23 Apr 2026 · 53m · Sean Brown, Tanguy Catlin, Sandra Durth, Lari Hämäläinen, Antoine Montard · 21 quoted facts
The McKinsey experts argue that generative and agentic AI is a true general purpose technology, comparable to steam and electricity. CEOs therefore cannot treat it as an IT upgrade. They need one view that joins strategy, technology and organization. Productivity gains alone will not create advantage, because competitors can buy the same models. Lasting value will come from redesigning workflows, innovating, and rebuilding the workforce and operating model around humans working with agents.
The central tension is speed versus slow economic payoff. Investment and model capability are moving extraordinarily fast, and coding has already flipped from skepticism to heavy use. History says full economic impact takes decades and needs complementary innovation. The guests also disagree on how painful the transition will be for workers. They agree that investing in human capabilities and running bold experiments are no-regret moves.
Antoine Montard and Tanguy Catlin argue that every competitor can access the same models, so productivity gains become table stakes that mostly flow to tool makers and consumers. Advantage comes from innovation unique to the company.
Catlin says simply rolling out AI yields about 20% productivity gains in software development. Firms that redesign end-to-end workflows, such as compressing two-week sprints into one day, capture far more.
Lari Hämäläinen says the length of tasks models can do rose from seconds in 2023 to over ten hours now. Developers who saw a 20% slowdown in tests later showed a 20% gain and resisted working without AI.
Hämäläinen imagines a firm with 50,000 employees and 500,000 agents, where agent costs could be double the payroll. Relying on one provider would hand an outsider control of that spending.
Sandra Durth notes about 30% of current work could be done by technology, spread across all roles. Catlin says about 75% of the workforce would need their jobs reconfigured, which makes skills and change capability decisive.
They mean it reshapes whole societies, industries and company performance, as steam and electricity did. Tanguy Catlin says it is not just another wave of technology.
Probably not on their own. Antoine Montard and Tanguy Catlin say competitors can use the same models and reach the same productivity, so lasting advantage must come from innovation specific to your business.
Lari Hämäläinen says a new generation of models could handle much longer tasks, crossing roughly a five-hour threshold last fall. Developers who once slowed down with AI now refuse to code without it.
Lari Hämäläinen says yes, especially if agents become a large share of the cost base. He says a company should not let one hyperscaler control its operating expenses, and the performance gap between top models has halved in a year.
Sandra Durth says investing in human capabilities is a no-regret move, so everyone should become AI fluent by doing, not just by training. Tanguy Catlin adds that HR, technology and strategy leaders should learn together.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Tanguy Catlin said US capital expenditure on data centers, hardware and energy would be about $7 trillion between 2026 and the end of 2030.
“about 7 trillion dollar of capital in expenditure across data center infrastructure, hardware, uh, energy between 2026 and the end of twenty thirty.”
Lari Hämäläinen said US data centers are projected to consume about 10 to 13% of US energy by 2030, up from about 2% the prior year.
“When we go to twenty thirty, the projection to currently is about ten to 13%, depending on the level we end up.”
Lari Hämäläinen said US electricity generation capacity has not really increased in the last thirty years.
“in an economy where actually the generation capacity has not increased really in the last thirty years.”
Lari Hämäläinen said about 50% of the generative AI workloads running in data centers now come from programming.
“50% of workloads today that are that gen AI uses, so that I use to run the data centers, actually come from programming.”
Lari Hämäläinen said the compute used to train leading AI models increases roughly four to five times each year.
“We're putting four to five times more compute to train the models.”
Lari Hämäläinen said the performance difference between models from OpenAI, Anthropic and Google has halved in the last year.
“if you take an OpenAI model, Anthropic model and Google model, it has halved in the last year”
Lari Hämäläinen described a scenario in which a company with 100,000 employees has 50,000 employees in three years, alongside 500,000 AI agents.
“If you have 100,000 employees as a CEO today, and let's say three years from now, which is actually in our projections, a viable scenario, you might have only 50,000 employees”
Antoine Montard said steam engines were first used to pump water out of mines and later replaced water mills in factories.
“If you think about steam engines, that were first used to pump water out of mines, and then they replaced water mills in factories.”
Antoine Montard cited Robert Solow's 1987 remark that the computer age was visible everywhere except in the productivity statistics.
“In the 1987, Robert Solow famously said that you can see the computer age everywhere, but in the productivity statistics”
Antoine Montard said e-commerce took about twenty years to reach roughly 10% of retail.
“e commerce took about twenty years to represent about 10% of, uh, retail.”
Antoine Montard said John Maynard Keynes predicted productivity growth would reduce the working week to fifteen hours by the year 2000, which did not happen.
“in 1920, uh, Keynes, uh, predicted that thanks to productivity growth, by the year 2,000, we would be only working fifteen hours a week.”
Antoine Montard said AI reduces transaction costs such as searching, evaluating, contracting and switching suppliers, which could undermine intermediary business models.
“AI fundamentally reduces transaction cost, that's the cost of searching, evaluating, uh, contracting, switching suppliers.”
Lari Hämäläinen said that in 2023 AI models could perform about twenty seconds of work by a human developer.
“in 2023, the models could perform twenty seconds of work by a human developer, which was negligible, right?”
