bad kind of advice to companies and from leaders, you know, that would just dictate, for example, you need to use 10% more AI this quarter. Literally, that's what they were being told. Right. And one of the things we know now, looking at companies and places that have introduced AI and tried to get it to work, first is that it's not a top-down exercise. It's very much a bottom-up exercise. And I think the big conclusion is that it takes tons of management help to get anything to work.

So to your particular question, if you think about what AI does, it and all these tools like search engines, they can help you do individual tasks and any job is made up of lots of different tasks. So for example, think of computer programmers, which is the one, one of the jobs that we think, oh, well, that's going to go away. Right. It turns out that programmers in organizations only spend about 30% of their time writing code. What are they doing the rest of the time? They're figuring out what are their internal clients

need. They figure out what is possible to do because what you want to do has to intersect this other stakeholder over here. You both have to agree on this. They negotiate budgets, et cetera. Right. So it does help them write code faster for sure. It doesn't eliminate the job of programmers that other 70% of work's got to be done by somebody. And what they're discovering now when they look at programming carefully is that, yeah, it generates code much faster and yeah, it adds a lot more new tasks to programmers are basically checking code. Right. And tracing down mistakes and trying to get them to work. Right.

So it didn't eliminate jobs even in that context, but it did change the way they do their work. What has the onset of the digital agent meant then for this process as we move forward here? Right. So the digital agent idea is just a large language model tool that you could pop in to take on a particular task. So let me give you an example of one of these where I saw it be introduced. And the idea was we wanted to speed up the ability to create our marketing materials. Right. And that means the goal here, pretty straightforward. It wasn't really cutting head count.

It was how can we get stuff out faster? We develop a product. Once we know where it is and it's going to do, we need to get that word out. How could you do that? Well, the first thing you got to do is you have to map out what we would call in management, the workflow. And that means, okay, you know, the project, you know what, here's what, you know, the new tool is going to do. Okay. We just got word from company. That's what it's going to do. What has to happen next? And then what happens to happen after that? And then we got to get it to marketing and what happens there. Somebody has got to sit down and map

out all those steps. And that is something most organizations have not done in forever. And in many places it's never been done for newer jobs. Then you have to think, okay, in each of those steps, who is doing that work? Then you have to sit down with that person and say, tell me precisely what you do. So this in psychology was known as job analysis. You're just digging into each job. Then you got to say with that employee, how can we stick an agent into any of these things? Any place here where a large language model would help you do something. And it's easiest to see if it's about

gathering information, like what do our competitors say they're doing right now? Okay. And then you say, okay, we're going to pop an agent in here. That employee has got to help you with that agent. That is to see what it's producing. If it's not giving you the right stuff, ask the prompt slightly differently. Got to monitor it to make sure it's not hallucinating or doing anything weird. Okay. And what do you find when that's done? It means that task has been taken over for that person. Maybe you're able to pop 10 tasks, 10 tasks taken over by large language models.

Then you have to think, okay, the jobs are different now. Do we have to recombine them in some way? It's unlikely, even if you take over a lot of tasks, that you're going to take over enough of any individual person that you don't need that person anymore. But it might be that you have to rearrange the work. Okay. So just to finish this thought, if you think about what's required there, it's a ton of management effort to get this thing to work. How then does the collection of data, which is obviously so important for so many companies, become impacted when you have work that maybe

was done by humans before, but can be done by AI through those processes now? Yeah. And I think that is one of the big questions and the big opportunities that we have not thought very much about. If someone were to ask, where's the real opportunity? It is inside these organizations themselves, the ability to use these tools to collect the information together and put it together in a framework where it could be analyzed. So that opens so many new opportunities. Right now, most organizations are drowning in data. I look particularly at the human resource area and they're overwhelmed with data that they can't

analyze. The reason they can't is because it sits in different databases. Those are not compatible with each other. Yeah. Cleaning them, putting together is a huge task. But one of the things large language models can do well is organize that data, put it into formats where it could talk to itself, each other, then you could analyze across them and get going and looking at some of that stuff. So we're not spending any much of any time on that, but it's such an obvious thing to do. What do you think then by doing this research is the thing that you want to look at going forward, but also the benefit

for an organization or a group like Organizational Psychologists who are thinking about, you know, the future of their workforces and how they work and how they react? Is there a connection there as you move forward, do you think? I think so. You know, I think the big task that organizations have to get their hands around is managing organization change. Right. So one of the things that is not going to help is to keep talking about how this is going to cut headcount and then ask your employees to help you introduce the tools that you're saying you're going to take their job away. Right. Yeah. You don't actually have to be very sophisticated

to know that that's not going to work very well. I think the other thing that we're finding is that the idea that low skilled jobs are going to be eliminated first and simplest is probably not true. And the reason for that is a lot of low skilled jobs have a lot of yes, no questions to them. You know, what is this paper coming in? Is it an appeal? Is it a new claim? That sort of thing. And the reason is those decisions are machine learning based best. You know, what does this look like? This sheet of paper. Those are quite expensive to build and they

require a lot of monitoring to get them to work. They work fabulously well. Right. And unlike large language models, they can do things where they can tell you, yeah, this is right and this is wrong. Right. Rather than this is kind of a generally what we know. And so because of that, a lot of the low skill things are not going to go away. And one of the reasons, additional reason, if you eliminate a minimum wage job, you haven't saved much money. Right. Sure. Yeah. But if you can augment by 10%, the work of an expensive professional, that's really valuable.

So I would guess. And one other thing, which I do think we are seeing is that these tools are getting pretty good at automating chatbot functions. Some of them, some of the results in healthcare indicating that people think the chatbots are more empathetic than the humans are. And that sounds unbelievable, but you know, the humans who are doing call center work are stressed. They're low wage jobs. They're often in foreign countries. But even there, apparently the evidence is that the hiring of people to do call center work in the Philippines and other countries where

it's going, it's actually gone up. So go figure, it's not an obvious thing that this is going to wipe out all these jobs. And I'll finish on this, Peter, because using the example that you gave of the medical industry with chatbots, they may be more empathetic, but there is a certain level that a chatbot can probably take you to without needing the input of a human being, whether it be a nurse, a doctor, whoever it might be, to be able to make sure that you get the right information, the right medication, et cetera. So you have to have that human component there in that industry and probably a lot

of others. Yeah. So actually we have been looking exactly at that in healthcare, actually doing it with my son, who's now a doctor. So looking at that, and I think, you know, for sure, you're right. One of the things you see is every one of these chatbot tools has a handoff to a human, right? And some of his patients just say, give me a human. And some of it is, it does just save time, which is a good thing. But in the most useful cases, they're doing things that were not done before. So give you an example, follow-up patients with calls. You know, if you've been to the doctor

recently, if you've had any procedure, you often get these little emails or texts, just ask you if you would respond to a five question, you know. Well, for a lot of this stuff, you need much more open-ended conversations. Nurses could do that. Nurses are pretty expensive and largely work nine to five. You see good chatbots can do it pretty well. And in some places, they can get you to monitor your health. Like, please take your blood pressure now. I want to see what it is and hook that up to your, to the AI system at the hospital so they can check. And that seems to work pretty well.

And it wasn't done before, right? So the most useful tasks to take over are things which we just weren't doing enough of, partly because it was expensive to do. Peter, great to talk to you as always. Thanks very much for your time today. Thank you, Dan. Peter Capelli, management professor here at the Wharton School.