
Loading summary
Podcast Host (Intro/Outro)
How do we experiment with AI in ways that are productive but also safe? Today's guest explains how he spurred new development projects with AI and recounts how various leaders he's spoken with think about the technology.
Andrew Palmer
I'm Andrew Palmer from the Economist and you're listening to me, Myself and AI.
Sam Ransbotham
Welcome to Me, Myself and AI, a podcast from MIT's Sloan Management Review. It's exploring the future of artificial intelligence. I'm Sam Ransbotham, professor of analytics at Boston College. I've been researching Data analytics and AI at MIT SMR since 2014 with research articles, annual industry reports, case studies, and now 13 seasons of podcast episodes. In each episode, corporate leaders, cutting edge researchers and AI policymakers join us to break down what separates AI hype from AI success. Today we're joined by Andrew Palmer. He's a senior editor at the Economist, where he's the author of the Bartleby column and the host of the Boss Class podcast. His current podcast season explores how the use of generative AI is changing management and jobs like ours. So, Andrew, welcome.
Andrew Palmer
Hi Sam, nice to be here.
Sam Ransbotham
So some of our listeners may not be familiar with the Economist or the Boss Class podcast. Can you give us a quick intro?
Andrew Palmer
The Economist, for almost all of its history, has been a weekly news magazine. Now, of course, we're available in lots of different formats. We're published out of London, but we're global in our scope and we cover economics, business, politics, science, technology and much more. And the Boss Class Podcast is a serial narrative series podcast that I host on management and the workplace. We've had three series to date and as you said, the last one was specifically devoted to this thorny topic of generative AI in the workplace.
Sam Ransbotham
It is thorny and I think you do a good job of exploring some of that thorniness. One of your colleagues, Ludwig Sigle, said on an episode, and I kind of pulled this out, the Economist embraces change. We think technology is good and should be used. I always find that pro innovation bias a little bit interesting because I have a background in computer security where we may have a little different bias about whether technology should be used. But in this case I think I agree with the Economist. How would you describe the journal's overall philosophy towards AI?
Andrew Palmer
I would say open minded experimentation is probably the way to think about it. So we have not rushed headlong into it. We have a variety of internal projects to see how we can use it in our journalistic processes. So for example, we fact check everything that we do. So there's a research team there which has to pull through a ton of stuff. Is it possible to make their lives easier while still having humans do the critical work of checking? Similarly, journalists have to conform to a style, guide, a particular way of working. So can we make it easier for them to check that their copy is doing what it should before it gets to editors, who then are the humans in the loop? So there's a lot of internal stuff. And then we have experimented with things like AI generated transcripts of podcasts that are available to people on our site. And we have more secretive, if I told you, I'd have to kill you kind of stuff around what we might be doing in two to three years. So there's a whole panoply of things that we're doing, but we're always very, very clear that we have a particular brand associated with high quality human intensive processes. And there's a lot riding on us getting this right. So we move fairly cautiously as well.
Sam Ransbotham
One of the ideas, I think, that came through a few episodes is this idea of a jagged frontier. That artificial intelligence has really amazed you in some areas, but also been unexpectedly disastrous in others. How does that affect the way you think about experimenting?
Andrew Palmer
I mean, I think it probably comes back to that overarching mindset of being cautious, so that you don't just thoughtlessly embrace the technology, let alone if it's public facing. So everything goes through a. An experimentation phase. And one of the things that's become apparent, and you'd see this in every kind of organization, I think, who's grappling with this, is that you need to have really experienced people in the loop. So for us, that editors who've been in the newsroom for a very long time, working out what counts as quality, providing feedback on the experiments that we run, so that over time it gets better and better and better, and asserting a pretty high bar for what counts as good enough. That's the way in which a mindset gets translated into actual processes for evaluating and checking. And it's a new way of working for us. Most of our history, the journalists have kind of controlled absolutely everything. And now it has to be much more collaborative, especially with this technology.
Sam Ransbotham
So what surprised you? What's been the biggest positive surprise and maybe the biggest disappointment?