Lari Hämäläinen said AI models crossed a task length of about five hours of human developer work in the previous fall.
“And last fall, we crossed about five hour task length threshold.”
Lari Hämäläinen said a group of developers that was 20% slower with AI in an earlier benchmark showed a 20% productivity increase when re-benchmarked in February.
“when the same group was ben benchmarked this February from minus 20%, so them slowing down, we actually have gone to 20% productivity increase.”
Tanguy Catlin said some organizations compressed two-week software development sprints to one day, with agents coding overnight and humans handling requirements and testing by day.
“We've seen organization move from sprints that would take two weeks to compress them all the way to just one day”
Tanguy Catlin said today's AI models run on chips designed two years ago and researched four years ago.
“The models we have today are based on chips that were designed two years ago and researched on four years ago.”
Sandra Durth said about thirty percent of current work could or should be done more economically by technology, spread across all occupations.
“What you see is thirty of current work could or should be economically more viably done by technology.”
Sandra Durth said some technology organizations redeployed 70% of their workforce within nine months.
“there's organizations in tech who've, you know, redeployed 70% of their workforce within nine months because they had to and it was possible.”
Tanguy Catlin cited Professor Agarwal of the University of Toronto, who argues that AI makes prediction cheap and so raises the value of complements such as data and judgment.
“Professor Agarwal from the University of Toronto, in his, uh, most recent book, made a very simple, um, representation of the fact that the machine right now is really, really good at prediction”
Lari Hämäläinen said the McKinsey Global Institute wrote a book called A Century of Plenty arguing that long-run technological change brings large societal and economic benefits.
“We wrote a book by MGI, where Hank is a director, called The Century of Plenty”
Sets up the episode's three lenses of strategy, technology and organization. It explains why executives find the signal hard to separate from the noise.
“executives are really struggling to make sense of all of what's happening.”
Covers the roughly $7 trillion US buildout and its energy demands. It also argues that existing internet services already supply a return.
“The level of investments here is quite unprecedented.”
Explains why model performance gaps are shrinking. It also explains why companies should avoid depending on a single provider.
“it's not necessarily a winner-take-it-all scenario”
Shows that transformative technologies take a long time to produce economic impact. Complementary innovation is what unlocks the gains.
“the technology becomes available very quickly, but the full impact”
Explains why productivity is only table stakes. Innovation and shifting value chains are where the money will accrue.
“value is unlikely to come from productivity alone.”
Uses developer productivity data to show how fast perception can change. It argues that real value requires redesigning workflows, not just deploying tools.
“when your organization tells you that, hey, we tried this, it didn't work, I think we should all say that, hey, hold on, let's try again.”
Examines how AI touches every role and why organizations must become flatter and more adaptable. It also covers how quickly people can be redeployed.
“you're just gonna make organizations wonderfully efficient at doing the wrong things.”
Each guest gives one piece of advice for integrating strategy, technology and organization. The answers focus on learning moves, bold experiments and technology architecture.
“bring your team to learn together, that's how you get multiple lens in the room.”
CEOs are increasingly recognizing that generative and agentic AI are just not another wave of technology. It is a true general purpose technology in that it will reshape societies, industries across geographies, and, most importantly, the performance trajectory of their own companies. And yet, when we spend time in conversation with them, there is a consistent theme that keeps coming back. And it is the fact that executives are really struggling to make sense of all of what's happening.
From McKinsey and Company, I'm Sean Brown, and welcome to our three hundredth episode of Inside the Strategy Room. That was Tanguy Catlin describing why generative and agentic AI represent something fundamentally different from other past waves of technology. And the challenges that business leaders face in understanding the implications and opportunities that this represents for their businesses. This week, we'll explore what that shift really means for CEOs and their organizations. I'm joined by Tanguy, along with Sandra Durth, Laurie Hamalainen and Antoine Montard to unpack the CEO agenda for thriving in what we're calling the agentic age.
We'll dig into the sheer scale and speed of AI investment, how leaders are approaching this moment, where they're seeing impact and value, and what it takes to move from experimentation to real, sustained performance gains. And throughout the conversation, we'll return to one core idea that succeeding in this new age requires integrating perspectives across three lenses strategy, technology, and organization, all at the same time. And now it's my pleasure to introduce our guests. Tanguy is a senior partner in our Boston office and a director of the McKinsey Global Institute, where he leads research on economic and business issues critical to the world's companies and policy leaders.
Tanguy, it's great to have you here today.
Thanks for having me, John.
And Sandra Durth is a partner in our Cologne office and a leader in our people and organizational performance practice. She helps organizations unlock lasting performance through transformations in talent, leadership and culture. Working at the intersection of technology and people transformation, Sandra leads our thinking on the agentic organization and what it means for the future of work. Sandra, thank you for joining us today.
Great to be here, Sean.
And Laurie Hamalainen is a senior partner in Seattle and leads our gen AI and agents work within Quantum Black, our AI practice. He works on large-scale strategic and operating model transformations, both for technology companies and companies adopting technology. Laurie, thanks so much for joining us today. Good to be here, And Antoine Montard is the capability director of our strategy practice based in Lisbon. He has extensive strategy experience in both advisory and corporate contexts and leads our strategy research and insight center, where he develops novel approaches for our strategy practice, building capabilities of our teams and clients, and he also serves as a strategy expert on client engagements.
