Will AI Take Your Job - or Make You Better at It? | Lindsey Raymond, Microsoft Research

Name: Will AI Take Your Job - or Make You Better at It? | Lindsey Raymond, Microsoft Research
Uploaded: 2025-06-18
Duration: 36 min 19 s

The hosts discuss how generative AI is changing the workplace, starting from the World Economic Forum's Future of Jobs Report 2025. They are joined by Lindsey Raymond, an economist at Microsoft Research who co-authored a study with Erik Brynjolfsson and Danielle Li on how a generative AI assistant affected productivity in customer support. She talks about which jobs are at risk, why the gains went mostly to less experienced workers, and how jobs may need to be redesigned.

Watch on YouTube

Transcript

Welcome to Agents of Tech, the podcast where we explore the cutting edge innovations shaping our world. Today, we are diving into one of the most pressing questions of our time: how is generative AI reshaping the workplace? According to the World Economic Forum's Future of Jobs Report 2025 the global labor market is undergoing a seismic shift. By 2030, an estimated 170 million new jobs will emerge, fueled by advancements in AI, the green transition, and demographic trends. Now, at the same time, 92 million jobs are expected to disappear, primarily due to automation. This means that while AI is driving job creation, it's also transforming and eliminating traditional roles. To explore this further, today we are joined by economist Lindsey Raymond, co-author of a recent study with Erik Brynjolfsson and Danielle Li on how generative AI impacts worker productivity.

The key here is sort of learning how to effectively work with an AI system. Just like learning to work with any sort of other team member, there are sort of things that work better, things that work less, reasons why you should use it in some tasks and not others, and I think workers who are able to like, effectively pick up that skill, which is definitely a skill, will be best placed to be the ones to benefit from the productivity enhancements these tools bring.

Now I am delighted to be joined in our brand new DC studio by my colleagues, Laila Rizvi and Stephen Horn. Welcome to our new studio. Yes, Agents of Tech finally has a home.

Isn't that amazing Autria?

Yeah, it's very exciting. Very exciting.

So, Stephen, I want to start with you because, you know, the jobs market as we mentioned, is set to be, you know, reformed by AI, whether we like it or not it's happening. As a CEO and an entrepreneur, what shifts are you already noticing in the workforce?

So I think that first point is right. Right. This is happening whether we like it or not. And the pace of change is to some quite bewildering. You know, we started off with, with chatbots and now we've got agents and, you know, before we before we know it we'll have robots and, you know, so the pace of change is quite bewildering and all these things will have an impact on the workplace. There's no getting away from that.

So what changes am I already seeing? Well, I see quite a lot of use of AI. Right. So a lot of emails being written by AI, a lot of summary reports being written by AI, a lot of, you know, stuff, right, being written by AI. The difficulty is sometimes is that piece that's done by AI has to fit into a wider workflow. And that's the non sexy bit if you like. That people aren't writing papers about people aren't getting excited about, but all workflows have different stages.

The other thing that I'm seeing is already I'm seeing reliance on AI, as I call it, the calculator effect. You know, when we brought calculators into the workplace, people stopped doing mental math. And it's the same now with AI because I can write that email for you. You don't bother doing it yourself. And before long why would you bother? You just press a button and get the computer to do it. And and as the computer gets more and more and more complex, in programs get more complex, then our ability to actually understand what's going on diminishes. Right. So, so so that's what I'm seeing at the moment. I'm seeing, you know, bewildering rate of change, lots of possibility, a lot of the things happening already that are just awesome, knowing that this is the route of travel. But there's a lot to figure out yet.

So, Laila you've been reading a lot about this. So what? What do you think?

Yeah, I've been reading a lot about it and kind of getting stuck in the weeds myself with all of these models.

You're one of these people I have been talking about, pressing buttons all the time.

