The World Economic Forum Annual Meeting 2026 includes examination of how AI’s integration into work environments affects organizational roles and labor dynamics.
At Davos 2026, leaders argued AI is shifting from a point tool to something closer to an “intelligent co-worker,” but warned that the framing matters less than the operational change it demands. Panelists emphasized that AI excels at tasks, not whole jobs, and that organizational rewiring is slower than predictions suggest. BCG’s Christoph Schweizer described AI’s move from “a relatively static query into a very dynamic query” that pulls internal knowledge, adds external signals, surfaces patterns, and drafts deliverables—creating a co-worker feel. Enrique Lores noted the same dynamic in supply-chain planning and autonomy (“a co-driver”) and urged realism about error rates: humans err too, and AI can outperform legacy processes in call centers.
Trust and safety require more than “training on the right data.” Munjal Shah argued for empirical “output testing,” redundancy (“models that check models”), and escalation protocols to humans. Kate Kallot stressed that in data-scarce contexts, premature automation can “perpetuate extractives,” and warned a looming compute divide could become an “AI jobs divide.” On talent, Kian Katanforoosh reframed upskilling as mentorship: not content, but goals, feedback, and a visible “bar for excellence.” The biggest adoption mistake: treating AI as a tech project instead of a CEO-led change in processes, incentives, and human strengths like judgment and empathy.
Hello and welcome to the World Economic Forum's panel on The Intelligent Coworker. AI today is evolving from a tool to a coworker, transforming how individuals work, how organizations operate, and how societies adapt. The big question we're going to try and get at today is what happens when AI takes a seat at the table in the workforce? We've got a great panel for you today. I'm glad everybody can be here to join us. Sitting to my left, we have Enrique Flores, president and CEO of HP, Christoph Schweizer, who's the chief executive officer of BCG, Kate Callow, who's the founder and chief executive officer of Amini AI. Whoops, I'm sorry. Munjal Shah, the founder and chief executive officer of Hippocratic AI, and Kian Canton Ferrucci, who is the founder and chief executive officer of Work ERA. Please welcome our esteemed guest. When we get started, I'd like to basically ask this question, which may be an obvious question given the topic of the panel, but when we're talking about AI as a coworker, what does that mean versus AI as a tool? You know, when we're getting agents in the workforce, when we're starting to, like, turn over autonomy to them, where's the line? Maybe I'll start with you and we can come back this way. What's what is what's the difference between a coworker and a tool for a for an AI?
So, personally, I'm not a fan of calling AI agents. Agents? Neither. I'm a fan of calling it a coworker.
Fair enough.
I can explain, I think that right now, if you look at the publications that come out of the Frontier Labs, they all concern tasks. AI is good at these tasks, is better at that task. Humans have jobs that can make up hundreds of tasks at times. And it turns out that AI being good at a set of tasks is very different than AI taking on an entire job. And in fact, most of the predictions on job X, Y, Z is going away over the last three years has been wrong. We still have translators. We still have customer support managers. We still have a lot of careers that we thought will go away way faster than they actually did. And in fact, for a job to change, you need people to change, organizations to change workflows, to rewire. And we know from cloud transformation that that can take decades. There are still businesses operating on pen and paper 20 years after digitalization. And that's why I like to call it agentic workflows. Professor ng also talks about that rather than co-worker.
Fair enough, fair enough please.
So again, we break down the framework into of coworker in three buckets Copilots autopilots and infinite Pilots and Copilots are kind of working with the human in the loop. Autopilots do the task autonomously. But typically a task you do today, and then infinite pilots is this idea of things you could never do until you had an infinite supply, or you had a very, very low cost. And I think that while AI is going to augment our abilities, if you fast forward five, ten years, you know there will be 8 billion humans or whatever number we'll have then, but there will be 80 billion AI and meaning they will, but they won't be doing the things we do today. I think there's only a little bit of that will be impacted. I think by and large, we're going to find all of these use cases of things we never thought to do until we had an infinite supply at low cost. So, you know, one example is we recently called 16,000 people in for this one Medicare Advantage plan. In the US on during a heat wave. And we called them at the hottest time of the day, and educated them on what to do, told them where the cooling center was. I think even in some cases, you know, helped them get there. And, and that was something you would have never done because you would have had to find thousands of people to call, educate, do an assessment. And, these are the types of things I think we're going to see a lot more of.
Kate.