Antoine, welcome to the session.
great to join you today.
So we're excited for the conversation today. And Tangi, let's start with a bit of context. You're spending a lot of time with CEOs right now who are grappling with the fundamentally disruptive nature of these technologies and how to succeed in this new age. What are some of the themes you're seeing repeatedly emerging in those conversations?
A certain vision, as you said, CEOs are increasingly recognizing that generative and agentic AI are just not another wave of technology. Actually, it is a true general purpose technology in that it will reshape societies, industries across geographies, and, most importantly, the performance trajectory of their own companies. And yet, when we spend time in conversation with them, and and we have many of those, uh, there is a consistent theme that keeps coming back. And it is the fact that executives are really struggling to make sense of all of what's happening.
There are some powerful signal, yet there is a lot of noise that is coming constantly and very fast. And so if we take a step back when when we listen to executives at the macro level, they see some profound shifts underway. Those could be geopolitical realignments as semiconductor manufacturing moves onshore, or massive capital investments, which we'll explore with Lari, that are reshaping economies, or breakthrough in industries like drug discovery. But at the same time, Sean, they are grappling with real risks cybersecurity threat, workforce disruption, physical constraints.
And so that is weighing on the promise of productivity gains that could be offsetting some of the aging populations or transform medicine, as I said earlier. But also some real concerns about job displacements and rising inequality. And then you go back to the micro level and the questions become more immediate. How will profit pool evolves in their industry? What will their competitive advantage be in the future? How much to invest, where to invest, how fast should they build their technology, should they partner, how to drive adoption of AI across the organizations while upscaling the workforce at scale.
And so they are facing this breath of questions, and they are trying to get sense of what are the signals, uh, and as you know, the market is crowded with solutions and strong point of view, so there's a lot of noise out there. So what we really wanted to do is find a forum where we could bring together three lenses the strategic lens, the technology lens, and the organizational lens into a coherent whole. And so our ambition today is not necessarily to answer every question, but instead it's to offer a more integrated perspective across those lenses, highlight where there is clear alignment, but also where there are thoughtful debates to remain.
Thank you, Tanguy. And as we think about the importance of integrating across these three lenses in your work with clients, what's the biggest barrier to having this happen? What what makes it the most challenging?
Well, I think there is, uh, first, an issue of executive education. This technology is quite new, few of our executives really understand, uh, it at the level required to be able to integrate those three lenses. And then there is an issue of fragmentations. Uh, it is not often that your CIO, your CEO, your head of strategy, your head of HR really spend some quality time together debating those questions, or at least with the frequency necessary.
I would also add Sean there that one thing we see is that technology has never moved faster than it has in the in the last five years or so. And, uh, I think the capability horizon of what we can do with the tools is increasing basically monthly. And I think we'll touch upon it today, but th- there's certain ambiguity related to that. And what, what actually was true a year ago might not actually be true anymore today.
And how do you then think about such such uncertainty in the context of a very large organization of tens of thousands or maybe over 100,000 employees? And I think that change management is a real challenge and something we'll explore a bit today.
Well, it's clear that AI is scaling rapidly. It's something that's in the press all the time. But maybe you could share a little bit of an overview around how these investments in AI are actually accelerating. And and how does the investment in this technology compare to other investments in past technologies?
The level of investments here is quite unprecedented. Uh, Sean, we are talking in the US. about 7 trillion dollar of capital in expenditure across data center infrastructure, hardware, uh, energy between 2026 and the end of twenty thirty. That that is very significant. And what is even more staggering is the fact that that investment in the US dwarfs any other investment elsewhere in the world. It's actually five times the combined investment of the rest of the G6 and China combined.
So with that level of investment, you start to see incredible level of progress that Lari, maybe you want to comment on.
I would just on the level of investments, I think the the pace at which we're going is amazing. Now, the obvious first question is that what does it imply? I think a good figure we've been debating is energy consumption. And, uh, if we look at simply US, we're moving from, I think last year was about 2% of energy being consumed by data centers. When we go to twenty thirty, the projection to currently is about ten to 13%, depending on the level we end up.
So we're five folding the energy needs from data centers in an economy where actually the generation capacity has not increased really in the last thirty years. And I think managing that, uh, and we can all imagine what it means for generation, infrastructure, transmission, distribution, and for the equipment supporting that. So it's a it's a huge society wide transformation. The second thing I would like to highlight is that there's been this debate on ROI. I think one framing, if we look at first of all internet today.
So all of us use search, most of us watch Netflix or similar streaming services. What basically has been happening in the last three to four years is that we moved the backbone of internet, so the algorithms feeding those services to run on gen AI technology. So previously, for example, a lot of the recommendation algorithms used to be traditional machine learning, right? Today, they all run on gen AI. So when people talk about ROI, it's often the unseen that is actually driving the capacity utilization, which is that all the services we use in our daily life just moved over.
Keep that in mind. Second, 50% of workloads today that are that gen AI uses, so that I use to run the data centers, actually come from programming. That 50% has literally emerged more or less in the last six months. So what happened is that we had a step function, which was to move the internet. In the last six months, we got another step function, which was programming, but only in the six months. So it's very recent.