Yeah. So what I'm seeing speaks to what both you and Autria are talking about is just the rate of change, which is exponential. I think each year it's increasing ten times. I mean, in the past three months, we've gone from a desktop kind of application that can write your emails to agents fully autonomously doing things for you. So I'm terrified each month, the rate at which it's accelerating because you can't keep up. And I think it's the first time technology is not changing intergenerationally, but rather intragenerationally. And I think that's also causing a lot of stress to people and people

And I think that also speaks to your point, Stephen, about integration of these AIs into processes. There's no point innovating all these AIs if there isn't anyone to apply them and synergize that gap between businesses and the technological innovation. So that's kind of where I sit or how I'm experiencing it. And the other thing I'm really interested in about AI in the workplace is this debate about who's it going to replace first, the numbers people or the creative people, because large language models suggest that language people, it's the humanities people. But historically it's actually maybe the numbers people because, you know, the AI can play chess better than any other human. And that's the numbers.

So it's interesting that you bring that up because according to that World Economic Forum's Future of Jobs report, the declining roles that we're seeing the most are occupations that include clerical roles. We're talking about accountants, cashiers, ticket clerks, administrative assistants, the secretarial positions. So it's interesting because I feel like you wouldn't at first blush, think of those being the positions that would be eliminated first, but those are the ones predicted to become obsolete almost.

Yeah, but, you know, I'm not sure I totally buy that.

Good.

I'm not sure I totally buy it because we talk about the rates of change, right? And we're talking about what we can see now based on how much. Right. So, what are you going to want to replace? You want to replace the people who cost you the most money.

Yeah. Right.

Right. Absolutely. So you're not going to want to replace your clerical assistants. You're going to want to replace your chief financial officer. So you're going to want to replace, you know, people further up the food chain. And we talk about, you know, numbers as opposed to creating because we're all creatives. Right?

I like to think we're safe. Right.

Yeah. Because we're all creators. And I just don't buy it. I think the pace that we're seeing with AI, I think that all jobs are going to be affected by it, which doesn't mean they're all going to be replaced.

But I want to go back to what you mentioned about the calculator effect. And when we did bring calculators into the workplace and into, you know, our school systems, people kind of forgot how to do basic math and multiply and divide, you know, and basic math skills. As somebody who is constantly hiring as a CEO and you're, you know, always evaluating the potential workforce, do you worry about people relying too much on AI and losing those critical thinking skills and abilities?

Yeah, absolutely. Right. Absolutely. And the more complex it gets, the harder it is anyway to understand what's going on, the easier it is to get the machine to do it. So now when you go on to almost any, you know, website, you have an AI written summary of what it is, and so that's what you go to. That's what the site generates for you. So no one's writing that. And you know, and you read it in a different way. And so very much so I think we're already learning losing those skills.

So when I'm employing people today, they come with skills I didn't have. Right. They code, they computer code. They've got a lot of skills I don't have. They don't have skills that I do have and they don't have those same reasoning skills. They don't have those same you call them creative skills. So. So yeah, I do worry about that. And as the pace goes faster and faster and faster, you know. And then let's talk about what Geoff Hinton said. Right? I mean, you know, that as these get agency, you know, they don't need us at all, right?

To explore this a little bit further, we are joined by economist Lindsey Raymond from Microsoft Research, co-author of a recent study with Erik Brynjolfsson and Danielle Li on how generative AI impacts worker productivity. Their research provides real world evidence of AI adoption and customer service, revealing that AI assistants can actually boost productivity by 15%, especially benefiting less experienced workers. What does this mean, though, for the broader job market? Could AI be a tool for upskilling workers across industries? And how do these findings fit into the larger economic conversation about AI and work? Let's dig in.

So, Lindsey, we want to get started with the question I'm sure you get all the time. The most common question: is AI going to take all of our jobs?

Yeah, I think the answer to that is no.

Good.

But there's a lot of concerns, I think, that are, well placed about how AI might change or take away some jobs. And I think that's definitely a concern because of all of the things that generative AI systems can do. It can achieve, like really great performance on a whole host of standardized exams. It can produce very compelling legal briefs, sort of lots of tasks that are part of the white collar work that is really primarily benefited from historical automation.

I think what really determines whether a job category will benefit or be at risk is whether or not the demand for whatever that produces, is kind of fixed, or there's more underlying. If the amount of legal briefs produced increases, that would be a good thing. So for instance, in the sciences, there's probably great demand for more scientific output. And so as people become more productive, it will not come at the expense of employment. But for a sector like, say, the call center industry, there's probably some fixed demand for the amount of technical support that people need. And so that will mean any productivity increases will probably come with some job loss.