I would also tend to agree. For me, AI is always a tool, not a co-worker. And the reason is because, like Qian mentioned, we were forecasting that, you know, radiologists at some point will disappear because computer vision and AI, started to surpass, recognizing certain diseases and reading, radiographies better than the human eye, right? But actually today there are more radiologists actually getting into the profession. And what happened is that the tool has made that profession sexy again. Right. And I'm thinking that I do believe that we're going to see this happening everywhere. We are seeing this happening in governments we work with as we start deploying AI systems and AI tools across governments, governments, workers, civil servants are finding a renewed interest of doing their work differently with a different interface. But that interface is a tool, not their coworker, because that tool cannot actually make value based decisions, cannot say for a specific government or for specific citizens, this is the best outcome because that tool doesn't have that context just yet.
I see yeah, I believe that was Geoffrey Hinton, right. Who said that? I think it was. How long ago was that now? Like ten years? Eight years?
Eight years ago.
Eight years. Yeah. Okay. Christoph.
So honestly, I'm not going to get into definitions. What is a co-worker or a tool? But I'm going to give you a practical example from our own firm. So at BCG we have 34,000 people. We've been around for 62 years. And for decades we've been collecting our expertise and benchmarks and data about all sorts of things in the business world at our clients that we serve. Now, we have also collected a lot of information in our knowledge management systems over many years. And now, with the advent of AI, you could use these knowledge management systems in a very different way. You could say, well, we are about to meet that pharma company that has 27 factories around the world. Can you please pull all the benchmarks from our gazillions of files and tell us what the best manufacturing sites look like, and tell us, how does this pharma company compare to that? That's when I really started to kick in, and you got a lot more out of it than through kind of the traditional queries. Where we are now is you can further augment that. And I do think you are now starting to be in a reality where it does feel like a co-worker, whether you call it that or not. So what happens? You get the things from our own database. You can then say, please also pull all the information that's publicly available from analyst reports and recent kind of, expert kind of assessments, etc.. And please, also, after looking at the data, tell me, what are the ten most unusual and remarkable patterns in the data of those 27 factories for that pharma company? And you can even take it one step further and say, well, please, now put that on a coherent set of eight slides that we could use for an initial workshop. So it goes from a relatively static query into a very dynamic query, and now it goes into augmented information, it goes into getting stuff done. And then it does feel like a co-worker. And I think we are at that point. It's a reality. We very actively and happily build our own agents, and we embrace it in many, many things that we do. And I think we are at this point that AI is becoming much more than a smart tool.
I want to come back and find out what some of those things you're using for are. But first, Enrique, please.
Yeah, I'll use also a couple of real examples when I use AI to prepare for earnings and to ask AI to help me to prepare questions. I'm using it as a tool when we replace, when we use AI to replace and to change. How do we do supply chain planning in the company? It feels more a more a co-worker than than a tool. Or if I go into a Waymo car and I ask the car to drive me to San Francisco, it is a coworker, a co-driver or a Co something. It is not a tool. So this is clearly something that we see is going to continue to grow more and more and differently from other panels. I think these are going to be really integrated into our companies.
It's amazing to me how normalized Waymo's have already become, how normalized the autonomous cars are. Like my kids ride in them all the time. And it's just it's a bizarre world. They take it for granted. But I want to come back to something you said. You talked about this heat wave where you did 16,000 automated calls. Obviously, when you're doing something like that, you have to make sure that whatever you want to call them that are making the calls are safe and reliable or can be trusted. Yeah. Talk to me a little bit about that. How does that happen?
