The Parlor Room returns for season 3 with a special edition: Hello AI—a series exploring how artificial intelligence is reshaping the business world. In the premiere episode, host and Harvard Business School Online Creative Director Chris Linnane talks with HBS Professor Iavor Bojinov about why AI adoption hinges on trust, how organizations can scale AI effectively, and what leaders must rethink to succeed in an AI-driven world.
Resources
Learn more from Bojinov in the HBS Online courses AI for Leaders and Data Science and AI for Decision Making, which he co-teaches with HBS Professor Karim Lakhani
Watch The Parlor Room Presents: Hello AI on YouTube
Transcript
Editor's Note: The following was prepared by a machine algorithm and may not perfectly reflect the interview's audio file.
Iavor Bojinov: The biggest risk with AI is that most organizations and people are obsessed with protecting job descriptions and not jobs.
Chris Linnane: Companies everywhere are racing to build AI, but what happens after it's built?
Iavor Bojinov: Why on earth if you've got this amazing AI system, why aren't people using it?
Chris Linnane: Why do so many powerful systems fail? Why do some succeed?
Iavor Bojinov: Who's costing you more? The humans or the AI.
Chris Linnane: I think the humans are costing you more on smaller mistakes. Ship it. You're going to ship it. I'm just going to go ahead and ship it. I'll tell you why.
Chris Linnane: Harvard Business School Professor Iav Bojinov studies this inside real companies.
Chris Linnane: What percentage of that has to be right?
Iavor Bojinov: Typically, we set the bar much higher for AI. The advice there in the development stages is really to build something that is designed to scale from day one. Don't assume that, oh, you can fix that after the fact, that you have to do this, all of these things from the beginning.
Chris Linnane: Welcome to The Parlor Room Presents Hello AI.
Chris Linnane: Before we start this episode, I wanted to come outside and take a moment to enjoy this beautiful weather and thank you for listening to this season of The Parlor Room. As you can see, if you're watching on video, I'm out on the campus, right on the banks of the Charles River. Again, thank you for listening or watching, and please consider following wherever you watch or listen to your podcasts. I hope you enjoy this episode.
Chris Linnane: Was there a moment where an AI system technically worked, but you realized that no one was trusting it, no one was using it?
Iavor Bojinov: Yeah. So this is a great question, and I love that we're starting the conversation here because I actually really lived this. So, a little bit about my background, I did a PhD in statistics, and so I was doing all of the stuff that everyone's talking about. I actually studied it, and I was doing it. And when I finished that PhD, I knew how to build AI systems. And so I said, "Okay, now it's time to go out into the real world."
LinkedIn was crazy enough to hire me. So I started working there on their experimentation team, started building their applied research group. And one of the projects that I was leading was to build this AI system for doing automated causal inference.
And that's a really hard problem. And what that basically means is if you imagine LinkedIn has an abundance of data, and what managers really care about is not just all of this data that you have, but they really care about, "Hey, if I make these changes or if I make these decisions, how is that going to affect the metrics that I care about, like revenue, engagement, et cetera?" And so what we noticed was there was this problem that data scientists were not really versed in these causal inference methods.
And so they would go, they would do some naive analysis. And I was in the room, and I distinctly remember this pretty junior data scientist going to a VP and saying, "If you make this change, you'll get a 250 percent increase in profit." And the VP is just sitting there looking at him like, "You're out of your mind." And so that was completely eroding trust in the whole data science organization. And so we set out to build a whole platform to actually do this type of analysis. And this is a very complicated type of analysis. People do PhDs on it. They spend many years on it, but there's a lot of methods have been developed.
And so we build this amazing... I mean, I was leading it, so amazing.
Chris Linnane: Amazing. I had heard about it.
Iavor Bojinov: They still use it. And I was there six, seven years ago. So it's still one of the key tools internally used in LinkedIn. We wrote a lot about it, actually. But we built this platform that massively automated the whole process. And so it reduced the amount of time it takes to do this type of analysis from, say, maybe two, three weeks, or even two months to a day. Okay?
Chris Linnane: Wow.
Iavor Bojinov: And not only that, it used the latest state-of-the-art methods. It had access to data that individuals wouldn't have because of privacy reasons. So this AI system could access this data, they could use it. And so it was more accurate. It was an order of magnitude faster. And we had done our pilot. It was really exciting. People were like, "Yeah, this is the best thing ever." The people who had used it, the four or five teams that we worked really closely with, we launched it, and no one used it besides the three or four people that I'd actually trained up on it. And we waited for sort of a month, no one used it. And this was a bit of a shock to me because my background is in statistics. We learned how to build methods. We learned how to build AI systems. No one told us that if you build it, they won't come.
So I was in this mindset that if I build it and if it's good, people are going to come to use it. And so that actually led to a whole stream of research and trying to understand why on earth, if you've got this amazing AI system, this amazing technology that works really, really well, why aren't people using it?
And so here's the thing. What we found is, and this is going to sound so basic, but it really came down to trust.
But trust is a very overloaded word. It means many, many different things. And if you ask people, "What do you mean by trust?" Trust. What exactly does that mean? But what we found is that there are basically three dimensions to trust in this setting when it comes to AI. There's trust in the algorithm itself, the model that's being used. And these are things that you can guess. Is it accurate? Does it give you consistent results? If I ask it twice, is it going to give me the same thing or is it going to give me something completely different? Is it transparent? Is it fair? Is it biased? All of those things around the model, that's one pillar. But it turns out that's actually not enough. There's two other dimensions that are actually really, really important.
And we really lived this on LinkedIn because we knew it worked. We could show it worked really well, but again, that still wasn't enough. The things that were missing were what I call trust in the developers and trust in the process. Trusting the developers is this idea that, do I trust the people that actually build the system? And there's a number of dimensions of that. The first dimension is, well, did that developer understand my needs and build something that would work for me?
And if the answer is no, then why would I use that? They didn't listen to me. They didn't speak to me. They have no idea if it's...me as a user, I have no idea if it's going to work for me. I'm just going to do what I know works. And so at LinkedIn, that was a huge failure because, as I said, we worked very closely with a couple of teams, and they knew it worked for them, but everyone else thought, "Oh, it just works in those special cases. It's not going to work for mine." Mine's really special. "You didn't speak to me. You don't understand my needs. It's not going to work for me. And so you should... I'm not going to even try to use it."
So that's one dimension. The second dimension of this developer that's getting even more important now is, are there any hidden intentions? So, for example, did that developer build the system to just replace me? Because if I don't trust the developer, and I don't think if I believe they just built this just to replace me after six months, of course I'm not going to use it. But that's very different from, does the model work? So that's the second lever. And then the third lever, which is sort of even more zoomed out is trusting the process. And here, this is fundamentally about what happens when it goes wrong, who's responsible? Am I on the hook for this or is there someone else that's going to be responsible? Or if I disagree with it, how do I handle that? What is the process for me disagreeing?
And so that's why I sort of learned that if you build an AI system, people are probably not going to use it. And if they're not going to use it, it's a failure in trust and it's usually one of those three dimensions. In the example I gave you, the LinkedIn example, we actually failed on all three dimensions in the trust in the AI system, yes, we showed the work, but it wasn't at all transparent.