Andrew Palmer
Well, I'll take an example of my own sort of playing around from the latest season as a kind of just a bit of a roller coaster. So almost every journalist on earth now is doing vibe coding, so I'm no different there. But at the time, it seemed like we were really breaking New ground. And so that's all recorded. But I'm a non coder, absolutely no idea what I'm doing. So you're the expert in this conversation for sure. But we have a styled guide which I mentioned to you, which is basically our bible on how you should write all sort of grammar, hatred of Americanisms, you'll be shocked to know all sorts of rules like that, which you can thumb through a big PDF to get to, you can leave it to your editor, but ideally you would have something that could basically just check your copy against this. Ludwig Siegele, who you mentioned earlier, had been waiting for a year to get developer time to build this. It was relatively simple, but it's just busy organization, lots of people with different things to do. This was low down the queue. And so the magic of this was I went away and in 75 minutes had built an extension which did check copy against the style guide. Now when I say I built, I was more like a sort of puppet. I had no idea really what I was doing. But by using Claude, this thing was generated. So that was kind of amazing to me. I felt like I'd achieved something which was totally beyond my purview. I just would have been utterly unable to do. And it clearly bypassed our kind of internal bureaucracies. Then the kind of disappointment that you mentioned is that actually we're not going to be able to just push this out magically to people the next day. So there's an awful lot of governance to think around on this. The behind the scenes stuff around what architecture, what software, what our data processes are all had to be thought through. In practice, what I had built wouldn't have worked scalably. So anyway, at that point a bunch of people who really knew what they were doing took it over and it did result in something fast. I definitely accelerated the process, gave them ideas to work with, but it was also clear that I wasn't going to be able to magically bypass all organizational processes and change things. I don't know if that's disappointing or not actually, but there are other examples where, okay, this thing is just giving me kind of nonsense after hallucinating all the stuff that you talk about in the show week in, week out. But I think that's probably the best example of moving from this sense of kind of euphoria, like a whole new world has opened up to being brought back down to earth.
Sam Ransbotham
Welcome back to another branded interview segment. I'm here with Vijoy Pandey, senior Vice President and General Manager at Outshift by Cisco People across AI and quantum computing debate about where progress comes from, from bigger systems or smaller systems. How does ALEC Shift approach that question? And what problems anchor your current work?
Vijoy Pandey
First of all, I'm excited to be back, Sam. The way I would describe the problem statement here is the biggest breakthroughs in computing have always come from connecting things together. So if you think about the history of computing, single servers became bigger, more powerful. But in addition, we also built the cloud software, went through the same progression. We built virtual machines and then went to microservices and serverless every time. The industry starts with scaling vertically because it's easier to do and start with, and then we turn around and we build distributed infrastructure. So right now we are hitting that same paradigm shift in two places simultaneously. The first one is in AI. Every major AI lab is pouring energy into vertically scaling models. More compute, more data, more parameters, building bigger, smarter individual agents. And that needs to continue. And it will continue. But there's a whole class of problems that no single agent, no matter how brilliant, will solve alone. When you need agents and humans communicating with each other. Not just communicating, but the reasoning together across organizational boundaries, sharing context, sharing knowledge over time, building on each other's work. Now that is a horizontal scaling problem. That's a path towards distributed super intelligence. And at outreach by Cisco, where I run this incubation engine, we are building the infrastructure for a distributed super intelligence and we are calling that the Internet of cognition. The second place where we are hitting this paradigm shift is in quantum computing. Now we all know fault tolerant quantum computing will require a million plus physical qubits. A million plus. You're not going to get to fit all of that in one machine anytime soon. Not in my lifetime, I don't think. Now quantum networking provides that scale out infrastructure, that distributed infrastructure for connecting quantum processors into this single logically large distributed quantum system. And we're building that full quantum networking stack from the ground up, from quantum entanglement sources to the universal quantum switch, through the software stack, the networking software stack that ties it all together and enables distributed quantum computing. And if this aligns with your viewpoint, head over to outshift.com we publish everything. We publish the architectures, we open source the code, the research, and we are actively looking for design partners who want to build either the Internet of cognition or the quantum network with us. So come and talk to us.
Sam Ransbotham
I think that's a nice. There's a truism in software. If you ask any software engineer, how far along are you is it close to done? Oh, they're 90% done. And that first 90% takes about half the time or a quarter of the time. And that last 10% is really hard. And it always takes a lot of time there. And I think that's sort of what we're seeing maybe with Vive coding, is that it's doing that 90% pretty quickly. It's making the initial screens or whatever. But like you say, there's a lot more to that process. And I think about extrapolating with artificial intelligence. I think we have a tendency to, like, draw those lines linearly or even exponentially. But, you know, diminishing returns may be the more normal shape, especially as you described there with all the other processes.
Andrew Palmer
Can I ask you something, Sapp? Are you seeing Vive coded apps for yourself and do you notice a difference?