So, you know, there's a misnomer out there that, if you train your AI on the right data, then it's safe. You know, I don't know if you guys remember, at the very beginning, there was this thing called PubMed. GPT trained only on PubMed, a nice high quality data source. It still made up stuff. And so what we realized was the first thing you have to do to ensure safety and kind of responsible AI is you have to do something that almost no horizontal model does, which is output testing. So we literally hired 7500 US licensed nurses because we operate largely within the scope of an RN in the US, don't prescribe or diagnose, but they do a ton of care management. And so we hired 1500 of them. We had them call the AI. We had them act like a patient. We had said, hey, you're a nurse. You know what mistakes it made? Tell us all the mistakes. And then one out of every 50 times they got a human nurse on the other line, and they marked all the mistakes they made. And we said, all right, we're not going to ship this until you. We are as safe as a human. And actually, we're now much safer than that. But we created an empirical output testing model, I call it, rather than an input criteria for whether or not your model is safe. And I think more and more people are going to do this. You can't really do it on a horizontal model because it's just an infinite number of outputs. But you can do it on a vertical model. You can do it if you roll out one use case at a time and you can test it. It's expensive, but I think ultimately you have to do some of these responsible things to do it. We also have triply redundant systems that keep checking who builds the safety system that doesn't have redundancy. So this idea of one model to rule them all, you know, this one ring to rule them all concept. I mean this this is not how any safety system is developed anywhere in any industrial application. So we have models that check models that check models. And so these are the, some of the, the techniques you have to do. And then the last part I'll just say is in a healthcare call, if I say I'm having chest pains or shortness of breath, you know, you need to run something like a triage protocol and assess, is this an immediate red transfer to a human being? And so we're we actually built in an automatic ability to run, triage and transfer the call immediately to a human. And so there are humans standing by. They're not in the loop. So it's not a human in the loop. But we will kick out to a human if there's an unsafe scenario because it can happen on any patient call.
Can I ask how often that happens? Like how often does it.
It varies widely. I mean, if it's a congestive heart failure post discharge from a hospital, you know, this is day two or day one. It can be one out of every ten calls. But if it's a blood pressure check in of a remote patient monitoring that the guy didn't upload in days, and we're calling to kind of get the data, it can be 1 in 1000, 1 in 10,000. So it widely varies. It's just really use case dependent.
Kate, I want to come to you. When we're talking about data, it's really important that this data not be biased, that it you know, that it's that it, you know, especially if it's in the workplace, that maybe it's where something's been data has been scarce or extractive even, how does that play out when in places where there's not such rich data sets.
Yeah. So from a global standpoint, and double clicking on Africa and other emerging economies, we are still living in a reality where most of our systems are operating on analog, paper based, very unstructured data, PDF that are fragmented and scattered everywhere. And even when you look at a government systems, most of our governments still operate on paper format today. So while there is an appetite to want to adopt a lot of these tools and intelligent coworkers and adopt AI at scale, there is still a lot of limitations that we have to solve before we can actually get there, because if we are not adapting and grounding those models in our realities, we are actually risking to perpetuate extractives and to perpetuate cycles where the model will make a confident decision on very incomplete records. So, for example, if your civil registries are not digitized, if your land registries are not digitized, how are you going to be able to manage your country efficiently and make decisions with regards to your citizens? Are you going to listen? The model that seems very intelligent and confidently tell you a decision? Or will you actually bring in civil servants and continue having the human in the loop for, for, for that purpose? So that's why we focused on really supporting government to build knowledge base, which are specific to each ministries, each governments, each entities that really help them digitize and transform their data. Because for us, it's the first step before they are able to adopt some of those tools.
We've been talking mostly here about, so far about basically training AI, but I think there's also a world where AI is training people and helping people become upskilled in some ways. Can you tell me a little bit about that and how that works in the work?
But, yeah, so what if you look at the industry and I'll give the perspective of the enterprise, you know, that's summary of 44 enterprises, 4500, the what you want. First, the outcome is you you want an ideal scenario where whatever skills is needed to the business, everybody has it internally. You have it there. It's a zero skill situation, zero skills gap. That's what you want at all time. If the skill gap grows, you want to close it and you want to do that with, an approach that involves everyone, that doesn't leave people with a fear factor that gives them the confidence that they can make it, and they're part of this. So that's where the question of AI mentorship comes in. How do you design a good AI mentor that can perform these outcomes? And I think a lot of people narrow the mentorship problem to learning, because we've all seen the last decade of we take a course and we finish a course, the adoption is not great. It's very one size fits all. So the problem of mentorship is actually not that much about learning. It's about something else. I'll give you a story from from our class at Stanford. One of the things that Stanford does very generously is they put our classes on YouTube once in a while, and on campus we might have 1000 students online. We have hundreds of thousands, millions sometimes. And if you were to call those students online and ask them, what's the difference between a Stanford student and how you feel, they would almost never tell you it's the material. They would tell you that it's they don't understand the bar for excellence, and they don't know how far they are from that bar. When the Stanford student has friends at OpenAI, at meta, at Google, they have a constant feedback loop. They know the rewards that they need to attain, and they know what an opportunity opening means. And so what a mentor needs to be is really able to set good goals for a worker, to set good rewards for a worker, and to assess them effectively so that they understand their gaps. If you actually do that well, the learning problem is not an issue. People with a big enough carrots are going to do their best in order to achieve that carrot.