How will AI actually if it doesn't replace us, how will it change the way that we do our jobs and within those jobs, what skills are likely to remain human skills if you like, and what, going to be taken by AI?

Thinking about journalism, what the model of production might shift a little bit towards is having the human be the editor rather than the producer. So in, producing, for instance, writing where the human still plays an important role in understanding out of all the possible versions of the paragraph that are produced, what is the right one to use? And I think that sort of like discretion and skill will become even more important as humans move towards a place where they're working with more and more AI copilot systems.

But I think the second part, that will change is often now when you have these copilot systems, they're sort of like a question and answer, or training tool. So like I ask ChatGPT questions all the time. And that can replace some types of interactions that used to happen with a manager or coworkers. And so I think that will also change collaboration patterns in workplaces as well.

Lindsey, I now want to kind of take a historical lens on it and compare to previous technological revolutions that we've seen. So during the internet boom, companies changed the game by creating software once and then selling it over and over. What we now call build once, sell twice. In the AI era, how do you think AI systems will scale in a similar way, and what impact could this have on scalability for businesses, and particularly how small businesses scale and how people work?

I think it's a great question we don't know a lot about yet. I think one key difference is that, it will probably change how quickly businesses can get to scale, because there are tasks that I used to need to hire people for, which, you know, working with people is always like a little bit of a slow process that I can now outsource to a generative AI system. So, for instance, a small business, say, like designing a business plan or doing accounting work. Previously I might have needed to hire as I started hire someone to help me with both of those tasks. Now, those are things that you can do reasonably well with generative AI.

And so I think it will make for any business that they can scale quite a bit faster than they could previously. We will probably see more of the large companies selling AI systems that service an enormous market with a fairly small number of people. There are so many returns to having large amounts of data, and there's really very low marginal cost of operating the LLM-based AI systems. So you will probably see more of the like winner take all markets.

That brings me quite nicely on to my next question. I mean, talking about scalability, you can do more with less. That's what technology enables. And it means that a small team of people can make a lot of money, I guess. And that kind of brings us on to the discussion about AI being a skill equalizer versus a skill divider. Do you think AI will widen or narrow income inequality?

Yeah, I think it really depends what you're looking at in terms of income equality. So one thing that's very different about this technology is adoption in the developing world is high, really high, even higher than it is in the US. And many of the like, people who work on freelancer platforms who say, you know, do outsource work for US small businesses through Upwork who are based in like India or the Philippines, they're already using generative AI. And I think these tools will sort of enable, probably enable them to do quite well in terms of raising their income, reducing the need to know the same language as the customer you're serving, and so kind of reduce income inequality across the world.

But I think that within the US there's a real possibility that it could widen it. Because many of the jobs where you maybe wanted to hire someone in the US because they spoke the same language as you. Now with these sorts of tools, you can probably outsource it to someone in a much lower wage country who can use these tools, like translate their writing into English. And so I think that is a real question about in the US, what are the skills that people can develop such that they still have a comparative advantage in some of the like, more routine cognitive, administrative work such that they also benefit from these types of tools.

So could you tell us a little bit about your research results? So in the 15% productivity gains, talk us through that.

So the study looked at, the introduction of a generative AI assistant in contact centers. And the assistant basically gives suggestions to tech support workers on how to resolve different problems, how to communicate with the customer, and the productivity metric, which is the number of issues these workers can resolve per hour, increased by 15%. And that was primarily driven by improvements among the less experienced, less skilled workers. So here within the call center, workers, we see the generative AI tool being a skill equalizer.

And one reason we think that is true is because the AI system was essentially trained on historical data about how to solve various types of customer support issues. And the primary worker who is like figuring out those solutions is the high-skill worker. They're kind of already implementing all of the best practices that the AI system learns. So when the AI turns those into suggestions, it's really the less experienced, less good workers who are not already doing that and so have gained from access to those suggestions.

So that is a technical explanation of why we see this skill equalization within this setting. And I think because that is a feature of how most AI systems are trained within some of our contacts, that's why we see a lot of these similar results across, also coding, also other, you know, legal analysis, you know, studies of like entrepreneurs using generative AI systems. So I think that is a common feature within narrowly defined contexts.