And I think the question that we're all asking is that will enterprise AI be the next step function? When you as CEOs and executives think, when you think your finance function, when you think your HR or your operations, whatever business you're in, will those workloads, and sorry to call them workloads, but will that work transfer to be gen AI paced? And will that fuel these data centers here? So, but I think so far we've seen the ROI, and I think that's why the investments feel justified.
And there's no actually there's a shortage of capacity on the market as we speak. What we're seeing, uh, are a couple things. So, first of all, we're actually seeing that the compute needed to put into the models increases roughly every year. We're putting four to five times more compute to train the models. Each of us is using in our daily lives, right? Four to five times more compute capacity. As a result of that, what has happened is that the number of models we're training is actually coming down because the sheer capacity needed to put into these 1 trillion parameter models is so big.
Uh, at the same time, what is happening is that we're three times improving the algorithmic efficiency of the models. So what that means is that a model that performed at a certain level a year ago, we can train three times more efficiently this year because of research, because of development, the way we optimize them. And what is resulting is that today, I think we have, I would call it like a really, a multitude of different applications.
So we have the reasoning and deep research agents, we have coding agents, you know, we start to have quite deep workflow integration into enterprises. And then, fourthly, I think it's fair to say that almost any benchmark on AI and how they compare to us humans, AI has already kind of made it obsolete, meaning they have surpassed the value of those benchmarks. So, in synthesis, we're putting huge amounts of investments so far, those investments actually have yielded returns, but not in enterprise.
And fourthly, I I think the development is, uh, is quite amazing and fast.
In terms of the different L L models, are you seeing sort of an a winner-take-all situation emerge? Or are CEOs placing big bets on one technology or one of the LLMs? Or should they really be looking to diversify their investments across a multiple different models?
So I have three arguments, and colleagues may challenge me because we have slightly different view, maybe of this. So one argument is that actually the performance difference between the LLMs. So if you take an OpenAI model, Anthropic model and Google model, it has halved in the last year, which might sound surprising, but if you think of it from the compute that goes in, we need to put an exponential amount of compute into this model. So that four to five times more every year.
What means is that the difference between the models is actually becoming smaller and smaller. I think, like when we think of the next generation of applications, then, which is how do they apply to your company? And which which models do you use, how do you use them, et cetera? Actually, the difference comes down to things like tonality. How does the model behave in different scenarios? It does not come anymore down to raw performance differences, and that is a very subjective thing.
It's almost like hiring an employee. Which model do I hire as a CEO? Because they have distinct characteristics and differences. Then I think, thirdly, I, I would say to the question of is it a winner-take-it-all scenario? One of the key questions for enterprises in the future is that do they want to use frontier models or not? So I'll give an example. If you have 100,000 employees as a CEO today, and let's say three years from now, which is actually in our projections, a viable scenario, you might have only 50,000 employees, depending on your business.
But you have 500,000 agents that act as employees. Do you want those agents to be from a hyperscaler or one company? Because that cost base that goes into the 500,000 agents might be twice as large as your employee cost base. And that's actually a very strategic question, because then you would actually give your company, you, you, somebody else would control your opex, and I don't think that's an acceptable scenario for any enterprise.
So I think, Sean, it's not necessarily a winner-take-it-all scenario, and companies especially need to make sure that they're not too dependent, in my mind, on just one provider.
And we've talked now a little bit about that technology lens, but maybe we can shift gears and focus now on strategy and how these technologies have evolved in the past. And maybe that might give us a little bit more insight into what might happen in the future.
Yes, absolutely, so I agree with everything we've been saying about the magnitude of the moment. We've seen the same movie a few times in the past, and I think it's really helpful to understand how that happened and what was the impact and the time scale. If you think about steam engines, that were first used to pump water out of mines, and then they replaced water mills in factories. But the real productivity gains only happened when factories got relocated to urban centers, and we had the the birth of manufacturing hubs, such as the textile hub in Manchester.
And even more so when, uh, we had steam locomotives that were invented that allowed much more complex supply chain and, uh, deeper access to markets. And so all of that took a quite a lot of time. Electricity is similar, uh, factories were redesigned around the electricity motor, away from the vertical uh, steam shaft. That allowed assembly lines and flow based designer factories that increased productivity quite a bit. And then you had complementary innovations such as the light bulb or refrigeration that really created a lot of value in the uh, in the economy, and all of that took quite a bit of time.
In the 1987, Robert Solow famously said that you can see the computer age everywhere, but in the productivity statistics, and that was in the context of the computer revolution. And even in eighty seven, uh, we couldn't really see the full impact of it. Arguably, we started seeing the impact in the nineties, with again innovation around ERPs and, uh, modern supply chains, and that took quite a bit of time and a lot of further innovation.
If you think about a bit closer to home and what we've seen over the last twenty five years, e commerce took about twenty years to represent about 10% of, uh, retail. And it took a lot of innovation to be able to deliver at scale, um, all of that volume, uh, of shopping. Uh, autonomous vehicle is similar, like it's been around for ten years. It's an amazing technology, but, you know, it's still very costly. They still struggles to manage edge edge cases.