And you wanted to weigh in here?
Please. I wanted to go back to the conversation about data and about accuracy, because I think it is very important because we need to be careful not to be more demanding with the AI workers or with the models than what we are with our own employees, because nobody is perfect. Everybody makes mistakes. And what we have learned, for example, we have been using AI models in our call centers. They are not perfect. Many times they give the wrong answer, but they are more accurate that the people that with the humans that we have had for many years providing that work and the satisfaction with our clients when they call is much higher. So I think we need to balance how demanding we are versus because we as humans make many mistakes regularly.
So just to pile on to that look, I mean, we work with thousands of clients in their AI adoption and scaling and getting to real impact. And, exactly in line with what you are saying, Enrique. I mean, the the technology is not the bottleneck. The models work also the agents, which are still early days work or they are going to work that does not make a company successful or fail when it comes to their AI journey. They will succeed if they really change how their people work. And to do that you need to change processes, organization, incentives, skills, leadership, culture. And frankly, that is a much, much, much bigger factor. And I mean, our number one conviction as we work with leading companies around the world when it comes to AI, is make this a CEO problem and don't delegate it somewhere down in the organization. I mean, this needs to bring together all these important aspects. And if CEOs take ownership, if they upskill themselves, if they make bold investments, upskill their organization, then you get into this positive flywheel where people all of a sudden see, oh, wow, this really helps me. Oh wow. This is great experience I love this. And then you get into a self reinforcing circle. And I feel at this point in time, the whole world still talks way too much about the technology question. I think we should talk a whole lot more about the skills and the positive sentiment and the change management around it. I think that will make it or break it.
I saw everybody nodding while he was saying that. Does anyone want to weigh in on?
Well, I think one thing on the skills topic, Karpathy has a very, good phrase he coined called AGI artificial jagged intelligence. And that's what we're finding all over the place. Like, you'd be so surprised at the things the AI does. Amazing. And the and you'd be like, are you an idiot? Like, how come you couldn't do that? So I'll give you an example on both sides. Like, you know, we also deployed our AI to convince patients, to do follow up tests, because a lot of times, like if you have a lung nodule, you're supposed to do a CT scan. We got 1700 patients that basically the health system had tried to get to do it, and they had refused, and they had sent him a letter, text, everything. And we called him. And the AI was so persuasive, it convinced 250 of them to do it. And, and these are these have been quote. Lost to follow up, which is the health care way of just saying we gave up, but there was actually a cancer in there. So it's like we saved a life, right. And and then but now think about but on the other hand they're like, oh, great. Well now you can also use this to do scheduling. I mean, sure, straight up these times are available. Okay. But the AI, it doesn't have common sense. So give me three times. Sure. Mr. Shaw, would you like seven, 705 or 710? Like we would space them out, like we would try out different things or. Oh, I want two times for me and my wife, but back to back, because we drive in together. Oh, God. I mean, GPT four zero. We benchmarked it on scheduling accuracy 23% of the time it will hallucinate. And you literally need model checking, model checking models to actually get that sub 1%, because you can't have a clinic where 24%, you know, one quarter of the people walking in and they're like, I'm here for my appointment. It's like, what appointment? And so that you would have thought, that's an easy skill and you would have thought persuasion is a hard skill, and yet it's exactly the opposite.
Christoph, I'm going to come back to you, because what you're talking about a little bit there, I believe, is helping boost productivity in the workplace. How can you tell if that's happening? How can you tell if you've set up agents or or whatever? Again, coworkers, whatever we're calling these AIS, we have operating internally and a successful way. Well.
I mean, the dream of every CEO is that you can perfectly measure productivity. I can tell you the vast majority of us actually find it pretty elusive. Yeah. You have a lot of fun talking to your IT function or to your marketing function, trying to get an explanation how productivity changed. Yeah, good luck with that. So what do you really measure. And we found that incredibly instructive and predictive of success. You measure. First of all, does a certain organization do the employees in that organization habitually use the toolkit? What we see is if people use an agent agentic, AI solution or any solution once a week, it's a distraction. It makes you less productive. Once people start using it multiple times every day, or at least once a day, it becomes a norm. You learn faster, you get experience, you get a whole lot more out of it. So you do track adoption and usage. The second thing is you do what all large companies do. You do employee satisfaction surveys. Do you actually enjoy this? And look, we have seen that at BCG ourselves. We now over the course of 2025, had an exponential increase in the number of people who said, this is a real help and I'm happier because of the AI tools I have. We are now at 77% of bcg's who say Thank God I have AI, and it was much lower. And so I do think you have to manage some of these adoption happiness appreciation factors and then you will get to productivity and eventually you can also measure that. But I do feel again, as I said earlier, the whole business world talks about parameters as if they were God given and perfectly precise, immeasurable. I think there's a lot of kind of measuring what your workforce does, how they work, how they feel. That is going to be very predictive for success.