And so these data scientists who are very technical users, they couldn't really see what was going on under the hood. So what did we do? We started running sessions where we exposed the code and explained, so turned that black box into a clear glass house so they could truly understand how this worked. That was one dimension. The second, I mentioned the trust in the developers, we didn't so much struggle with replacement because data scientists do many, many things. This was just one specific task, but this personalization, that was a major failure on my part. And so what I did was I basically had to go and have individual conversations with literally a hundred data scientists and say, "Hey, tell me your problem. Let me show you how this actually works for you. And let me tell you that we really are listening to you and if it doesn't work for you, we're going to implement it so that it does work." And then for the process, it was very ambiguous, which around what happens if you run an analysis through our AI system and then you follow on the recommendations and it's wrong.
Who's at fault? It was unclear. And so what we did was we built a whole certification led that said, once you build the analysis, if you get it certified through our process and we review it and we give you a thumbs up, we're responsible and you're no longer at fault if something goes wrong because we reviewed it and we own that central responsibility. And so once we build the trust in the tool, build the trust in the developers, build the trust in the process, three, four months down the line, this was being used by many, many of the data scientists. And as I said, this is something that's actually still in use today. They've developed it, they've improved it, et cetera. But hopefully this gives you an example of, yeah, this happens all the time, but you can fix it if you really diagnose why you're not getting the adoption.
Chris Linnane: So you had those three other factors that you added in. Once you implemented those, was it like a wave of people using it.
Iavor Bojinov: Massive wave. And then we actually, we weren't quite prepared for it because, as I said, we changed this process so that we were actually reviewing things. And so we became the bottleneck because we had so many people doing these analyses and they would come to us and they would say, "Oh, I have this amazing analysis I needed to review." And so we actually had to... There was at one point, initially there was just this huge spike. So we had a couple of people working almost full-time on just reviewing these analyses because we really just built a tool to build an AI system that people really trusted and they were able to get behind it.
Chris Linnane: So the system's only as good as these factors of trust.
Iavor Bojinov: Absolutely. And this was pre the age of Gen AI, but this is something that we see time and time again every single time you look at any AI system that is not getting adoption and you go out there and you try to diagnose it, I can guarantee you it will be one or more of these trust factors. And I mean, I'll give you one concrete example. So we just finished a case, which is already approved, it's in the works, it will come out online very soon on Microsoft's internal deployment of Copilot to the sales organization.
Chris Linnane: Interesting.
Iavor Bojinov: And the case is all about, it's actually an ABC case. The A case is all about they launched Copilot, they built it, right? They launched it and they got this initial great wave of adoption that then sort of plummeted. And again, it sort of asks the question of why aren't people using it? And if you look at it, you can map it again to there was a lack of trust in the tool. It just wasn't clear if this tool worked very well at the time. This is a couple years ago, like in the early days of Copilot. It wasn't clear if it was actually working very well. It wasn't really clear to the sales organization, what were the intentions of the leadership team? And this right around the time Microsoft was doing some major restructuring, they were doing some layoffs. And so there were people genuinely were afraid that if the AI system got too good, it would start to replace them.
And there was lack of trust in the process, which was, "Hey, if I follow this AI system and I make a mistake and something, who's at fault? Am I a fault? Well, if I'm a fault, I'm not going to use this thing that I don't fully understand or fully trust. I need you guys to take some responsibility at least away from me. Otherwise, why would I do this?" So we just see this coming up in pretty much every adoption story that I've seen.
Chris Linnane: Yeah. And imagine that's three strikes, three trust strikes right there.
Iavor Bojinov: Exactly.
Chris Linnane: Wow. So you touched on responsibility and I want to get back to that soon because I have more about that. But when leaders worry about AI risk, are they more uneasy about the system or about owning the decisions?
Iavor Bojinov: Yeah, I think it's a little bit of both, and it depends at what level. So I think there are sort of two levels here when it comes to AI risk. You have the really high level, the broader AI risk, which is at the organizational level, my sort of security team is in charge of all of my AI systems. And as I start to lean more and more heavily into AI systems, there's a greater risk because I'm starting to automate more and more and more. So that's one piece of it is this overall aggregated risk across all of my AI systems. That's one part of it. And then the second part of it, of course, is of course every single model and every single application and use case of AI comes with its own associated risks.
And so I think leaders are worried if you're a very senior leader, you're of course worried about the aggregated, right? If I'm the CEO, I'm worried about the aggregated. And maybe there may be a couple of very important algorithms that my company heavily relies on and those I might also be worried about, but as a senior leader, that's kind of it. And then as you sort of go deeper into the organization and you look at the owners and sort of the managers of those AI systems, they of course start to worry very operationally about the specific application that they own. And so I think there's a lot of worry on all of those pieces of it.
And so again, it really does depend on the level that you're sitting in. And it's a different type of risk. At the CEO level, it's not just the risk of, is a mistake going to cost me some revenue, but it's also regulatory risk, which is, if I use these set of algorithms, am I going to be in breach of some sort of compliance? We know the regulations are evolving. In the US, there isn't as much as there is in Europe. And so if I'm a multinational and I deploy something, I'm worried that is the European Union with the AI Act or GDPR, is that going to suddenly come and affect me? In the US, I'm worried about the California acts. Texas has their own legislation. So I'm worried about those things because if you look at the fees that you can get, their percentage of global revenue in Europe at least. And so if your LinkedIn at the time was owned by Microsoft, and so if LinkedIn makes a mistake, it's Microsoft global revenue that gets hit.
And so that's kind of that really high level risk. That's, I think, what leaders are very conscious of because it's a trade-off, which is you want to experiment, you want to move quickly because the benefits are potentially tremendous. But if you get hit by some fines, and there's a lot of companies that are being hit by GDPR fines, AI fines, those numbers are very quickly growing. You just don't want to fall victim to that. And then again, I think individual employees and lower down managers who are responsible for specific applications, they worry more about, is this performing as it should? Is it delivering the value that I promised, et cetera.
Chris Linnane: I almost wonder if you had an org chart graphic of decisions and risks in certain industries. I know there are risks in a lot of industries, in certain industries that were pretty safe, I bet you put little exclamation points near all the risk points, nowadays it's probably just loaded with risk the minute you pull AI into your organization compared to the past.
Iavor Bojinov: Absolutely. And we wrote a super interesting case on JP Morgan and sort of the initial... First, it started off with the initial banning of generative AI to sort of the building and opening it up internally to the launch of their proprietary coach system, which is now sort of integrated and deployed to all of these people. And they had to actually change their risk and governance process when it came to generative AI. And one of the key lessons there was that in the past with traditional machine learning, every time you had a problem, you would build a bespoke model to solve it. So in the example I gave with LinkedIn, we built a bespoke solution for doing this very specific type of analysis. And so every single time you build one of those, you would go through a full compliance and a risk assessment. And that's great.