So I'm an entrepreneur and I use AI systems. But one of the things that I see quite a lot is very high skilled people using AI systems to do what I would call medium skilled work. So people who may be writing something or putting together a business plan and they might be very high skilled, but they will use the large language model to, you know, to write a good chunk of it. So do you think that you're going to continue to see this, that AI helps those skilled people disproportionately? Or do you think that's going to change?

It sort of depends a bit on the context. So there are places like where you described where the lower skilled worker can effectively understand when to follow the AI suggestion or when not. So for instance, writing a business plan like in general, even if you're not particularly great, you kind of have an idea of what makes sense versus what's not. There are definitely instances where the best workers are the best at understanding when to follow the AI recommendations, or and when not to, and so they benefit more than the lower skill workers. And so I think really the value of that, like expert discretion determines will AI end up being a skill equalizer or kind of widen skill gaps within any sort of task or job?

Lindsey, I want to go back to something that you mentioned earlier when you described, you know, looking at AI and the output and the content that's generated, as you know, the worker becomes more the editor than the producer. So given that, how do you think workers should prepare for a future where AI is increasingly present in the workplace? Do you think that there are specific skills or mindsets that workers need to adopt that would be most valuable when it comes to being able to utilize AI in the workplace?

And so I think part of that is, firms understanding that they actually need to train and understand how to effectively use these tools in their workplace, and also workers understanding that this is probably such a new area that they may need to try out and experiment and sort of do a little bit of research on their own to figure out, how to use these effectively, because this is all very new. And I think that this sort of like set of best practices about how to integrate these systems, is not use AI for everything. It is certainly going to require some skill on the user's part to understand how to effectively write prompts, how to effectively edit the responses, and like where you really need to check the AI output.

Well, we talked about that to some extent at the beginning, didn't we? We talked about the importance as well, of having the AI that we're using integrated into the company's own workflow, which is, to some extent what you're talking about. There with the customer service agents. One of the things Lindsey I was struck by was in your research was when you said that even when you take some of those AI modules out of the way, the customer service workers were still working to that improved productivity rate. I'm wondering, do you see there has been a use of AI in training?

Definitely. I think that was one of the most interesting aspects of the paper, because generally worker training is really hard to do well. It's not something that generally people feel like has been very effective. And then often it relies on people which are sort of costly and difficult to scale. And sometimes you have great teachers and not so great teachers. And so there's just a lot of, differences in, you know, how well the teaching session goes.

And so I think the two initially we thought on one hand, maybe the workers are just blindly following recommendations. And so, like, they're really building less human capital than they would in a world where they didn't have this tool. On the other hand, the human managers in the call centers are supposed to review transcripts with each worker for about an hour every week, and they're supposed to say, listen, like in this situation, the right answer was to do this, or you should have done X. But of course, doing one hour a week of feedback is much less effective than having real time suggestions. And then often, humans will sort of lean on very general statements like, you should try to just resolve your calls faster and like, that's not very helpful. You need like specific, actionable, personalized, "in this specific case you should have done X". And that's what the AI does.

And so on balance, it seems like the benefit of access to this sort of real time suggestions, which we know from education research, is very effective in, learning outweighed any sort of like reliance on a machine. And I think that is like very interesting because training or teaching anything has been really, really hard to scale or do effectively. And so that is a very exciting finding where you can imagine building these sort of tools. So in the paper, that tool was not built to actually train workers. And you can imagine if you did that as a specific goal, you could have much larger effects. And I think that's a very exciting direction, where introductory tools could really benefit workers.

And, Lindsey, we've been talking, I guess, a lot about improvement of skills training new workers and low skill workers. But I think, I want to talk more about the top brass, that kind of top performing workers in the workplace and the impact it has on them, because I've actually read quite a lot in various domains. I read a paper amongst material scientists where there was a study that it was very helpful the AI tool for the low performing material scientists, but it took away the joy of work for the top performing ones. And, one surprising finding in your study was that the AI actually lowered the conversation quality of top performing workers. So why do you think that happens? And how should companies manage AI adoption to ensure that the top performing workers maintain their top performance and aren't demotivated either?