Let me share a data point that kind of supports what Christopher was saying. Every year we conduct a survey that we call the Workforce Relationship Relationship Index. And we basically ask people all over the world, multiple companies and countries, how do they feel about their work? Is they are they able to meet their own personal goals and professional goals? And something we have learned is that those using AI regularly in their work have a higher satisfaction level than those that don't. So this shows that really are starting to see the value that these tools bring. And they really feel not only more productive, but also better about what they do.
The AI is having a very positive 360 review, is that right?
Exactly.
By the way, you say that jokingly, it's actually, I mean, if you follow the title of this panel that it's a coworker. I mean, any coworker that's a human, you spend time to recruit it, you find it, then onboard it, instruct it, ramp it up, give it, give feedback, train it, and then promote it for bigger use. Or sometimes also let it go. You have to handle any agentic AI or coworker AI solution the same way, and you should not be negligible on that whole early part of it. There is onboarding, there is training, and you can measure the contribution as you measure it for any human being. And I think, I mean, if you take that whole coworker thesis seriously and we do, then I mean, treat it as such.
I was joking, but I agree with you. Kate. I'm going to come back to you here. To your earlier points about data scarcity. I'm curious if you have thoughts about where the line is in terms of, of having AI like, augment human judgment and where we should we should stick to humans.
So, I've been sitting on the Global Leadership Council for reimagining Aid, and I'm no aid expert. Right. But the reason why they decided to have someone focused on AI in the council was because, everybody believes that today we need to rethink, deconstruct and reconstruct how we're thinking about development and how we're thinking about having technology embedded into the ways we're making decisions right now. When you think about development, we see countries in the global north making decisions on where to invest in problems in the global south, but without really integrating the data that's on the ground. One of the biggest example of that has been USAID. So when USA was dismantled, there is an economist who started working on kind of like building a knowledge base of all the projects that USAID had done in the past and kind of building a knowledge management system to say, listen, we've we've been doing this for decades and decades and decades. There is a treasure trove of data of how those projects have performed on the ground. Where did we invest the money and how can we use these insights to be able to make better decisions and reconstruct that system in a better way? So that, for me, is a great use of AI, where I think we shouldn't cross the line is in making sure that is in having AI making decisions over where which citizens, for example, or which community in your country is more will deserve more to receive that capital than any others, because there's still uniqueness and opportunities and challenges that the model is not able to understand. And this is where all our uniqueness as different cultures, communities, countries is coming into play. But for us to get there, I want to come back to a bit of the elephant in the room, which for me is compute capacity, because I feel like we've been talking from a place of abundance. But in the countries I operate in, we're not talking about upskilling workforce, we're just looking into the fact that we're entering a new digital economy. And you have we forecasted over the next ten years to add 1.2 billion people in emerging markets who will reach the age of entering the workforce. But only 400 million of jobs are forecasted to be created.
That's that's that's a big gap.
It's a big gap. 800 million, most of them in Africa. What am I going to do with them? Right. If I don't actually provide them the infrastructure, the compute capacity, if I don't provide them, access, if I let a digital divide, which is compounded by a data divide become a compute divide, it's going to run right towards becoming an AI. Jobs divide. And we can't let that happen. So I think we need to kind of have mixed emotions when we addressing that question and also understand that there are realities on the ground that need to be addressed, and that's infrastructure, that's sovereignty, that's bringing GPUs. We were talking about this right before. You've been struggling to get GPUs in Africa, bringing GPUs and upskilling our youths to be able to use those systems, because it's becoming the default interface for all of your companies. We also need to be able to benefit from that.
How do you how do you make AI sovereignty happen if you have an opinion on this?