The roads are still the same, the rules around the roads are still the same, and there's still a lot of innovation to be happening, uh, to enable that technology. And so what we've seen again and again and again is that the technology becomes available very quickly, but the full impact, the economic impact from that technology, take quite a bit of time and a lot of innovation to come through in the economy.
And Antoine, that gap between the initial adoption of the technology and the time when it actually delivers real value is obviously of great importance to CEOs. And as Tanguy related earlier, they're grappling with questions like, where is the ROI actually going to come from? And it seems that there are two groups that are looking for this ROI. The companies that are making these massive capital investments in the foundational models that Tanguy talked about earlier, but also the corporations that are now investing in adopting these AI models in their own businesses.
Could you maybe share some insights on how this AI value creation might play out over time?
Absolutely, so, uh, I'll, I'll take a few different angles, and we call that waves of of AI technology and its impact. The first one is productivity, and we were talking a lot about productivity here, and and you can think about productivity in terms of automating tasks and improving the quality and cost of those tasks, all the way to workflows and systems. Typically, productivity gains accrue either to the tool manufacturers, so in these cases, those foundation models, or consumers, right?
But very rarely to companies themselves. There may be some temporary advantage if you're getting ahead faster than your peers in your industry, but value is unlikely to come from productivity alone. It's basically setting up the table stake in the industry and you have no choice but to go after it, but it's unlikely to become a a big, uh, source of, uh, of value. I think innovation is kind of the the next layer, and I've talked about the steam locomotive and refrigeration and light bulb.
That's the equivalent here. What are the innovation that are made possible with AI, and what are new businesses or business models, products and services we could create? So we're starting to see, uh, drugs that are, uh, driven by AI research. We're seeing kind of AI tutoring, we can see real time offering tailored with AI. There's many different innovation or robotics that, you know, are coming through. This is where growth and value creation will be the most acute, and to a large extent, the future remains to be created.
We don't know yet where will most of value be created. And then the the the third aspect is about transaction cost. AI fundamentally reduces transaction cost, that's the cost of searching, evaluating, uh, contracting, switching suppliers. And, you know, with today's technology, you can imagine having an AI switching utility provider or or your banks. And all of that actually puts in question all of those business models that relied on aggregation, brokering, uh, intermediation, and might fundamentally shift a lot of those value chains and the control points will be moving.
That's more of a where will the uh value accrues in the value chain. And again, here there's a lot of unknowns, but that and innovation are where really the money will will accrue in the future.
And Antoine, if I can add just one thing that I think is important right now. Most organizations are obsessively focused on productivity, uh, and I think that one of the takeaway from what you are saying is that's where you lose the battle, not where you win the battle. And one thing that I think is important to understand is the reason you don't win the battle on productivity is your competitors have access to the same models. And so they will pursue the same strategy, and they will gain the same level of productivity versus your innovation, something that's quite unique to you.
That's absolutely right.
And as we think about these different waves and then moving through each wave, Sandra, a question for you. These early productivity gains that Antoine described more as table stakes are often the ones that make employees in organizations the most nervous. And then, as Antoine was talking about the subsequent value creation from innovation, employees can often get more excited about that. So, as you think about advising a CEO who's leading their company from the productivity wave to the innovation wave, how do you ensure that the organization comes along?
And that people remain engaged and excited about that longer-term future that Antoine might described, while they're also dealing with the short-term pain of restructuring roles in the overall organization?
Yeah, I think it's a great question, and part of it is being honest and visionary about what's coming. And while we call it waves, that also means thinking about what are some of the unlocks. I think overall, we're also in a world where, at least in, you know, developed worlds, where, you know, we've reached peak child. There, you know, we're losing, you know, what is it like, one of GDP growth a year, just because of an aging population.
So there is also a call for AI to help solve some of that labor shortage challenge we're, we're focused, we have in the developed world. So I don't think productivity is necessarily a bad thing. The big question is, as we deploy AI and technology in order to increase productivity, which by definition is good because it, you know, increases livelihoods. Can we redeploy human capacity quickly enough? And this is where the topic that Lari brought up earlier comes in, right?
This question of can we actually build the human capabilities for the future? Can we redeploy human capacity towards innovation and towards growth quickly enough in order for it not to become a conundrum?
Thank you, Sandra, and Tanguy, Antoine, Lari, anything you'd add there?
Well, I, Antoine, maybe you want to explain why, because I think it's very counterintuitive. Why is productivity going to create jobs versus what Lari was saying earlier, which is, you know, there will be a massive reduction in the size of the enterprises.
Yeah, yeah, again, going back to history in 1920, uh, Keynes, uh, predicted that thanks to productivity growth, by the year 2,000, we would be only working fifteen hours a week. And yet we are, and and we're working quite a bit more. And the the reason is that productivity leads to lower cost, uh, and that expands, uh, the demand frontier, and the elasticity of human demand is pretty high. Thanks to that, we've had increasing, uh, standard of living, and that further, uh, demand basically created jobs.
And that's this dynamic that is, you know, productivity gains and growth is fundamentally a positive force in the economy.
I, I actually wanna, I wanna challenge that a bit, like Antoine, uh, we. We wrote a book by MGI, where Hank is a director, called The Century of Plenty, meaning, uh, exactly, uh, great book, by the way, if anyone wants to read. Which exactly argues that when we take a very long horizon, we're getting to extremely positive societal and economic benefits, uh, due to development. I I think the thing to remember is that what we see and feel is the short to midterm.