There are different characteristics of a job, so there are many jobs which are similar to call centers, where people are just massively overburdened. So for instance, like a nurse in a very busy emergency room or like most doctors. And so in that sort of setting, having a tool that reduces their cognitive overload is probably going to improve job satisfaction. While in a case like material science, where you have researchers who go into that profession because they love doing science, probably not. You know, they have a very cushy work life compared to a lot of other workers. The tool might take away aspects of the job that they find, give them the most meaning. But they're not ones who are like, wow, I end every day and I'm so, so burned out and so stressed because I'm so overloaded, I have no time to myself. I think depending on which of the two worlds you are in, really probably depends on whether the introduction of the tool will make you happier or not.

In terms of demotivating the best workers, because you're in an environment where they're already so capacity constrained in terms of the time they can spend on each call. What we think is happening is instead of the call center worker taking additional time to craft the best response, which is probably still better than the best machine response, they will take the slightly less good machine response, right? Because they just have so much. They have this queue of people who are waiting. They have to get through as many as possible. It's like a very stressful environment. And so that is a setting where unless you sort of change people's incentives, you can lead to, people just taking the kind of quick and easy AI solution instead of crafting the best. So one way around that would be to change the, worker's compensation structure to maybe give them an extra bonus for, quote, "call quality" or times when they didn't take the AI recommendation, they crafted their own to sort of account for that effect.

I have a follow up question to Laila's, one of the things that strikes me from what you're saying and from my own experience, is that a lot of this comes down to actually the structure of a job. Where you know what we're talking about, are a lot of functions that are within jobs, but we're not talking about jobs. I mean, there are some jobs maybe within customer service where the tasks are fairly regulated in that way. But for a lot of us and maybe the material science person that we're talking about, they will have different functions that are within a wrapper that we call a job. And I think part of this is the way that jobs are designed, and therefore for these functions is a bit like when you go back to 1906, when cars were made by a single person who would, you know, put everything together, and then by the 1920s, then we've got a production line where the entire way of a job has changed. So I wonder if you think enough thought is being given into that, which is how jobs are going to need to change and how to actually do that.

Yeah, I think this is something that we are all figuring out in real time. So, for instance, in the call center industry, what has changed is now there tends to be more of a front line AI solution, and the humans will only deal with the things that kind of get past that autonomous AI chatbot, essentially. So they've kind of moved towards handling more of the exception cases or moved to more complicated tasks like upselling or sales roles. And even sometimes companies will now have agents who are just assigned to help train the models or to monitor the output, help it understand how to resolve new questions.

But I think systematically, we'll sort of see this, process of job redesign happening in real time. As firms sort of realize where the benefits of having the human are and what are the sort of new tasks that humans now need to do, because we have these sort of tools like before we were using these AI agents, there is really no need to have a call center worker training model. But now that is kind of a new task that has emerged out of this, and I think this really, this sort of job redesign, because everything is moving so quickly and is so new there's not really like a thoughtful or set of principles that might guide, like when you need to think about job redesign. And I think companies are probably sort of like scrambling a bit. And so this is, I think, a really important piece for future study to understand what's happening.

You're making me feel better now because I'm scrambling a bit and I'm not scrambling on my own.

Lindsey, before we let you go, I want to get to one more point that your study found. And that was lower attrition rates and higher customer satisfaction when utilizing AI assistants. So my question is, do you think that the implementation of AI could actually make work more enjoyable and fulfilling for employees or do you think that there is a concern that work may become a little bit more robotic, and people might be working a little bit more in silos and not having that collaboration, and therefore that level of enjoyment.

This is a great question because I'm doing new work on this question exactly. Basically, the concern here is that, on one hand, maybe these sort of tools will automate some of the parts of jobs that give people the most meaning or satisfaction, and they may sort of substitute for some human interactions, like I'm interacting less with my manager or less with my colleagues, and I'm more like interacting with this AI assistant. I think it really depends on the type of job, and some of these job design questions. So in a world where you are like a really overburdened doctor in the ER and these sort of systems come in and help you summarize patient records and diagnose patients more quickly, I think that will probably have a more positive effect than, say, you know, the material science case. And it really just depends on like, how overburdened are you and how salient and how important are the tasks that are being automated.