So that's what we focus on in my in my company, right. We build sovereign AI for countries in the global South. We work with governments to transform their data. We support them to understand, you know, that you don't have to build gigawatt factory to actually be able to have your own AI system. There is a different way to do it, and it's to understand what is the minimum viable compute infrastructure that you need to be able to keep your critical data. I'm not talking about the entirety of the country's data, but just your critical data as they relate to your citizens, to health care, in your country, sovereign in your country and within your borders. And then there is another layer to that which is, you know, being able to access some of the models to be able to fine tune, to be able to integrate some of those localized data pipelines in those models. So they are reflective of your countries and your realities. And then the first customer is the government, because governments are still driving a lot of the innovation and a lot of the the economies in the global South. And governments have to understand that they need to be able they need to be the first ones to uptake and show the rest of the ecosystem how it's done, and then be able to support developers, startups and innovators with access to compute.
Very good. I'm going to change the subject a little bit and come to you now, because what we're talking about here is making sure that humans have agency, I think, how do you make sure how do companies make sure that their workers have agency that that the that the, the AI systems that we're bringing into the workforce are, are are benefiting the workers themselves?
Yeah. Yeah. Agencies is trending. It's an important topic you can hire for it. But most organizations today, have to build internally. There's just not enough AI native talent out there. And they're all sucked by the hyperscalers and the AI startups. So you have no choice but to build internally. Now, to build those type of behaviors, you sort of need to reinvent HR, reinvent, lend. Last decade, lend and HR were seen as a benefit. It's learning as a benefit. It's self-directed learning. Here's the gym. Become a bodybuilder. We know it's not. It's not as easy as it seems. So what I've seen companies do that worked is be much more top down. The companies that make learning a business imperative trigger certain changes internally. And oftentimes companies are scared to do that, like, oh, I'm going to require you to have an AI driver license. It turns out the response that I've seen from workers is we would rather be required to do something and rewarded for it. And if the company is doing that, it's probably because they're not going to fire us. They're going to, in fact, invest in us for the future. And so I think the companies that are top down that are very clear about rewards are the ones who manage to get agency out of their employees. Additionally, going back to Kate's point, we don't fully know the skills that are going to be useful for the future, but we can't separate those skills into two categories the durable skills. The ones that we know better are going to be useful ten years from now. Problem solving, critical thinking, communication, coding. In my opinion. We also know there are perishable skills that change every six months. And so companies have to understand what is their strategy in durable and perishable skills and how you handle these differently because they're so different. And that's also an important consideration.
It's interesting you're a coding bull. Still people are falling out of fashion I've heard.
Yeah. No, I think everybody needs to be able to code. I'm not saying write code. I'm saying, you know, work with one of those tools and build your own stuff. And Andrew has been talking about that a lot. I fully agree.
Enrique, I want to come to you on something. There's I think there's another elephant in the room, which is that you're starting to hear about job losses in terms of entry level jobs and a lot of the, you know, a lot of the way that, that you grow as a, as a, as a person within an organization is by doing these entry level jobs, you learn from them, you get somewhere else. My first job was I was a fact checker. And, you know, it helped me learn how to be a better reporter, how to be an editor. So if if we're assigning some of this, you know, for lack of a better word, term, grunt work to, you know, to AI agents or having them make calls or having them, you know, organize tasks, that maybe the ways that people have traditionally learned, how can we make sure that the, you know, that early career workers are able to learn enough about the professions that they can grow in them?
I think the question is even broader than that, because what is very hard to predict today is what is going to be the structure of an enterprise five, ten, 15 years from now. If you think about how any of our companies is organized, is almost a pyramid and you have entry level workers, and then they you start making progress as you spend more time or you become more proficient, what is going to be the structure of a company that has been fully transformed by AI is going to be radically different. Today, we are organized by functions that are driven by the processes that we have been using during the last 30, 40 years. As Christoph was saying, AI is going to be start by transforming these processes. When you transform them, you are not going to need these functions anymore. You may need different types of functions. And what is the impact on the overall organization is something that we will be learning altogether during the next years. What I think the only thing we can know is going to be different from what we have seen until now.