The short to midterm means that what does the next decade look like? And, uh, I think beyond what I would call the utopian view on technology, which is that everything will be good, every disease will be cured and everybody will have everything. There might be a more difficult future ahead because development is not linear. And, uh, what I mean by that is that change is different on a societal level and on an individual level. And I think what we've struggled and what we keep struggling is that how does individual redeploy themselves?
So if you've been a call center agent for thirty years, are you the one who's gonna drive innovation? Maybe, but how do we help people get there? I think that's the question because we're seeing the breadth of this technology, meaning it's applied to every aspect of society at once. And never has such massive wave of technology adoption existed simultaneously. But on that, let me give an example. And I, I work with several organizations where we started applying gen AI three years ago, especially to development.
And most of my clients, so most of you heard from your developers that this technology does not add value, right? That's what we heard. They benchmark developers and what the actual productivity of the models applied to their work was key learning. Everybody expected that they have productivity increase when we, when they were observed, actually productivity declined by 2020 percent. Fast forward then to next year, when the same group was ben benchmarked this February from minus 20%, so them slowing down, we actually have gone to 20% productivity increase.
And this is the same group of developers. More importantly, though, which is, I think, is even a bigger finding, they had they had real trouble getting these same developers to participate in the. And the reason they had, uh, they had problems was that the developers did not want to participate in the study, right? They said that I don't wanna code anymore without AI. So we had done a one si one eighty in what less than one year, in nine months, basically, on what the benefits of the technology was.
I think the learning number one for every executive is that when your organization tells you that, hey, we tried this, it didn't work, I think we should all say that, hey, hold on, let's try again.
Thanks, Lari, that makes a lot of sense. But what explains that massive inflection point between the productivity losses the first time it was tested, and then the subsequent gains? What was driving the growth? Was it that the programmers got better at working with the models, or was something else going on?
Uh, y, yes and no, Sean. A. Actually, there's a real technical development that occurred last year. Let me cover it. So what we saw is that end of this year, over 20% of all code will be AI written. But this came only in the last six months. And the reason it, Sean, came to your question was that we simply got a new model, nothing else happened, we got a new model. And the way I like to explain it that three years ago, in 2023, the models could perform twenty seconds of work by a human developer, which was negligible, right?
It's not useful. Twenty four, we got to one to two minutes, twenty five. Last year we got to twenty minutes, twenty minutes to half an hour. This year, we're above ten hours. And last fall, we crossed about five hour task length threshold. Nobody actually knows, Sean, like, why that was the threshold. But it seems to be that the real world value view of using these models and having them as almost like a co-programmer to you became so valuable.
That even people who were saying that, hey, it hasn't been valuable. They suddenly said that, hey, I'm gonna use the model because it it really increases my output so much, and it's really useful in my day-to-day work. And that's the change. The key question we're now facing, and I think, which is a question for my colleagues and all of you, is that will this same phenomenon apply to our companies? Because what, what is right now happening, is that we're teaching the models how to be an accountant, we're teaching the models about corporate law, uh, et cetera, et cetera.
The list goes on how to do recruiting. So can we apply the models very soon, maybe already this year, to a lot of the processes in our organizations? And will that be a similar threshold as we had in programming? And Lari,
And so sorry, Tanguy, it looks like you wanted to say something here.
Yeah, I, I. I wanted to jump in because Lari said that there was only one driver for this significant change. I think there is a second one, coming back to what Antoine was saying, which is the system is adjusting. So Lari was talking about 20% gain in productivity. Organizations that are just deploying the tools are getting 20% productivity. That's wonderful, but organizations that are fundamentally changing their approach to software development life cycle or product development life cycle are able to increase increase their productivity by orders of magnitude above 20%.
We've seen organization move from sprints that would take two weeks to compress them all the way to just one day, where the agents are coding overnight and the requirement gathering testing happens by humans during the day. And we are even, uh, breaking those frontiers. So one of the thing that I I think is quite important to remember for the audience is the simple deployment of the technology might sound compelling, but you're gonna get 20% of the dollar of the value that you can capture.
The real value will require you fundamentally redesigning end to end workflows to capture the potential of that technology. And that redesign of workflow is complicated. Actually, most organizations have not codified those workflows. It's all in the head of the employees, and so it's a significant amount of time and effort. So I think, Lari, we, we, we certainly subscribe to the fact that the models are going to continue improve. The models we have today are based on chips that were designed two years ago and researched on four years ago.
So there is a lot of upside to the development of, uh, of the models. But the real question for the audience is, are you ready to take a blank sheet of paper and start redesigning your workflows to capture the value from that technology?
Sandra, I'd love to bring you in here. What? What's your perspective on the impact of AI on productivity and that lag involved in getting to the actual value creation?
I have so many things to say, including wanting to answer Lari's question on the transition phase and the transition costs. It's so interesting, cause that question on will there be more or less work is probably one of the most contested questions. And if you read like what right, like the leading academics say about that, you have everything from, you know, black to white as the answer. And I think the p the. The reason is that they make such fundamentally different assumptions about what will accelerate technological adoption and also how easy it will be to redeploy human capacity towards more value adding activity.