But when it comes to generative AI, usually you have a single model that can be used for many different types of applications. And so what was happening at JP Morgan, and I've seen this in many other organizations, especially ones that are really sort of struggling with gen AI adoption right now, is every single time there's a slightly different use case, it goes for a full review process. So for example, maybe I want to use it to translate my English documents to French. That goes through a review process. Now I want to use this to translate to German. That goes for another review. Okay, I want to do Japanese, completely different review process. And you can imagine the risks are actually the same. And so the big aha moment for JPMorgan there was actually we should approve functional patterns when it comes to generative AI. So like translation is a functional pattern.
Data analysis is a functional pattern. And then there's many different applications to that, right? I could be doing data analysis here, I could be doing it over there. I don't need to get the whole thing pre-approved.
So that was the big sort of unlock where it went from, I shouldn't just put the exclamation marks next to each application, but I should think about the functional patterns of how people are using it and handle risk at that level and then trust that I build systems that can handle various applications within the same functional pattern.
Chris Linnane: And that gets it to a slightly more manageable scale.
Iavor Bojinov: Way more manageable, especially right now in this world where the cost of building your own custom agent or custom chatbot is almost negligible and anyone can do it. You can't go and review every single one of those applications. You'll do nothing else.
Chris Linnane: Yeah.
Iavor Bojinov: It might be more work than just doing-
Way more work.
Chris Linnane: ... from scratch.
Iavor Bojinov: Exactly.
Chris Linnane: So when an AI system makes a bad call, who feels it first, even if they're not officially responsible for the bad call?
Iavor Bojinov: It's a great question. It's usually the users. And if you think about for most organizations, there's basically two types of AI systems that they deploy. One is external facing, which is their customers. So think about Netflix, their ranking algorithm, right? If the ranking algorithm stops working, who's the first person to see it? The person watching Netflix, and they're the ones that are going to start to change their behavior. And so when it's external facing, you need to have very careful metrics because your end users see this first, and so you need to be able to detect those degradations as quickly as possible. And it can be very costly if it goes wrong. There are many, many examples where companies have basically messed up and it has cost them a lot of money.
Because as you have more and more fully automated systems that are fully relying on AI without... It's just really hard to monitor some of these complicated things like a ranking algorithm. It's really hard to know when that breaks and to build observability layers on, is this really working? Because yeah, it could be giving you rankings, but maybe just someone flipped the sign and it's just showing you the things you don't want to watch. Every measure is correct. It's showing you exactly what it was supposed to show you, but the system doesn't know that someone flipped a sign somewhere and now is showing you the worst possible things for you.
Chris Linnane: What a great idea. The things I don't want. Show me everything I don't want, but it could happen.
Iavor Bojinov: It could happen. And so you would be the first one to feel it. And I think beyond Netflix, whenever it comes to these types of consumer facing AI, it's usually the consumers that feel it first. When it's internal facing, which is the other time, remember I said two buckets, internal, external facing, the risks associated with those are of course quite different. But with internal facing, again, it's the end users. But there, the risk is a little bit less because in the example I gave of LinkedIn, which is if that product stopped working, it's highly skilled data scientists that would say, "Hey, I'm seeing this issue," and then they would sort of reach out to you. So the risk are much lower. Now, of course, there are some, if you're using this to manage your internal supply chains, of course that could be quite risky. If you're someone like Amazon that heavily relies on fully automated supply chains, again, you would sort of start to feel it. And so there's been actually a lot of work on, be a little bit technical for audience for a second.
There's a lot of work on what's now sort of sequential testing or continuous monitoring to basically try to catch these types of degradations as quickly as possible. And so we did some work, for example, with Netflix where we were able, and this is a system we built, we wrote a paper, it was deployed, it's used in Netflix right now, and we were able to detect these types of degradations in the first few hours, as opposed to having to wait a day, two days, or even a couple of weeks before the... Because if it's a really big thing, you see it. But if it's a more subtle thing, it can take quite some time before your actual company data starts to see this. And so you have to build systems that are able to continuously monitor everything that's going on, so that if something goes wrong, you can catch it as quickly as possible.
Chris Linnane: Where does that end? So you have a system that monitors the system, and then you need a system to monitor that system. How do you put a cap on this and feel like we're just going to go for it? Because I know in the little work I've done, and I've done a good amount, but compared to this stuff, little, I feel like I can't put enough safety measures in because it feels like it's just sliding on me at all times.
Iavor Bojinov: Yeah, that's such a good question. And it's a really hard question to answer. And so I'll do the classic HBS thing where I won't really answer your question, but I;ll talk about something else and we'll see how that goes. No, no, but seriously, I think there are many methods out there to try to monitor degradations. And what you have to figure out is the saying, if you can measure it, then you can actually act upon it. You have to think about where are your real risks? And usually that comes in the place of something to do with revenue. If it's starting to affect your revenue for social media companies, it's something like engagement. And so you want to build very stringent guardrails that flag if something is going wrong, but they need to be sort of flexible enough because you might not necessarily need to shut things down.
You just need to sort of alert someone that's to know that they should go and look at this and try to figure out what's happening. And so the way I think about it is build whatever system you want, think about the risks where things can go wrong, and then design metrics that can detect these types of mistakes, and then just build an automated system to monitor that for you. And then you should be good. Now, as we talked a little bit about earlier about the hierarchy, you can then imagine you actually need the system to monitor all of these systems, but that's actually not that hard because if you build a good observation layer, you can actually quite easily just look across all of these and sort of scan them. And that gives you a broad sense of the overall risk that your company is engaged in.
Chris Linnane: So there's some level of redundancy of watching.
Iavor Bojinov: Oh, absolutely.
Chris Linnane: It's not like you can build the perfect model that-
Iavor Bojinov: You can't. This is the reality with AI systems is they're never going to be 100% right, but people are not 100% right. And if you think about how we've organized work over the past few hundred years is we've tried to build redundancy so that when a person makes a mistake, there's usually some other, their manager or someone else looks at that output, especially if it's an important output, to make sure that gets caught. And you just build as many redundancies as you can afford.
Chris Linnane: When AI systems start acting on their own, what's the hardest thing that leaders have to get comfortable with?
Iavor Bojinov: So I think... I'm going to rephrase that a little bit because I think it's not a binary.
Chris Linnane: Okay.
Iavor Bojinov: It's not that these systems act on their own. Yes, no. It's a matter of how much agency they have. And so you can imagine some systems have very little agency or almost none, and then other systems have a lot of agency to do a lot of things on their own. And as you sort of dial up the amount of agency that you give a system or what we are now starting to call AI agents, which are basically AI models that have agency to act. That's the definition that Marco Air and Citi and we teach a course on competing in the age of AI and we talk a lot about AI agents and there's always this debate, what's an AI agent? And our definition is just very simple. An AI system that has agency, right? It's not a binary thing, most systems are agent, and then it's just about how much agency you have.
And so what happens when you have very little agency, the system is basically an assistant. Think of your chatbot. You go out, you ask a question, then you go and respond. And what's happening with these low agency systems is that they're basically... They're starting to augment work. You're not really changing how you get your work done. You're just maybe speaking to it a little bit more than in the past, maybe you would've had one of your colleagues review this email if it's a really important email, now you just have ChatGPT do it. It's not really a fundamental shift in work.