Can I jump in? Please do. I think we're getting a glimpse into what that looks like, but I completely agree with you that, like, we're all you can always see the first order impacts of a new tech, but it's really hard to see the second order impact. But, one of the things we're starting to see is, is this abundance thesis that we really haven't, like, let's take healthcare the what's the ideal healthcare staffing. And it's really 1 to 1 education has that same property, right? One teacher to one student is the best way to learn. And you know, so but we only have in the US 5 million nurses for, you know, 300 and whatever million people, 360 million people. And so, but, you know, I remember my mom, she was diagnosed with high blood pressure six years ago, and she came home from the doctor and told me she was fine. She didn't tell any of us, and she didn't take her blood pressure med. And so then fast forward six years and we're in the hospital one night because she had 220 blood pressure and had the beginnings of congestive heart failure. And I felt horrible as a child. I said, oh my God, I should have been on her. But I didn't know because she came in. She said, oh, my doctor said, I'm the healthiest 78 year old he knows. Okay, great mom. But nobody from the hospital called him. Nobody from the doctors said, hey, Mrs. Shaw, you didn't refill your prescription for your high blood pressure meds. In fact, you haven't refilled it for. What's your blood pressure, ma'am? Can you take it right now? And so I think there's an infinite abundance we can absorb in health care. And so the big idea starts to. I mean, I think we have to rethink and say, what's the ideal staffing level that gives us the best health outcomes in that case and this vertical. And maybe it's so large that that's what the AI is doing rather than us. And then what are the humans doing? Maybe the humans are supervising all the AI's. Maybe the humans are the escalation point for all the AIS. Maybe the humans are helping to come up with novel use cases. Like, I think for all of human history, we have assumed scarcity in almost everything. We go to war over scarcity, pretty much. You know, we fight over islands that are not called the name of the thing. They are over scarcity. But like ultimately, for the first time in human history, we might have infinite abundance. And I think we haven't even begun to think that way, because our entire civilizations and our ways of thinking are built off of scarcity. And for the first time, we have to move to this new framework, and we have to bring it to all parts of the world.
I appreciate that, and I'm we're getting close on time, and I'm going to come to audience questions in just a moment. But before we do, Christoph, I want to ask you because you've got a really broad view of a lot of clients. You, you run a giant enterprise yourself, and then you see all these other organizations. What trends are you seeing people in the workplace taking up in terms of bringing agents in, in terms of bringing, you know, bringing in these coworkers or not coworkers? What are you seeing start to take off across the industry?
Well, what is fascinating for me is how much the functional mix of our AI work for clients has evolved over the past 12, 18 months. Initially, when I came to the world, there was a lot of use case around the call center, the customer contact center. Automate that. Make the agent the human agent more productive, better scripted, etc. very plausible, very intuitive. It's happening over the last 1218 months. We do now see that the use of AI and AI is going into the deep technical, value added areas of the most sophisticated companies. So what do I mean by that? It goes into pharmaceutical R&D, regulatory, clinical trial management. It goes into at the tech companies, the coding and software maintenance, documentation testing. It goes into sophisticated underwriting at major insurance companies. It goes into the marketing, content production and campaign design at the world's best consumer goods companies. So I mean, with all due respect to call centers, it's really important. But we are now seeing it go into the functions that either make you a winner or a loser in your industry. That's drastic, a massive change. The second change is what I said earlier. I mean, people are realizing, okay, I have a technology problem to solve, but I do really have an organizational and workforce problem to solve. And then the third thing is this whole question, Enrique, that you are also teeing up. I mean, an organization is not something static. That is an org chart. I mean, there are human beings in every one of these boxes who do things that qualify them to also eventually do other things. There's learning. There is kind of progression, there is move into other roles, etc.. And, I feel this whole question, how many people do you need and in what role? Okay, sure. I mean, we all have a job to manage that. I think we all have much bigger questions. What is the qualitative career path of a person in healthcare of a nurse? What's the career path of a person in HR, in consulting, in sales and marketing, in technology going to look like how are my coders of the year 2035 getting trained today so that they are world class? Then I think this qualitative question is becoming bigger than the quantitative one, even though from a societal perspective, of course there will be lots of talk about this, the quantitative.
This has been a very optimistic panel and I appreciate it. I want to turn to the room now. If you have a question, please raise your hand. We can bring the mic around. Yes. Right here.
So compared to.
800 person job gap, this is a very prosaic question. Excuse me but we started with AI doesn't do jobs, it does tasks. And then we heard that it you need to completely rethink skills and career pathing. A small question, but I think it's critical. A lot of companies are not getting ROI on their AI investments, because people don't know what to do with the six minute productivity gain or the 36 minutes. Right? We don't we don't save three months. We save bits of pieces of time here. What do we do about that soon?