And then the question of will this substitute for human interactions? You can imagine a world in which, carefully redesigned set of interactions where you have this AI assistant and it helps you answer maybe some of the more routine questions, like all the stupid questions that you're too afraid to ask the manager, that could help, so that you can spend more time actually with, like, higher level questions. You know, how should I approach a problem with these characteristics? There are like two very different worlds we can end up with depending on all of the choices we make now. And so that's why this is like one of the most important questions to try to study.

I'll be interested to find out what your research reveals about that. Final question to you today. What do you think are the biggest unknowns right now when it comes to the impact AI will have on the workforce?

Right now, most of the work has focused on very short term outcomes. So my paper looks at like six productivity outcomes after about six months to a year, which is kind of the longest, but not really that long. And you can imagine that if the AI systems are introduced and it makes people really hate their jobs or makes them feel more siloed, they'll sort of quit, or over the long run, that the sort of gains from these systems will be much lower than anticipated. So I think actually asking, you know, how sustainable are these productivity improvements? And are there other things that are happening that we need to either think about or understand such that, we, like, fully understand the long run implications of adoption of these tools? That's a really important, first question.

And of course, now we're sort of moving towards more autonomous AI systems. So like the example I gave you earlier, now it's possible to build an autonomous chatbot that's reasonably good and that's starting to be rolled out. How will those sorts of tools like fully autonomous ones affect employment and wages? I think it is going to be a really first order question over the next five years.

Fascinating. All right. Well, thank you so much for your time today and for all of your insights and for sharing your research with us. We appreciate it. Thanks very much.

Thanks so much.

Well, I thought that was really interesting and also encouraging. I think one of the big, concerns about AI is everybody thinks it's going to completely replace me. It's going to take my job. But I think what her research showed was, you know, there's there's still a place for the human component, don't you think?

Yeah. I mean, I think what I got from her is it all depends. So I think that it's, you know, it depends on the context. It depends on the job. It depends on whether you're a low performing worker and a high performing worker. It all depends. But I thought the discussion and what Stephen kind of brought to light about design, we need to design our jobs in a new way and incentivize people in a new way to prioritize quality and not compromise on that.

We're using a, you know, a lexicon of the past, really. So we're talking about all these new things that are going to change the way the job functions. But we're talking about jobs as if they haven't changed since the 18th century. You know what I mean? Right. And I think what we are going to see is entirely different jobs.

I thought it was interesting that she brought up, you know, maybe you incentivize somebody monetarily for not using AI. That's a fascinating way to look at it.

Yeah. Well, it actually brings me back to, one of our previous episodes on cybersecurity where, Dr. Damon McCoy was talking about building frameworks where you incentivize bigger firms to prioritize cybersecurity, and it differs in the US and UK. So I think it's the same with AI, really, and we haven't really thought about that macro level design yet because everyone's using it individually, kind of secretly in various ways.

And we also haven't talked about incentives, that incentives are going to play a very large part, both on a human level for individuals, how their incentives are aligned, but also for companies what their incentives are going to be, because, you know, if it's purely a profit motion and motive, then, you know, this design might not work out as well as we want it to.

And it was also encouraging that she mentioned, you know, she sees more on the job training for how certain firms and companies would like for you to implement AI.

Laila put together something which, you know, literally pulled stuff for you so that all you have to do is write stuff in a box and stuff behind it will pull it together. You don't have to worry about how it pulls it together. And I could see that for training straight away. So, you know, now it takes us maybe two weeks to onboard a new sales researcher. And we could do that in probably hours.

That does it for today's episode. If you are enjoying our discussion, please don't forget to like and subscribe as this really helps us to improve the show for you. Thank you so much for joining us here today, and thank you guys for being a part of our inaugural recording here in our brand new DC studio. What an honour.

So exciting. Thank you both for your time today.

Thanks and goodbye. Goodbye. Goodbye.

Agents of Tech is brought to you by WebsEdge. We want to thank our studio engineer, Dan Henriquez, our editor Abbie Harries and producer Cath Sheehan.