So these questions around what tasks are actually substitutable, right? Your question on can we translate from coding to other knowledge work? What happens to the cost of compute? Because at what point in time does it actually make more sense for a human to do that? And do we want Opex to be massively, uh, in, you know, one type of labor force versus another? But also, you know, what will governance, this government responses do to some of these questions?
I think there's so many uncertainties, which also means there's a lot of things for us to shape. The one thing that, through everything that has been said, has been very clearly no regret, is investing into human capabilities. Because no matter what scenario, no matter how fast or slow it goes, no matter what the government does, the the work we will be doing as humans will be fundamentally different three years from now than it is today. And no matter how quickly or slowly, and what industry or geography you're in, making sure your employees are AI fluent and you're actually not giving up on building expertise and institutional knowledge and judgment and system thinking hand in hand with technological adoption.
To me, that's like one of the no regrets that everyone should be focused on at the moment.
Sandra, given what you've just described, though, that that the nature of the work is going to change, is it possible that the work is actually going to become more stressful? Because some of the mundane tasks that maybe gave a worker a little bit of a mental break over the course of their day are now gonna go away. So now you're just full on all the time managing yourself and your agents. And could that possibly lead to that Kenzian world that Antoine described earlier, where someone working four hours a day can actually still be as productive as someone who maybe worked twelve hours a day before AI?
But then how do you, as a CEO, think about what that right balance is? So that you're not having people work at that accelerated pace at twelve hours a day and burning out your entire organization as you're adopting and rewiring for gen AI?
There, uh, a lot of, uh, questions in that question. I think everything's possible, right? Give me a crystal ball and let me tell you what the future holds. I don't know, right? I do what I sense in myself, right? I am not even digitally native. So to me, having not been a technologist and starting to learn how to work with tech, this technology to do my work better. It actually requires time, it requires interacting with the technology in ways that I wouldn't have done three or four years ago.
So to your question on, you know, what does it do to humans? I do think creating the space for that learning and learning, not as like formal learning, but learning by doing is really important. The other part to your question on burnout, I think the entire question around brain health. Not just for us in terms of the intensity of what's hitting us, but also for the next generation that's coming into the workforce. And are they prepared for all of this?
It's a really important question. And how do you actually create individual resilience and adaptability in people? How do you invest into that muscle of humans? I think it's another really important question. So I I think there's a a lot of, um, aspects to consider that are like the second and third bounce of the ball questions versus the. While we now have a technology, we can automate tasks we should. Because I think if you just to tangy your point, if you just automate existing processes with new technologies, you're just gonna make organizations wonderfully efficient at doing the wrong things.
And so making you can't reimagine and reinvent yourself while you're doing business as usual. You have to create the space to actually do that work with your organization.
And and Sandra, maybe coming back to something that Sean said earlier, there is the whole element of how you upscale your organization so they are ready to adopt the technology. But but Sean, you said, what's going to happen to the knowledge worker? And Sandra, you have perspective about the fact that organization have been designed around knowledge workers. If their value changes, what will happen to organization design?
Yeah, I love, I love the question, as you know, and I. I could be very visionary in terms of organizations being designed for the industrial era, in terms of like command and control and predictability, which is very far from the world we're in, where change is constant all the time. And we actually need to create an organizational system that is much more adaptable to the constant change that's coming. Which also means that if you think about the potential for AI to take friction and transaction costs out of a system, you can imagine organizational models where you are organizing your valuable human and technology resources and you're deploying them to solve valuable business problems.
And you're no longer constrained to where people sit and to rigid boxes and roles, because you have rapid decision making and learning cycles embedded in an organization. So this entire construct on designing for adaptability and scaling versus designing for stability will come at us now. Do I see many organizations who've done that at scale? Yet? No, I do see many organizations that are experimenting with operating model experiments, where they're taking, you know, capabilities from different, you know, knowledge domains into one team to solve a problem.
So we, we are seeing, um, more organizations do that, and those organizations that are able to do that and have more of a muscle of being more fluid and redeploying resources, they're actually having a much easier time of the value capture equation. I think
It, it's very interesting, Sandra. I wanna add one thing, so we're, we're, we're seeing two things that are happening. So one is that there's this definite power user phenomenon, power user being that it's about one of AI users in an organizations that actually consume and use AI way above others. And and what we found is that they're the ones who figure out new ways of working, they do something unexpected with the systems and tools. The second is that, to your point around, it's hard to find the space to do things.
Uh, I think what we started to think, with lack of better word, is it's almost like skunk working, a new organization, skunk working, meaning that instead of trying to change what you have and introducing change to existing teams. Of course, you enable those employees with tools and so forth, but you build a new and, uh, obviously, everything cannot be built a new. But a large organization can select one, two, three areas they invest in, put twenty people there, the power users, and have them actually try to build.
What would an AI native process or workflow look like? And I I think we live in an era that, given that uncertainty, every organization should have few significant skunk works that they try to rebuild. Because that introduces learning and also creates the narrative of change to an organization.
But I think it will be it will be incredible because changing an organization's structure is one of the hardest things to do. And so we'll have organization much flatter, an organization that will no longer have expertise in knowledge as the dominant axis. It will be much more about the ability for colleagues, human and agents, to work together towards outcome. And that pivot is one that I think many of our audience will need to reflect on. That will be one of the biggest challenge they'll face, and one of the critical enabler to their success.