And so for most people, what we found is once you get past the initial hurdle, which I think most people already passed that hurdle over the last year and a half too, you become okay with it, you start using a system. But something really drastic starts to happen as you dial up that agency and you start to get into these systems that are starting to perform tasks that either were previously not at all done by people because it was too expensive or you just didn't have the resources or were actually done by people and now you're actually starting to change the work that someone is doing.
And if this isn't done well, you can imagine that this goes against people's work identities and they can be really against using this. Again, sort of comes back to the trust, like, do we trust the people who build it? Why did they build it? What do they want with it? And by the way, this doesn't have to be with agenetic systems. This can actually happen with traditional ML models. So we wrote this case like two years ago on Pernod Ricard, one of the biggest outcome makers.
And we were looking at their digital transformation, we were looking at the AI deployment, and we looked at a number of very traditional AI models. One of them was basically giving you next product recommendation. So when you go to a shop and you try to sell them something, it would say, "Hey, if they don't want to buy this, give them this other thing." That was one product. And then there was another product which was basically trying to tell marketing leaders how they should split their budget. And the first product was just the recommendation and people didn't have to follow it. They put incentives in place to follow it Because the recommendations were actually very, very good and if people follow them, it actually worked. And so they put in incentives which were, "Hey, if you follow it and you don't meet your quota, you still get your bonus, but if you don't follow it and you don't meet your quota, you don't get your bonus."
Chris Linnane: Interesting. This is a smart move.
Iavor Bojinov: Very smart.
Chris Linnane: Yeah.
Iavor Bojinov: Great, great, great move on their behalf. And so they got lots of adoption. That was great. The other part of it was you could imagine that system had a lot of agency because it was actually starting to determine how much spend you would give. And it was fundamentally changing the role that a person has. Because the previous example, the salesperson would still go there. They would still cultivate a relationship. But with the marketing modeling to say, "Okay, you should put this much budget here, this much budget on this brand, this much budget on this brand." That was usually the job of the CMO of the area. And that was going completely against their whole identity. And so of course they didn't want to adopt it. And so that's what I mean by agency is this continuous thing. As you dial it up, it becomes harder and people struggle with it because it starts to push against the type of work that they were previously doing.
Now, it's not all lost. There is a huge opportunity here because, and I'm sure we'll talk a little bit more about agenetic AI in a moment, but just to sort of jump the gun on that one. With agenetic AI, as you move away from sort of specific task-based automation and you think about how can you architect a whole operating model that's agent first. And what I mean by that is people don't go and use tools, they use agents and those agents use tools and act on their behalf. You're fundamentally rewriting work. And so yes, people are giving up some work that they've previously done, but you're actually fundamentally changing their work. So let me make this a little bit more concrete. We wrote a case, and it's under review, but the case I'm sure will be approved on this amazing company was in France. And this is what I'm going to say is basically open source and public knowledge, so I'm not disclosing anything. But they basically build this whole agenetic operating system, which they've open sourced called Coday. You can look it up, you can get access to it.
There's articles on it. And what the system does is it's basically an orchestrator of many agents and those agents are designed to complete very specific tasks. And each person that works there basically goes to the central agent, tells it to do work, and that agent goes and completes the work on their behalf.
And this is all the way from coding to document creation for release notes, even to reviewing basic terms of contracts, et cetera. And they use it in almost every part of the organization. And what they found there is that there's this major transition happening where people were moving away from being sort of producers of content where I'm the producer of this code to being sort of content managers where I manage the agents that wrote that code, I'm reviewing it, I'm editing it, and I'm sort of pushing it out. And what that has really done for this company is it's allowed their employees to transition from being very reactive like, "Oh, there's this bug, I need you to fix it," to being proactive and saying, "Where is there going to be a bug? Let's see how we can design the code base and catch that stuff even before it happens."
And so I think there's a way of increasing that agency without losing the emotional attachment to work, but you have to do it very, very carefully and you need to think more broadly than just a specific task. You need to really think about how can I... Maybe the analogy here is, how can I protect people's people and not their job descriptions. Right? And when you do that and you have flexible people and a flexible culture that are okay with having a role that's actually evolving and their job is sort of changing, that's where you can actually move away from this emotional fear of, "Oh, this is taking this specific task because actually I'm much bigger than that. I'm much broader and I can evolve my own role."
Chris Linnane: That's super interesting way to look at too, because a lot of times the job description is specifically what you're holding your value to and the organization and the person don't feel very clear about how much that can move. Sounds like that's where we have to head.
Iavor Bojinov: Exactly. And I think this is, if I think about the biggest risk with AI right now and sort of the broader adoption at the societal level is most organization and people are obsessed with protecting job descriptions and not jobs. And if we move away from that and say, "Actually, we're in this world where it's no longer you get a specific job description and you're going to do that for the next 20 years, but actually you have a job that's going to be evolving. And it's your role to figure out how you can leverage technology in not just AI, but more broadly technology to reimagine your role and move to this world of like, what are the jobs that actually need to be done as opposed to, oh, this is how I do this job."
Chris Linnane: Yeah, it's interesting. So you had just mentioned that company, or I'm sure a lot of companies do this, just the end-to-end agenetic aspect of that. What percentage of that has to be right in order for it to be considered good enough? Humans probably in the loop and looking at it, making sure it meets certain standards, but is it like, "Hey, if 50% of this is good, we keep moving," or does it have to be like 90% that has to... What's acceptable for our organization?
Iavor Bojinov: That's an amazing question. Typically, we set the bar much higher for AI. For example, self-driving cars are significantly safer than drivers, especially drivers that are tired, maybe have had a drink. Waymo is way safer than any of the San Francisco drivers. And yet the bar for self-driving cars is an order of magnitude higher than for human held cars. I think there's several reasons for that. One is we never actually look at how good people are. Of course, we do reviews, but if you look at most professions, there isn't a measure of how often did my manager make the right decision.
And so of course, if we don't know, we just say, "Oh, if you ask people to estimate, are you above average, below average, or average?" The vast majority of people say they're above average. They can't all be above average. And so because we all think we're above average, we say, "Oh, okay, this AI system has to be much better than me and so, etc." That's a very superficial answer, and I don't think that's quite right. I think there are multiple dimensions to this. One is, of course, how right it is, but really there what matters is what's the cost of a mistake, but there's other dimensions, for example, speed.
If I have something that's very low risk and a person can get it right 90% of the time, but it takes them two days to do the work and they have another system, they have an AI system that can do it in five minutes, but it's right only 80% of the time. So it's actually worse than a person, but it takes them so much quicker. As an organization, I might say, "You know what? I'm okay with making these mistakes and then having to pay some price for those mistakes because now I just increase my capacity by a thousand percent." And these are the differences we're talking about here. It's not a two or 3% improvement. In some examples, you're getting a thousand, 5,000% improvements in terms of speed over people. Usually it's the same quality, if not higher, but even if there's a lower quality, that's a cost consideration that some CIOs and leaders are actually okay doing that calculus.