I think it starts from really redesigning the processes. I think when this happens is because you are using AI to help you on your daily job, to really see the big impact, redesign your process and once you have redesign it, think how you will use AI and technology. Then the impact will be very different.
We're seeing a lot of impact on the go ahead, a couple of numbers. We're seeing a lot of impact on these emergent use cases. So one of our health systems spent $200 and got back $1.3 million. Another one spent $10,000 convincing patients to move back to their in-house pharmacy for their specialty meds. Got back $1.5 million a year. Another use case, that we saw was $2,000. And in the when I mentioned earlier about the lung nodule screening, they got back $2 million. So, maybe we have to think differently, because you're right, the what I call fragmentation of ROI doesn't allow you to capture it. But again, when you go to the emergent, then it's very different because nobody's doing it. So you're not fragmenting, you're not getting fragments.
At the same time. These three, five, 15 minutes have a big impact in how your team member feels about what he or she does. And this is very valuable as well. So you may not be able to improve or to reduce the cost by 15%, but that person feeling better is going to have an impact in the company also.
So I'm not going to talk about how to measure ROI. I think it really depends on the use case, but what I can say is a lot of companies are in echo chambers, and the difference between top tech AI companies and the rest is becoming increasingly drastic. I'll give you an example. Prompt engineering has a very different definition in the top tech AI companies versus others. Some people think of it as how do I guide the model? Some people think of it as zero shot, few shot prompt chain of thought prompts, agentic workflows, retrieval, augmented generation reasoning. It's just that there was six minutes. My guess is maybe in this organization it looked like six minutes a day. In another organizations, it would look like the agent is constantly working behind the scenes because it's just been built different.
And I think we have time for one more question. If anyone has one. No. Okay. Well, I'm going to I'm going to actually leave you guys with a question, which, we've talked a lot about everything that's going right. What, what are the mistakes that organizations are making right now when they're trying to bring agents into the workforce? If we can just get a quick answer from each of you, because we're very close to being out of time.
So I look a lot into talent decisions. Talent decisions are shots in the dark. They're almost all wrong today. We make really poor talent decisions hiring, firing, performance management, upskilling and reskilling. It's very wrong. Internal mobility. And I think in the next few years, we're going to start seeing AI be way better than the best humans, the least biased humans at making talent decisions. And going back to what Christoph said, the context that's going to go into all the agents, you need some context around the skills of the person if you want to help them effectively. That will come from AI, not humans. And I would imagine that in a few years it might even be illegal to not use an AI for interviewing. I wouldn't be surprised because we know it's worst.
John.
So I think that you have to focus on experimentation. Too many times we're deploying AI the way we deployed software, but software was a specific intelligence that did that one thing. So the pilot with ten people in your organization would extrapolate to the experience with the software, with everybody else. But like if you get a if you put in a prompt in a ChatGPT and get a bad answer, do you throw away ChatGPT? No. You try a different prompt. And so you don't want to accidentally give up on the entire initiative because of your first beachhead was the wrong beachhead. And so really, we're recommending you do ten use cases in parallel, because actually some work, some don't work. And it's kind of hard to predict at this stage which ones will and won't. So that would be my my advice to folks on what we've seen.
Kate on my side, thinking that all of this is just a magic wand and everything. Just as soon as you deploy, everything will roll. I think we also need to rethink how we reskill some of our workforce, especially in the case where if we consider that agents are going are going to be co-workers, you have to move some of those individual contributors into learning how to manage some of those. Right. So how do you help your workforce get to that next layer? Is going to be my advice for everybody.
So many things can go wrong, but I think there are two that stand out. The first one is you try to or you handle this as a tech challenge. You just use a technology and you try to get a tech solution. You will get a tech solution, but you're not going to change much for the better. The second thing that can go wrong is you lose sight of what makes humans special. And I do feel exactly in line with what you said earlier. There are things that humans do that are unique and will be unique for a long period in time, an ability to deal with ambiguity, judgment, empathy, care, relatability, a willingness to motivate others in a way that they do things differently going forward. They won't disappear. In my mind, they will only get more important if you miss that in your organization. I think you're in real trouble.
Enrique, take us home.
Yes, I think we. When we talk about AI, many times we think about the cost implications this is going to have. It will have and it will be positive, but also it will help companies to move faster, to innovate better, to provide better customer experience. And we need to make sure we look at all these elements, not just as the cost impact of it.
Well, we are out of time. Thank you all so much. This has been a great discussion and I really appreciate it.
Thank you.