Sam Ransbotham
I'm super into coding in general. That's what I do as a hobby and I'm very. I love it. And what I find is almost exactly what you've described, that this ability to prototype something quickly is amazing. You can just throw up something and get a sniff test of, is this worth and then further investment without, like you say, waiting on your bureaucracy to come along and, you know, take a year's backlog. But you can't fool yourself into thinking that that's actually going to be production code. And so, you know, you learn a lot from that. But I don't know about actually dropping anything in there into production because by the time I get the code, I've rewritten every bit of it before it actually goes into something that's production. But he gave me that sniff test pretty quickly.
Andrew Palmer
Okay. Yeah. One of the people we spoke to was Anton Osicka, who's the boss of Lovable, one of these big vibe coding platforms, Swedish based. And he had this nice phrase, demo don't memo, as a kind of way of thinking about. This is really good for prototyping. Don't do PowerPoints, don't write long documents, just build the thing and show what it can do. But he was also very clear that you do not want to be putting stuff directly into production environments, which sounds totally consistent with what you've said.
Sam Ransbotham
Yeah, there's something interesting too, about. And I know you talk about jobs a lot, but one of the things you touched on there was this delay that you had waiting for the style checker to come along through the normal bureaucratic processes and you were able to sort of get a quicker smell on that. My sense is that most organizations have giant backlogs of projects.
Andrew Palmer
Yeah. I'm reminded of a conversation with Hannah Calhoun, who's the head of AI at Indeed, which is you may indeed have had her on your so big jobs marketplace, as everyone knows. And she described the organizational problems that come from suddenly everyone can generate a lot of code, which means a ton of code has to be reviewed by people. So it's a bit like a sort of waterbed effect. Right. I mean, lumps of work are disappearing in one place and they're popping up somewhere else. And I guess the question is, I think you said that you were basically kind of do the sniff test, but then you're writing everything from scratch again, and is there a risk of going halfway down a route with vibe coding and then realizing, actually, we didn't want to start from here, so we're going to need to redo it again? So thinking that through is part of the organizational challenge of this. There's clearly a technical challenge, but actually that's sort of how we design things to avoid really bad outcomes is kind of also a blocker to fast progress.
Sam Ransbotham
We don't have to know that yet, though. I mean, we're all just, like you said, you're experimenting, everyone's experimenting with this. We don't have to know exactly how these are going to fit into organizational processes yet without learning some and all these tools that we were talking about, they're going to improve. I mean, people like Lovable that you mentioned and others, they're going to get better over time. And I think we are sort of in charge of our destiny here in terms of how that plays out and what we choose to invest in and what we don't choose to invest in.
Andrew Palmer
I do agree with that in theory. And then there are a whole load of incentives at work in the system which in practice constrain your freedom to maneuver. So if you have a bunch of people in the C suite who are under a lot of pressure to realize productivity gains, it's quite possible that they'll run. I mean, we've seen it, right? I mean, that was part of the story of the initial years of generative AI is running towards stuff without necessarily thinking about the organizational consequences.
Sam Ransbotham
So talk about those incentives. What incentives do you see out there within organizations that are in tension?
Andrew Palmer
Yeah. So, I mean, lots of boards are putting a lot of pressure on CEOs and their peers to come up with very material returns on investment. That can be cost savings in terms of letting people go, or it can genuinely be a big jump in output. Neither of those things are necessarily great. It May be that you're cutting people prematurely, it may be that you're actually sacrificing quality over quantity. And so I'm trying to think of an example. So Johnson and Johnson was another guest on the show and they had a very conscious sort of let a thousand flowers bloom approach to generative AI in the initial years. And I don't think they had sort of disastrous outcomes, but they also quickly found that there were kind of this problem of bottlenecks emerging in various places.
Sam Ransbotham
That's your waterbed.
Andrew Palmer
Yeah, exactly. It's the waterbed again. Right. So around that very big organization, you had lots and lots of people independently come up with the same kind of workflow to improve. So something like invoicing, making that faster, then all that would do would be to create a whole bunch of invoices that would then land on finance who weren't expecting them. So there was a sort of chaos problem. But more importantly, the lack of prioritization meant that there was a lot of work being done and only 15% of projects were delivering 85% of the value. And the rest of it, maybe there was benefit, people learning, etc. Etc. But in terms of a payoff, not so much. So they have now pivoted to a much more kind of priority driven centralized approach where there's a central AI council. Sounds a bit Star wars, but central AI council, A central data council signing off on stuff. But it feels much more intentional, much more directed. I do think you see surveys with C level executives saying we haven't seen massive productivity gains yet in the next three years. We expect to. And at some point they're either going to have to say, well, none are coming, or they're going to be bound into a kind of like, we're going to make this happen. And yeah, the incentives to sort of show results, they're pretty powerful.