Chris Linnane: Yeah. And it seems like the more you experiment with the AI, the better your judgment gets on these things. Because I could see, especially this happened to me and it happened to a lot of my colleagues is the magic trick of doing it the first time is just like, "Oh, that's so cool." Even if it was 30% right and you spent more time fixing the 70%, the magic trick was impressive enough. I feel like that was kind of the early days of this. And once you get used to it, you start realizing that's 30%, that's not good enough. I got to figure out how to get it to 80 every time.
Iavor Bojinov: Yeah. And I think, I mean, this is also the other challenge here as well is when it comes to the broader adoption of these AI systems is that typically when you change work, there is a productivity dip. And we saw it with IT. When IT sort of started getting everywhere and adopted, there was actually an initial productivity dip when people figured out how to do it because they were used to writing things on paper, they knew how those process worked, and suddenly you were getting them to do a different process. And so there's that initial dip, which is exactly what you said, which is initially when you use it may take you longer just fixing it.
But if you keep persevering, then you figure out how to actually use it effectively and then you can get to this much higher productivity state and maybe even quality state of work. But you have to be aware that there is that dip. And then at the organizational level, there's broad questions around like, how do you make sure you have the right incentives to reduce the time it takes to navigate that, but also to make sure that people are willing to take that risk and they have time to actually go become worse for a moment so that they can actually get better. And most companies don't do that.
Chris Linnane: Yeah. It's very important, right? Because without that, no one's going to take the risk. So they got to feel like they're safe taking learning. I mean, riding a bike for the first time, you're not going to get it. It'll fall over, but if someone took your bike away the minute you fell the first time, it wouldn't feel good.
Iavor Bojinov: Yeah, absolutely. And the challenge with gen AI, the general purpose generative AI tools that we have out there is there isn't a guide that tells you, "This is how you should use it." And so every person needs to figure out that for themselves. And this was actually one of the things that came out in this case study on Microsoft was, if you think about how you do typical deployment strategies without AI, with IT, usually I show you how it works and then you go and use it. With Gen AI, I show you how it works. You tinker around a little bit, you have a couple of those situations where you actually had to spend a bit longer to figure this out and you say, "That's not worth it." I don't want to do that. And so there's this whole phase of experimentation where you have to figure out where is this going to work well, where is this not going to work well, and then how can I really use it in the places that it's going to work well? How can I get the most out of it? And then you have to figure out how can I actually communicate that to the whole organization? Because if you find one method, I find another method, how do I actually share those learnings? So organizations have to do a lot of things that they haven't really had to do in the past with other waves of technology. In addition to everything that we already talked about, about building trust and the models, the developers, the process, like how do you actually do the incentives and everything else? How do you do organizational learning? I mean, fundamentally, actually, if you think about this agenetic operating core, fundamentally it's rewriting what organizational learning is. In the past, organizational learning is really about the knowledge of the employees as well as some of the documentation that gets produced. With AI systems, you can actually capture a lot of the tacit knowledge that lives in people because you can monitor how they work. And a lot of that actually, then the knowledge, the knowledge base actually starts to live in this agenetic system. And so the whole system of organizational learning becomes much more rigorous and AI centered, but most companies are not really ready for such a transition yet.
Chris Linnane: Now you'd mentioned identifying when things worked well and when they don't work well, this kind of leads into like, why do so many AI experiments end with a slide deck instead of a real decision?
Iavor Bojinov: Yeah. And so really good question. There's a number of failures around this, and you can sort of think about all of the failures in the whole process of developing an AI system. And there's basically, at a very high level, there's like five stages to any deployment of an AI system. The first stage is selection, so figuring out what you should be doing. The second stage is the actual development. So building it, making sure it works, evaluating it in your little data sets. The first stage is actually experimentation, making sure it works in the real world. And very often, there is a gap between the lab study of, "Hey, does this work on my local computer where I build it, or does that actually scale when I actually deploy it to all of my customers?" The answer is not always yes. There's many examples of that. After the experimentation, then you've shown that it actually works at scales, it's the adoption phase. We talked a lot about the adoption, then the final stage is the management. And so you can imagine there's failures at every single one of those points. The management is that sort of, you got adoption that was working and then eventually it sort of trailed off. So we won't talk too much about that, but let's talk about the development phase. One of the reasons why you might fail there is you might not have the necessary data to actually do it. You might not be able to build a model that's as accurate as you want it to be.
The other reason is you may, and I learned this analogy from my colleague Karim Lakhani, and I love this example, which is very often when people are building pilots, they try to build a pilot that works for the pilot, but they don't design it in a way to scale. And so his example is, if you want to get an elephant, you can't start with a squirrel. No matter how much food you give to a squirrel, it's never going to become an elephant. If you want to get an elephant, you need to start with a baby elephant, and then you need to feed it and grow it. And this is the same thing with pilots, which is very often, people build a pilot that is not designed to ever actually scale. And so yes, it works. You got the square. Yeah, it works, but turning the square into an elephant doesn't work. And so you need to start from scratch. And so a lot of people basically are unwilling to do that work. And so the advice there in the development stage is really build something that is designed to scale from day one. Don't assume that, oh, you can fix that after the fact that you have to do this, all of these things from the beginning.
And this isn't just like the infrastructure, but like things like privacy, right? There's this whole area around privacy by design. You can't just bolt on how your AI system handles data after you've built it and just hope it's private. You have to think about that from the very, very beginning. And so that's an example of where this, in the development phase, it will fail, and then it just turns into a slide deck where you say, "Well, we tried to build it, this and this happened," etc. And again, you can sort of look at it. We talked a lot about the adoption and how it failed there. There's a lot of ways to fix it in the experimentation phase. And this is where basically going, I think about it as going from the lab to the real world. A lot of things fail there and sometimes that's actually okay because in the lab, something might look really, really good and it might look like it's working, but then you deploy it to the real world and then there's unintended interactions that might be happening.
So, for example, maybe I've used Netflix as another example. Maybe in historical data, this recommendation engine was just doing a really, really good job of predicting what people actually wanted to watch. But when you actually deployed it, you saw that usage over like a two-week period actually went down. And the reason why it might be, well, it's really good at picking exactly what you want to watch and that's great, but that basically means I'm going to show you just the same content. And sometimes you just get bored of that and you need a little bit of something new. And so that's an example where, you know what, I'm okay with that failure because we actually learned something from it and so now we can design a better AI system. And so again, when it comes to ending up on a slide deck, you need to understand at what point did it fail. And also if you learn something from it, that's okay, right? Because it means the next iteration of whatever it is you're doing actually is going to be better. And so it's not necessarily a failure, it's actually a mode of learning.
Chris Linnane: So now I'd like to try to play a game.
Iavor Bojinov: I love a good game.
Chris Linnane: Make it even more concrete. Okay. I have three scenarios.
Iavor Bojinov: Oh, you've got cards as well.
Chris Linnane: I've got cards. This is like my late-night David Letterman, Jimmy Kimmel thing.
Iavor Bojinov: I love it.
Chris Linnane: My card's here. I'm going to read some scenarios to you.
Iavor Bojinov: Okay.
Chris Linnane: And it gets even weirder because I've got three paddles for you, Iav. And for the people who are watching, they can see this, and the people at home, I'll talk you through it. There are three paddles, and I'm going to ask you, I'm going to build a scenario, and I'm going to ask you, in that scenario, I would ship that product, which is a check mark. I'd pause that product, which is a pause thing, and then X, I would kill that product.
So I'll read the scenario. Think it through for a moment, because that'll give our listeners and our viewers a second to pitch in their own thoughts as well, and then you can tell me which one you think applies to this. And I'm not trying to brag or anything, but I spent a lot of time on those.
Iavor Bojinov: These are really good.
Chris Linnane: They're really, really, really good.
Iavor Bojinov: Really high quality.
Chris Linnane: Really high quality.
Iavor Bojinov: I think kindergarten level here for sure.
Chris Linnane: It could be second year. Yep. It could be second-year kindergarten. That can happen. All right. So, okay.
Scenario one, this is called the black box agent. A customer-facing AI agent improves conversion rates in testing. No one can clearly explain why it works. It behaves consistently, but its decision logic is opaque.
Legal is uneasy, and marketing wants to scale it. Should you ship it, pause it, or kill it?
Iavor Bojinov: That's a very interesting question.
Chris Linnane: Yeah. And you could change your mind if you learned more information. I'm giving you limited information on purpose.
Iavor Bojinov: Am I allowed to ask...
Chris Linnane: You can ask me a question, but I don't have a straight line with the company, so I'm going to give you false information about this getting information. But go ahead. Give it a try.
Iavor Bojinov: I think, straight away, I'm going to pause this one.
Chris Linnane: Okay. Why do you want to pause that?
Iavor Bojinov: And the reason I'm going to pause it is we need to understand what is our exposure and the risk here because we need to understand, yes, it is increasing conversion, and that's great. And I can understand why marketing is so excited, but that's great. I love all my marketing friends. But suppose the European Union hits me with a huge fine.
Chris Linnane: And you talked about this earlier.
Iavor Bojinov: And we talked about this earlier. I'm not going to be happy about that. And so I need to understand, and there's some questions that maybe not necessarily... I don't need to know exact, because if I'm using a large language model, I want to know exactly why it's giving these recommendations, but I need to understand that we have strong enough guardrails around data privacy, and I have a strong enough... You see these sort of multi agent systems where you have an LLM that's sort of monitoring. I need to make sure that that monitoring layer is strong enough that it's going to shut it down if I start to see some crazy behavior. And then what I'll probably do is figure out, again, I don't know too much about the company. If it's like a B2B company, maybe I'll start to deploy this with some either... Depends how it goes.
It could be with either sort of trusted partners, or it could be in a space that I don't really care about.
Maybe I'm an enterprise, I only really go after enterprise and maybe I have some SMBs and I don't really care too much about them. And if I feel like there isn't too much legal risk and I have good ways of stopping this, I'm okay with this. And so just to be very clear, the reason why I'm pausing it is not the fact that I don't understand the logic. I'm pausing it because I don't know enough about the guardrails.
Chris Linnane: It's the black box.
Iavor Bojinov: It's the black box. You want to open that up, and you want to see it. And sometimes it's going to be murky because again, we don't know why... I mean, mathematically, yes, we know that, but you can't really control it in a deterministic way. And so you have to be okay with that ambiguity, and so you just need to build guardrails around that.
Chris Linnane: Beautiful. Alright.
Iavor Bojinov: And maybe you take it off the pause once you learn more information.
Chris Linnane: Totally.
Iavor Bojinov: And I think in reality, you end up shipping it. Again, there's so many external factors, right? Am I really cash strapped and I really need to actually... You don't have the resources and maybe it's okay. I think there's many out of considerations.
Chris Linnane: Okay. So we're onto scenario number two. Okay. This one's called the agent with rare expensive mistakes. The title's a little scary, but an AI agent manages inventory. On average, it improves efficiency and reduced costs. Once or twice a quarter, it makes a mistake that costs tens of thousands of dollars.
Humans used to make fewer large mistakes, but many small ones. Would you ship this, pause this, or kill this?
Iavor Bojinov: Okay. On average over the year, who's costing me more? The humans or the AI system.
Chris Linnane: I think the humans are costing you more and making smaller mistakes.
Iavor Bojinov: When I add it all up, they ship it.
Chris Linnane: You're going to ship it. Alright.
Iavor Bojinov: I'm just going to go ahead and ship it. And I'll tell you why. AI systems, we can improve them. And if right now that's saving me money across the year, I'm okay with shipping it because I just reduce my costs. Sure, two times a year it would cost me 10,000, but if I'm summing up every week, I'm losing 100, 500, 200, 1,000, and that's actually costing me more. I'm okay with that. With the understanding that as we ship it, we're going to keep working on that model, and hopefully we can reduce the amount of time that this is going to make those major, major mistakes. Maybe there's some way of having a person review it, or maybe there's some other signals we can detect, maybe there's some better data. I think we can fix that. I can't fix the people because probably if I'm an organization that's been around for a few years, say maybe a few decades, I've probably trained my people to the max. That's as good as they're going to get. And so I would just go straight away and ship it.
Chris Linnane: You'd ship it, but if you had a real bad month where all of a sudden the AI was making a bunch of mistakes, you pause it.
Iavor Bojinov: Exactly. And we talked about the monitoring. If I understand in historical data that there's two times a year, I'm expected to really mess up. Well, first, hopefully we can figure out why, but if we even can't figure that out, we start to monitor that. And if it happens like two or three times, you shut it down, you go back to people. I'll tell you what, there's a big risk though, which is as you shift work from a person to an AI system, the people in the past that were doing this job, they start to lose the skill necessary to complete it. And we're already starting to see this even in really high stakes decision making like healthcare where we're starting to rely more and more on AI systems. There's this whole phenomena where one, people are getting these skilled because they're just not used to doing this without the AI.
And two, as the AI gets better and better and better, even if we are told we have to review this, we start to fall asleep at the wheel and we start to actually either not read the output. We just go, "I mean, the last 10 times it was right, just keep going because I don't want to read this anymore." So there's a huge risk as well. But right now, I mean, you sort of look at the economy, of course, tech sector and parts of it are doing really well, but I know a lot of companies out there are definitely struggling with tariffs and all these other things. If I can cut costs, I'm going to cut cost right now.
Chris Linnane: Fair enough. Okay. Last one. The tool with shallow trust. This is something you do with early on.
Iavor Bojinov: Oh, trust. Okay.
Chris Linnane: An internal AI tool is used daily, but no one relies on it without heavy checking. Usage is high, but trust is low. It has never caused a major failure.
Iavor Bojinov: So I guess I would wonder if we don't trust it, why are people using it so much?
Chris Linnane: It's a great question.