Sam Ransbotham
Well, I think more cynical take is I think we see a lot of job cuts due to AI announcements that I deeply suspect we've got AI as a scapegoat for. You know, you have a choice of saying, all right, I made a really poor decision and over hired, or I made really poor decisions, or it's this AI thing and it's really nice to point to the exogenous thing. I want to come back to something that you mentioned before about. We were talking about coding, but I think it is a bigger issue in general. Maybe we could expand on that. Your waterbed is that you created an ease of creation like with these tools that the ability to generate, I Mean, it's even in the name Generative AI. We don't have an evaluative AI. Right. That's not a huge passive trend out there. Gen AI is the topic not evaluating. And I think we're seeing that certainly in science. Open repositories like Arxiv have been overwhelmed with submissions because the hard part is less about the generating and more about the figuring out what's worth consuming, particularly in our time based attention economy. How are we going to work around that? Offer some hope? Do you have any thoughts about what's going to happen about the world when we are able to generate everything so quickly?
Andrew Palmer
Well, so the example that comes to mind, and the one that I kind of just played around with by myself was in the recruitment space where there is this sort of uncontrolled generation of content on the candidate side and then on the recruiter side you are forced to use AI to cope with this bombardment. And so there's this sort of strange arms race. So in my example of this was I just signed up to an auto apply software provider, put in some sort of details, really scant details. So I did sort of, you know, I could have spent more time on it and sort of went off, pottered around for like an hour, came back and found that I'd applied to 100 jobs. I had idea, no, no idea what they were. And they included being the head of operations for the City of New York, director of the Iran Afghanistan Veterans Association. I mean, you can probably tell from the accent why I'm not the obvious person for this. And so it was a sort of, it seemed sort of ludicrous that you could fill up people's inboxes. So I'd like to apologize to the Iran Afghanistan Veterans association for doing so. But obviously on the other side of that then you need to have this automated response. And so no one is happy with this. It's like a really bad equilibrium that's been generated and you can see that in other places too. So what's the way out of it? So I guess more humans might be one way out of it. You could intentionally insert humans into processes on the recruiter side and kind of see if that works in some way, but that doesn't feel very scalable. You could be transparent about your kind of AI policy so saying it's fine to use AI, but this is how we want you to use it, or to have an AI specific question that proves how would you use your AI fluency to best advantage, whatever it might be. So there's Some evidence, it's anecdotal, that being transparent about AI use can actually reduce the amount of kind of bogus applications. And then I guess the very long term answer, which is a little too nirvana like for me to buy totally, but it is the one give us the dream. Well, this is what people who kind of like in this world say is that eventually the AI is going to be so good that it is going to hunt out candidates. Right. There'll be what they call reverse apply. So you don't as a candidate need to worry about applying to anything. The AI is going to know from your entry on a site like indeed, exactly what you are suited for, understand your preferences and experience and basically they'll come to you. No need for all the kind of slop that's already in the system. Maybe, maybe that will be the case over time. I'd really like to know whether you think that's plausible, but it seems like we're going to spend a long time getting there and in the meantime it's just sort of bad for everyone.
Sam Ransbotham
Yeah, I mean, I think we've maybe make some analogies to like the deep fake and the deep fake detection that's going on every time we improve the deep faking, we improve the deepfake detection and we go back and forth. And so as you generate applications for a job, you'll have the application for job detector. We've seen this with for example, search engine automation, where you put the right keywords on your website and you get higher in the search engines. We work through many dynamics like, like this. But I think actually on your episode you had a recent graduate, Kat Harrison Gaze, who was talking about the experiences of applying for the job market. Can you give some grounding in the example?
Andrew Palmer
Yeah. So Katz was at Oxford. So in a UK context, very, very prestigious university, clearly very smart, should be someone who employers want. There are lots of things feeding into this. We shouldn't blame this only on AI, but her experience is sort of. She described two worries. One is the process of applying and how do I navigate this? And she has a suspicion of AI, she's worried about cognitive dependence. She sort of valued her own ability to think things through. So regarded. It was like, if you touch this thing, it's going to infect me and change. My ability to think was kind of part of it. And then the second worry was just more generally, if you think about a career in decades, where the heck do I put my chips? What is the career that makes sense going forward? And so both of those Seem to me to be totally reasonable worries. It is really hard to navigate the recruitment process right now. I don't think not touching AI is the answer, by the way. So the advice to Kat was use AI, don't shy away from it. That's not the way to think about it. And the advice for employers was here's this very smart person who is determined to think independently. Take a look at her.