Iavor Bojinov: Right? If people are using it so much, they must be getting some value out of it. I think this is the example you said, which is you sort of use it, gets you 30 percent there, and then you need to do a bunch of work to fix it. I think I'm going to keep this going. I'm going to keep this going because if it already... We talked about how hard it is to get adoption. If you build a tool and you already have adoption, that means there is some level of trust in it. Maybe we sort of talked about those three pillars of trust. Maybe there's a lot of trust in the developers. Maybe there's already a lot of trust in the processes that you've put in place, and maybe there isn't that much trust in the AI model itself, but over time we're seeing this just in sort of...
Actually, in almost every category in AI, we're seeing exponential improvement in performance and capabilities.
And so I'm okay just riding that wave and saying, right now it might not be that good, but in six months, six months it will be good and I don't want to kill the excitement and sort of the culture that I've built around people using this. I'm going to keep leaning in on them and I'm going to put a lot of pressure on the team that's developed this to say, "Hey guys. You've got the adoption, you've done kind of the hard work. Make sure you keep it reinvesting in this tool and keep making it better and better and better."
Chris Linnane: Now we didn't come across anything that you killed, but that's okay. You tell me in like-
Iavor Bojinov: What would I kill?
Chris Linnane: What would you kill? Give me an example of something-
Iavor Bojinov: What did you want me to kill?
Chris Linnane: Anything bad idea, any bad idea. So, a bad use of AI that you instantly would be like, "We're not going with that anymore."
Iavor Bojinov: So that's really hard because the question that the leader should ask is not necessarily, is this something AI can do? It's like, should AI be doing this? And it really depends on the company, the culture, the data.
For one company, we talked about giving a lot of autonomy to some systems. One company just may not have the culture where they're ready for the employees to handle this type of heavy automation. And for those, I would shut it down because if your coach is not ready and your employees are going to revolt, you're not ready for that.
Chris Linnane: Okay.
Iavor Bojinov: You need to go back and figure out how can you do this in a more gradual step. So that's one big category, like it's your culture and are your employees ready. There are of course some things which are legally you're not allowed to do, and obviously you want to stay away from those. If there's regulations, say you can't do that, please don't do that. If the risks and the cost of mistakes are extremely high, I would be happy with some augmentation, but really you really want to keep that decision maker involved and have them actually keep practicing because if they stop doing that, they're going to get worse at it. And so those are the types of places where if the AI system really is not able to add that much value. And there's actually a bunch of examples of this where you have an AI system that's really good.
You have people that are really good. Doctors do this all the time. When you give this AI system to the doctor, the doctor performs worse than they did on their own and it performs worse than the AI system because the AI system makes them double guess themselves. They change their mind. And there's this whole quite bit of research on this, this whole paradox. And those are cases I would shut those down if I see that overall performance is actually dropping even if independently they're actually good.
Chris Linnane: That's super interesting. I've heard that about the doctor scenario too. I guess that makes sense.
Iavor Bojinov: Yeah. I mean, you just start to second guess yourself. And we don't want people who are used to making very high stakes decision quickly to really start to second guess themselves and to stop practicing that muscle because our jobs are not life or death. We can take some time, but if you're a ER surgeon and you have 15 seconds to make this decision and you start to second guess yourself because the AI system said something else and the air system's 85% of the time is right, then that starts to create a problem and I would shut it down.
Chris Linnane: If you thought of it differently, it's almost like would you ever put two doctors in an emergency room and then tie their legs together and make them move together? You'd be like, no, no, no, because they're always going to get in each other's way and their decisions and their planning.
Iavor Bojinov: Exactly.
Chris Linnane: It's a version of that.
Iavor Bojinov: Exactly. I think that's a great example.
Chris Linnane: I'd like to see an emergency room like that. That would be great. I would watch that show. There's enough doctor shows. This is the new one. Yes, the worst one yet. All right. So what's the feel of an organization when all of the pieces are going, I guess, going smoothly? So the evaluation, the experimenting.
Iavor Bojinov: I don't know if there's an organization that's going to-
Chris Linnane: No one's done it yet.
Iavor Bojinov: No one's done it yet, but there's some organizations that are getting close to this and I think it feels amazing because what's... And here's how I think about it, which is humans are forced to do a lot of things that we're not good at. And if you fundamentally take a step back and think about what is human intelligence versus what is artificial intelligence. We are very creative beings. We love social interactions.
We are really good at having conversations, engaging, connecting with each other. We're really bad at sticking to very prescribed, very regimented things. So I'll give you a very concrete example. When's the last time you opened up a map to figure out how to get from A to B?
Chris Linnane: So long ago.
Iavor Bojinov: Right.
Chris Linnane: So long ago.
Iavor Bojinov: We're really bad at that. Our minds don't work in that very algorithmic way. Of course, there's some examples and there's some amazing... And I have an undergraduate in maths, I did statistics. I actually do function in this sort of very logical mathematical way, but I recognize that we are a small, small minority. Most people don't want to have to do that type of work. They don't want to have to repeat the same task. It's boring. They want to be creative. They want to be social. They want to be interacting. And so when you get it right, the AI system takes over the tasks that people were bad at, and we've had to force ourselves to do because we didn't have any other way of doing it. So that we can then focus on the things that we're actually good. The problem solving, the creativity, the selling. Think about sales.
For example, I'm going to get these numbers wrong. So I want to give you the exact numbers, but it is in this case that we wrote on Microsoft. But the majority of time of their sales organization, and not 100%, something like 75% is probably getting this... The number isn't the case, so we can look it up, but more than 50% of the time was spent on administrative tasks. So not actually going and having an interaction with their clients, but either writing some report about it, putting together the slide deck, et cetera. That's not what these people want to do. They want to go and have conversations with their customers. They want to figure out how to solve their problems, what are the right tools for them, what are the right profit-sharing ideas so that they can really help their customers grow. But because we needed to have consistent work that we talked about the fail-safes and the monitoring that people have on people, because we had to do that, we created all these bureaucracies. And so you just spend most of your time doing things which are not actually fun or your job. And with AI, sort of in the B case, once we saw the successful deployment, there was actually a complete pivot where they're able to start spending something like 70, 75% of their time with their customers because a lot of that backend work was being automated. And so I think the feeling when you get it all right is you actually build a more human-centric organization where the humans do the things they're actually good at. And then the rest of the work that needs to get done gets done by AI systems.
So that's why I'm not that worried about jobs. I'm not that worried about the economy. I think the transition's going to be extremely painful because, as we talked about, people are so obsessed with the job description, they're not obsessed with themselves, which is to me very confusing, but hey, that's how it is. And it's just going to allow us fundamentally to just be more human. And so the most AI-centric organization that has an agenetic operating model is actually the most human-centric organization because the people do what they're good at. They communicate with each other in a much more level playing field where if there's a misunderstanding, there's an agent to help translate. So if a software engineer is being to a PM, the AI system can actually do that translation so you have better communication because software engineers sometimes struggle to explain what they mean and PMs struggle to explain why this is the framing of the product to the software engineer. And so by having an AI system there that can do that translation, has the complimentary skills, you just build a much more engaged and human-centric organization.
Chris Linnane: This is fascinating. That brought something up. When you talked about maps, I don't know if this connects in any way, but I grew up really... I loved looking at maps.
Iavor Bojinov: Yeah, same.