Sam Ransbotham
What you're describing, I think is as we move, I think towards machine to machine interaction, formerly back in the old days, people would walk around from office to office and apply for a job and maybe drop off a resume and meet someone in person that became online. What you're sort of describing is a future where your factor talks to my factor, your agent talks to my agent. And much like the ball players who have agents to negotiate on their behalf, you're kind of describing where you'd have recruiters negotiating on the behalf of the company, agents negotiating on behalf of the employer. And there's not necessarily any need for those to be humans. And that can be a mistake. Machine to machine interaction, is that the future we're headed towards?
Andrew Palmer
I mean, I kind of hope not. I guess it depends. What's the point at which humans come? First stage. I can totally understand that it's almost essential right at this point because it becomes hard to see how it can scale. But as long as there are humans in workforces. And by the way, this is a process which is done exceptionally badly by humans right now. It's a really difficult process to get right, but a lot of it can be improved by humans taking the time to test whether someone is a good cultural fit by being honest about like this job is good for these reasons but bad for these. Right? I mean those kinds of things require. They don't have to be done by humans, but they generally work better if you have humans talking things through. So humans have to be in the loop at some point. I think as long as we are working with other people. Cultural fit, the values of an organization, all of those things are totally essential to a good hire. So you could imagine a kind of machine led skills based process testing whether someone can do the job, but actually what motivates them, why are they joining? Would they be a good fit? Do I want to work with this person? All of those kinds of things. I don't feel like a kind of an agent is the right way to answer those questions.
Sam Ransbotham
Maybe it's just because that's what I'm comfortable with in terms of people that I Work with wanting to. But as you say, we do a poor job of that historically. We make biased decisions. We make decisions based off of attributes we should not make those decisions on. And I find it somewhat appealing that perhaps the increased automation could help us at least see that or, or at least sort of raise potential candidates that our biases may have kept us from.
Andrew Palmer
That's true. Danny Kahneman obviously was very pro the algorithm being in every process and would bet that that was way better than any human. I think probably it's a combination of both, right? I mean, you have a rules based algorithmic way of stripping out bias in the way that you ask questions, whatever it is still fundamentally there's something important about getting on with someone that is quite an important part of a hire. Can I ask on arms races? Sorry to turn the tables, but you said you're in cybersecurity, right?
Sam Ransbotham
Yeah, that was my dissertation research back in the day. Yeah.
Andrew Palmer
Okay, so that sort of arms race problem, right, you've got an AI sniffing out weaknesses, an AI trying to patch them. What's the end point with that? And what's the role of the human in that world?
Sam Ransbotham
I'd love to know the answer to what that's going to be, but certainly it's big. And we've seen so much automation on things that used to be human penetration testings and these sorts of things. That's all become largely automated. But there's something fundamentally clever about people that figure out ways of both protecting and designing incentives that I think still is winning out in many ways. Now what happens is as we get the incentives clear and we get a structured set of rules, that the machines sort of take it to the hyper refined level. But then someone will come up with a different approach or a different idea. And unfortunately, you know, in security, the common problem is that the humans are the weakest link. We could attack your password all day long, but it's going to be much easier just to go in your office and take a look at what you've got written on the sticky note on your desk or trick you into revealing it. So I think we're seeing that the mortals may be the weaker link here. We need machines because only machines can react that quickly. But I'll pull back to one of your episodes on the oura ring. One way of attacking a company would be a refund attack. I mean, that's not what we think of as a traditional cyber attack. But a refund attack might be complaining about something wrong with a product in order to get a Refund. And you had the example, the Oura ring that I think an AI, what do you describe it? The AI agent diagnosed the issue, checked the policy, and then ordered a replacement.
Andrew Palmer
Yeah, it was exactly that. We were talking to Mike Krieger, who's one of the people at Anthropic in the new products division, co founder of Instagram, and he was sort of talking about what was on their roadmap and obviously, you know, inevitably started talking about agentic stuff. And he recounted like his best sort of customer service experience at that point was Oura ring seemed to have a battery problem, interacted only with an agent, and this thing basically sort of decided for itself you are entitled to a new ring, asked for his address and off it was packaged. And he regarded that as the single greatest customer experience he'd ever had. To your point though, of what's to stop the AI just giving you batteries for life. And that's in the guidelines. Right. It's sort of like if there's a certain threshold number of or percentage rate of refunds that are given out by an agent, then you stop. Right. And you're kicking it to a human or another AI model. So there's a lot in the governance there. And on that I had a conversation with the Brett Taylor, who's the chairman of OpenAI, but the founder of Sierra, which is another customer service agentic startup. And again, metrics were super important in kind of getting this right. So initially they were thinking, okay, so any call that doesn't require the agent to hand off to a human is a successful call. And then they realized, okay, well this is totally gameable. Right. So we'll just never hand off to a human. 100% success, everyone's happy. And obviously that's not right. So now they have a combined metric of proportion of handoff calls, but also net promoter scores from customers who've experience interacting with the agent. And so that blended metric seems to work and that sort of feels totally obvious, but it is kind of the bread and butter of implementation of good management. Like, what's the problem you're trying to solve? How do you measure success? It's like totally fundamental to the this being got right. And that's a simple example of how they've got to a decent metric.
Sam Ransbotham
Yeah, that measurement's a big thing. And I think one thing that's happening is that these tools are helping us measure things differently so we can measure things that we never could measure before. We're collecting a lot more data and that's great. But it is pointing out that many of the existing measures may be, I think your phrase was gamable, that once people figure out what that metric is, then they will do something. Let me switch to you. I mean, our show is me, myself and AI. How did you get interested in this? Tell us a bit about your background.
Andrew Palmer
So I've been with the magazine since 2007. Multiple writing and editing jobs covering a variety of things from Latin America to finance. And we have a data journalism team, Britain job, et cetera, et cetera. Most recently I've been writing on management and work and most of the what we do here at the Economist is kind of looking from the outside in. It's like big impersonal forces at work. Macroeconomic, geopolitical, technological. I'm the person kind of inside the workplace looking out and sort of looking at humans as a byproduct of that, like how we all interact. So I've been doing that for a while and very obviously AI is affecting the workplace and affecting us as employees and affecting managers and raising all sorts of questions. So it was a very natural thing for me to start to write about it. And I think it's just one of those super interesting intersections of you've got this incredibly scientific cold sort of machine technology and then you have this unbelievably messy soup of emotions which is what a human is. And putting them together is like just really interesting to observe. So a little sneak preview, but it won't be a sneak preview by the time this goes out. I'm writing this week on the one thing that everyone can agree on is that it's great that AI is going to get rid of grunt work or drudge work. I'm not totally sure about that because for humans drudgery in the right dose is really good. It's really good. There's a bit of agency because you can get stuff done, you can, you can kind of relax a little bit because you can't be tote on the entire time. There's some evidence that mind wandering is really good for creativity. So there's a problem with the idea of like we're all going to be maxing out on higher order tasks the entire time. We're just not built for that. So that's the kind of territory that I'm in.
Sam Ransbotham
I think that's pretty fascinating because it is true that only by sort of pausing and thinking and reflecting and it takes a bit of not constant being on to have those sorts of ideas. And that's certainly counter to the, oh, you'll get rid of the drudgery now. Same hand. There's certainly lots of our jobs that are true drudgery. And so I think the trick as we try to figure all this out is how much drudgery is the Goldilocks amount.
Andrew Palmer
Yeah. And again, you get back to incentives. Right. A manager of a certain mindset might think, okay, puttering around is not something I want anyone on my payroll to be doing ever. Let's go out there and do cognitively intense stuff all the time. It'll be amazing. So there's also a kind of mindset shift there. We couldn't get rid of drudgery before. It was just part of life. What if it's an option to get rid of it all? How would you think about that? I spent some time at an air traffic control center once where their job is to think about what's the optimal performance environment that you do not want these people to be below their A game. And so there's a really interesting balance they're trying to strike there between you don't want to overload, so limited session timings and mandatory breaks, but you also don't want to understimulate because too much boredom, basically your attention starts to veer off in ways which are not. Not great if you're in charge of air traffic. So they have thought about it. It's sort of human factors discipline. And weirdly, it's probably coming to every office just in a much less high stakes way.
Sam Ransbotham
One of the things that you mentioned there was that before we didn't have the ability to automate these things and so we didn't have to make any choices. Now that we have these abilities, we have to make some hard decisions. What kind of skills help people make those choices? These decisions that, you know, I'm in university, help me out here. What should we be helping students learn? What kinds of skills help them make those sorts of decisions?