Chris Linnane: I love it. I love the... The idea, when I figured out what a topographical map was, I was like, "This is the most amazing thing to look at." But I don't like following maps. So it's just strange to think I can't make money just looking at apps at maps. I can't make a career just being like, "I appreciate maps so much. I go to a map store." But someone made maps for a long time so we could functionally use them, but now they've been replaced. So we just talked, I'm going on a little bit of a tangent. They've been replaced by these systems a long time ago and now AI. Now when we talk about job descriptions changing, people enjoy some of the tasks they do, and while that gets changed, now it's like, all right, well, my job has changed, but you took the heart out of my job. So what's the thinking there? People could potentially stay employed if they can bear down and deal with it, but the idea of their purpose changes for them.
Iavor Bojinov: Exactly.
Chris Linnane: And that's a little scary, a little sad.
Iavor Bojinov: It's really scary. It's really sad. And I think especially if you made a career of making maps, that's what you did. You made maps. Back in the 70s. And I love maps as well. If you come to my office, I have an old map of Harvard Square, which I love and it's great to look at.
Chris Linnane: It's artwork.
Iavor Bojinov: It's art. Absolutely. Absolutely. I think it's going to be really painful for those people. I think one of the ways that I've seen companies approach this, and this is going to change as we move towards more agenetic operating models. But for now, what they've basically have said to employees is, "You should figure out what are the boring parts of your work, figure out how to use AI for it." Keep doing the stuff you love and just do less of the stuff you hate. That's great for the next few years.
Chris Linnane: Sure.
Iavor Bojinov: But I think at some point, what's inevitably going to happen is that there will be parts of people's jobs that they love doing. And I think this is the Pernod Ricard example of the CMO, she loved figuring out how much budget to put to each of the brands, how to activate them.
Chris Linnane: She's got a very good attitude.
Iavor Bojinov: She's been doing that her whole time. That is what she's been doing. And 10 years ago, she was promoted to this. She had control over everything. She put 10 million here, five million here, gave us so much control over how the other brands interact with her. And the challenge is, well, the AI system's better. It's actually better at picking this than you because it has way more data and way more information. And so I think in those situations, leaders are left with a very difficult decision of what do you do with those people, which is if they're unwilling to give up some part of their work that they really love and they identify with to an AI system that is just far superior.
Chris Linnane: Wow.
Iavor Bojinov: What do you need to do? We probably need to get rid of them.
Chris Linnane: Yikes. So that transition is going to be-
Iavor Bojinov: It's going to be-
Chris Linnane: Underway to an extent, I'm sure, but it's only going to get more difficult.
Iavor Bojinov: It's going to get more and more difficult. And not naming any names of companies, but in tech, there's been a number of layoffs across very large organizations. And what I hear anecdotally is there's definitely some correlation between the people who are unwilling to be a little bit flexible around, "Hey, let's start using AI more to do these types of works." Those are the people that are probably going to be the first ones to go because if you're trying to... And this is an existential crisis for a lot of these companies, which is right now we're seeing a whole new wave of AI native, not digital native, AI native startups that operate in a fundamentally different way from how all of these other companies are operating.
And what we saw when digital companies came for non-digital companies 15 years ago is that digital companies mostly won. And so if you're a company, even a digital native company, someone like Google, you look around and you see perplexity, they're starting to nip your business, right? You see open AI, they start to nip your business. You're seeing lots of other companies that are just organizing in a completely different way, way lower margin of cost. You start to ask the question of, how long can I allow people to keep doing what they love if that thing can be better done using AI?
Chris Linnane: So looking towards the future a little bit. Let's cheer up. This is cheer up, but it's also kind of-
Iavor Bojinov: Oh, no.
Chris Linnane: Well, no, it's not bad news. So a few years from now, what do you think organizations will be most embarrassed by their current approach to AI?
Iavor Bojinov: I think the thing they'll be most embarrassed about is being so focused on narrow specific use cases and not trying to think about how to rearchitect their firms. And my favorite question, I do a lot of executive teaching, especially for CIOs, techie persons. I spend time with CIOs and the question I always ask them is, "Hey, if you were building your company, your tech stack from scratch today, would you architect it in the same way? Would your operating model be the same?" And they're all like, "Absolutely not. I would throw everything out and I would do this, this, this, this, this." I'm like, "Okay, great. You have your North Star That's where you want to get to." I realized that you can't just snap your fingers and magically it will happen. It's a painful journey, but most of these CIOs have done it with the cloud, they've done it with digital transformations.
It can be done. So I think the thing people will be embarrassed about is that they went, "Oh, I'm going to do a little use case here. I'm going to do a little use case here, but I'm not going to rearchitect the whole thing." And the example in sort of digital transformation times is you look at all the companies competing against Amazon. What did they do? They all had websites, but they didn't change their operations. They didn't change how they actually function. It took Amazon a couple of iterations to actually figure out how to pull this off, but when they did, no one else could compete. And so you're going to see the exact same thing, which is people are doing the same thing that people are doing in the '90s, which is I have a website, I have an AI agent, but you haven't changed anything. And so I think in a few years time, they're going to look back and be like, wow, why didn't I actually figure out my North Star of how I should orchestrate my whole organization being agenetic first and then take the steps to get there instead of just being like, "Oh, this is a really cool use case, so let's do that. Oh, that saved me 10 million. That's great. Everyone's really happy about it." But it's one narrow use case. I haven't actually changed anything about how the organization functions.
Chris Linnane: That's outstanding perspective. The idea of I have a website, you think you're winning and now you're like, "I have AI." You think you're winning, but the big picture needs to be clear, needs to be in focus.
Iavor Bojinov: Exactly. You need to change your operating model.
Chris Linnane: So when does this stop? When does AI stop? When does the high speed out of control feeling of every week is new? When does it either slow down or hopefully really slow down? Will that ever happen?
Iavor Bojinov: I don't think so.
Chris Linnane: Oh, my God.
Iavor Bojinov: I'm sorry.
Chris Linnane: We're always chasing our tails now.
Iavor Bojinov: And a couple of my colleagues call this... They call it the new sort of change athlete. I'm sort of moving away from AI transformation, that term, because when you think of digital transformation, it's a very clear end state. We go from being analog, we use paper to being digital. Great. We did it. We're digital. Job done.
Chris Linnane: Yeah. Different case here.
Iavor Bojinov: With AI, because of how it's moving, I don't see an end state. I think the end state is we become comfortable living in an ever-evolving dynamic world where capabilities are changing, our roles are changing, the work that we do is changing. And by the way, AI allows you to actually expand the type of work you can do. We've got some great research on this where we did it with the IG group and we basically showed that people can actually expand the type of work they do with AI. So the study we did was we had basically content creators in marketing, we had generalists in marketing, and then we had data scientists. And with AI, the generalists in marketing could perform as well as the specialist sort of content creators that really work on writing articles, a generalist could actually do it. And so these people are able to expand the type of work that they can do. And so what we're going to see is a lot more fluid, flatter organizations where people focus around problems to be solved, jobs to be done, and they sort of keep transitioning the work that they're doing in a much more fluid way.