Andrew Palmer
I actually think, oddly enough, I use the term human factors, which is its own discipline. Right. But it's always been quite narrowly defined, like sort of how do you get the best out of an air fighter pilot or whatever it is. But I do think there's something about understanding human performance that is really important in understanding how the sort of complementarity of humans and machines works. There's something around management discipline itself. How do you design a good process? How do you avoid bottlenecks? Regulating the flow of work in a way that you don't have? Everything just bunching up somewhere else in the system requires you to think at an organizational level and to think about systems and processes. And so basically systems thinking, all management training, all of that is super useful, I think, in thinking this through. And a bit of introspection, right, which is maybe the best thing about this technology, which is that once you start to use it, it forces you to be introspective about what am I good at, where am I likely to have a sustainable advantage, what do I like to do, what do I not do? I really like the idea of doing 100% higher order stuff. Is that credible for me? So I don't know how you train self awareness, but in the process of trying the thing out, you start to inevitably, I think, have those kind of internal conversations.
Sam Ransbotham
And they're useful even to answer those questions. We're going to need a lot more information about people and about how they work. And that's a little bit back to your measurement thing. And that in many ways comes in conflict with my desires to keep my own personal information guarded. At the same time, these systems could probably help me understand a little bit about my own self and how I work and what situations I work best. I think overall that one of the things I think you're bringing out sort of in many different cases and many different examples, is this need for nuance, this need for not always going too far, not always doing too little experiment some, but don't over rely. And I think that's really the overall theme that I got from your podcast, which is this idea of bringing some nuance to all these decisions. And I think that nuance doesn't always play well in our current do these 10 things to make you a better AI person. I think a theme from what I've pulled from your work is clarity and nuance. And I think that's really hard.
Andrew Palmer
It is really hard. I mean, another way of framing it, but it's like the worst way to market the podcast is be boring. So master the essentials, the basics. What is the thing that you're aiming for? AI adoption is not a metric in its own right. So what's the problem you're trying to solve? Think about workflows, end to end. Rather than a single thing. Work might just be being redistributed. All of those kind of questions are. They're just common sense. But as usual, that takes you quite a long way.
Sam Ransbotham
Yeah. Andrew, thanks for taking the time to talk with us. Listeners want to hear your podcast or hear an experiment where you've created a bot to create your own voice in a podcast. I thought that was a fun example. Season 3 of Boss Class is going on right now with the Economist. Thanks for taking the time to talk with us.
Andrew Palmer
Thanks Sam. It's been fun.
Sam Ransbotham
Thanks for listening today. On our next episode, we'll shift gears to health care and speak with Carla Goular, Perl, Chief Medical Officer at Philips, about how AI is enhancing, not limiting, the human element in the medical space. Please join us.
Podcast Host (Intro/Outro)
Thanks for listening to me, myself and AI. Our show is able to continue in large part due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple podcast review or a rating on Spotify and share our show with others you think might find it interesting and helpful.
Podcast: Me, Myself, and AI
Host: MIT Sloan Management Review
Episode Date: May 19, 2026
Guest: Andrew Palmer (Senior Editor, The Economist; Host, Boss Class Podcast)
This episode explores the real-world complexities, opportunities, and pitfalls of adopting generative AI in the workplace. Sam Ransbotham welcomes Andrew Palmer, Senior Editor at The Economist and host of the Boss Class podcast, to discuss what distinguishes genuine AI progress from hype, why the future demands nuanced thinking, and how careful experimentation blends with organizational pressures for rapid productivity gains. Drawing on experiences at The Economist, insights from other leaders and companies, and policy perspectives, the discussion delves into AI’s impact on workflows, management, and job design.
Philosophy of Experimentation & Caution
"We have a particular brand associated with high quality human intensive processes. And there's a lot riding on us getting this right. So we move fairly cautiously as well."
(Andrew Palmer, 03:17)
Collaborative Quality Control
"You need to have really experienced people in the loop... working out what counts as quality, providing feedback on the experiments that we run, so that over time it gets better and better and better, and asserting a pretty high bar for what counts as good enough."
(Andrew Palmer, 04:26)
Successes and Disappointments
“I went away and in 75 minutes had built an extension which did check copy against the style guide...But in practice, what I had built wouldn't have worked scalably... It did result in something fast... but it was also clear that I wasn't going to magically bypass all organizational processes.”
(Andrew Palmer, 07:05)
The Limits of Rapid Prototyping
“This ability to prototype something quickly is amazing...But you can't fool yourself into thinking that that's actually going to be production code.”
(Sam Ransbotham, 12:33)
False Expectations of Productivity
“Lots of boards are putting a lot of pressure on CEOs...to come up with very material returns on investment. That can be cost savings ... or it can... be a big jump in output. Neither of those things are necessarily great. It may be that you're cutting people prematurely, it may be that you're actually sacrificing quality over quantity.”
(Andrew Palmer, 16:17)
Letting Many Flowers Bloom vs. Central Management
“There was a lot of work being done and only 15% of projects were delivering 85% of the value.”
(Andrew Palmer, 17:47)
Generative vs. Evaluative AI, and the Attention Economy
“Gen AI is the topic not evaluating. And I think we're seeing that certainly in science. Open repositories like Arxiv have been overwhelmed with submissions because the hard part is less about the generating and more about the figuring out what's worth consuming.”
(Sam Ransbotham, 18:58)
Arms Race in Job Applications and Filtering
“There is this sort of uncontrolled generation of content on the candidate side and then on the recruiter side you are forced to use AI to cope with this bombardment. And so there's this sort of strange arms race.”
(Andrew Palmer, 20:02)
Potential Solutions to the Overload
Human Touch and Cultural Fit
“What motivates them, why are they joining? Would they be a good fit? Do I want to work with this person? All of those kinds of things. I don't feel like a kind of an agent is the right way to answer those questions.”
(Andrew Palmer, 26:43)
Strengths and Weaknesses of Human vs. Machine Judgment
Governing AI with Blended Metrics
“Initially they were thinking... any call that doesn't require the agent to hand off to a human is a successful call. And then they realized... this is totally gameable...So now they have a combined metric of proportion of handoff calls, but also net promoter scores from customers who've experience interacting with the agent.”
(Andrew Palmer, 30:27)
Measurement as a Double-Edged Sword
Human Factors and Systems Thinking
“There's something about understanding human performance that is really important in understanding how the sort of complementarity of humans and machines works. ...So basically systems thinking, all management training, all of that is super useful, I think, in thinking this through.”
(Andrew Palmer, 36:41)
Value of Drudgery and the Goldilocks Problem
On AI in Journalism:
"Open minded experimentation is probably the way to think about it."
(Andrew Palmer, 02:32)
On Euphoria and Disillusionment:
“That was kind of amazing to me. I felt like I'd achieved something which was totally beyond my purview...Then the kind of disappointment that you mentioned is that actually we're not going to be able to just push this out magically to people the next day.”
(Andrew Palmer, 07:09)
On Organizational Incentives:
“We have to experiment, everyone's experimenting with this. We don't have to know exactly how these are going to fit into organizational processes yet without learning some.”
(Sam Ransbotham, 15:05)
On Applying AI in Recruitment:
“It seemed sort of ludicrous that you could fill up people's inboxes. So I'd like to apologize to the Iran Afghanistan Veterans association for doing so.”
(Andrew Palmer, 21:04)
On Nuance:
“A theme from what I've pulled from your work is clarity and nuance. And I think that's really hard.”
(Sam Ransbotham, 38:37)
Advice for Organizations:
“Master the essentials, the basics. What is the thing that you're aiming for? AI adoption is not a metric in its own right. So what's the problem you're trying to solve? Think about workflows, end to end.”
(Andrew Palmer, 38:58)
| Timestamp | Segment Description | |-----------|-----------------------------------------------------------------------| | 01:25 | Introducing The Economist & Boss Class podcast | | 03:47 | The “jagged frontier” of AI in practice | | 05:17 | Euphoria and disappointment of rapid AI prototyping | | 11:30 | The 90/10 rule in software and rapid AI prototyping | | 13:13 | “Demo don’t memo” and prototyping vs. production | | 16:11 | Organizational pressures and incentives in AI adoption | | 17:47 | Johnson & Johnson: from “let 1000 flowers bloom” to centralized AI | | 18:58 | Content generation vs. evaluation, information overload | | 20:02 | AI arms race in recruitment, Palmer's “auto-apply” experiment | | 24:49 | Human vs. machine roles in hiring and agent-based recruitment | | 29:45 | Customer service automation, agentic AI, and blended success metrics | | 36:41 | Required skills: human factors, systems thinking, introspection | | 38:37 | The importance (and difficulty) of nuance in AI adoption | | 38:58 | “AI adoption is not a metric in its own right” |
(For more, listen to Season 3 of Boss Class with Andrew Palmer or visit MIT Sloan Management Review’s "Me, Myself, and AI.")