And also in the process of getting there, the question will come up around what's your bottleneck, right? And one of them will be governance, building of trust, right? So there's a couple of things you have to think about as you build for scalability. That previously might have been an afterthought, that now actually have to become a major as you think about the design.
And Sandra, do you have any research you can share around how broadly people throughout an organization are going to be impacted? It's gonna pretty much touch everybody, right?
Yeah, and if you look at just the work done today without having done the reimagination that Tanguy talks about, that is absolutely necessary. What you see is thirty of current work could or should be economically more viably done by technology. This is including regulation and society and what technology can do, et cetera. But that thirty isn't thirty of today's jobs or occupations, that thirty is actually spread across everything, right? So if I take something like a manager, I think, uh, Harvard wrote that sixty of managers are today spending fifty of their time, or more than fifty of their time, on tasks that can be automated today, right?
So that impact is across everyone and across every level, unlike prior technology. So it's the speed Lari talked about, the scale enterprise-wide, and then the scope will depend very much on the type of work you do. But ultimately, what you end up seeing as you s you, you go up the value chain from AI capabilities is actually human capabilities will also move up the value chain. So a lot more people with those more judgment, system recognition, higher cognitive thinking, you don't need the lower cognitive skills as much.
Which then, to second bounce of the ball, brings up the question, how do you actually build that in the leaders of the future? Because you can't just stop hiring for you, you can't stop hiring high potentials into your organization, otherwise you're kind of constraining your future, too.
And and Sandra, just if I can just add one thing, Sean, maybe two. One is seventy five is really something that people should reflect on. Is that seventy five of the workforce? If you want to capture productivity gain, you're gonna need to change their jobs because there is not enough to automate for you to be able to remove them from the workforce. You need to reconfigure the work. I think I hope every C H R O is listening to us right now, because that is not something that our society has done.
What what is also very important, there will be tons and tons of new jobs that we cannot even imagine today. And I think you need to really start thinking about what those jobs are and what are the skill set that they w that those people will need, and where are you gonna source them from? I mean, Lari can talk about the type of role that we see to oversee agentic system in software development life cycle, but the skill set of those who design those systems is quite new.
Uh, so there is a massive, massive opportunity to identify those future roles and invest early in creating the training program and the development for those people to be successful.
And a related question to that is, how quickly can these transitions really happen? And as a CEO, how do you lead your organization in a way so that you can be on the winning side of this? Sandra, maybe you could just take us through. And Antoine referred to previously some of these general purpose technologies, but how do humans actually adapt? Personally, I'm more of an optimist. I think they adapt pretty quickly, and I'd love your take on that.
Yeah, and I think as we saw during Corona, right, when you have to, suddenly it becomes possible. Part of the challenge is not everyone's in software, right? So not every organization has understood yet how disruptive this is. And so the call for action on this actually being so fundamental a solve for the competitiveness of the future. I'm not sure every top team has really, like, you know, said that is true and put behind those words actions, including budget implications, et cetera, that in order to do that well, um, but I think if we have to, we do, we move, we redeploy quickly.
And there's organizations in tech who've, you know, redeployed 70% of their workforce within nine months because they had to and it was possible. Now, I don't wanna underplay that. There will be a responsibility that employers will have for certain populations that are deeply impacted, where that might not be possible. What I'm saying right now, but I don't think because there's that population, we should shy away from doing it where it is possible, where it is possible, we should do it quickly.
And that means, again, I don't think you teach AI, you do AI, like, what are the mechanisms? You make sure that every single employee in your organization does AI. That's the question that leaders should be thinking about right now.
And Sandra, just just one more point, because we we keep talking about employees as people who are employed. But many organizations have also a large dose of contractors. They might have, you know, use of Gccs. I think the whole employment ecosystem is going to evolve. If you drive enormous productivity gain in your enterprise and you want to protect the employment, one of the things you could do is start cutting the job that you have outsource of given vendors.
So there is a whole rearchitecture that goes way beyond the employee base that I think audience will need to think about.
Thank you both. And and this does sound like a massive change management issue. And as we alluded to earlier, Sandra, there's both the potential for employee fatigue with the constant change, and then also with the end result in in their ways of working. And maybe you could comment on both of those in terms of how do you navigate that successfully?
Yeah, I think we're also living in a world where a lot of is hitting us at the same time, right? So we're talking a lot about AI, but there's also other external factors, which, and you know, I talked about the pandemic just a minute ago, right? Like, we've, we've come through these waves of change, and so many people are exhausted and are, you know, constantly running, um, but in, you know, despite that employee fatigue, there's no question that we have to change more rapidly than we have in the past.
And so to me, that question of those organizations that I see who actually have built a change capability. I mentioned this earlier, who have built that learning muscle, who have turned into more of a learning organization, where that's the norm, there's practices around it. They're having an easier time, not just capturing value, but also bringing their employees along and redeploying employees towards new roles. That actually turn exciting. Because to your question earlier, I don't wanna be doing the manual, repetitive stuff either, right?
I'd rather be going out for a walk to integrate learnings. Um, so I think, uh, this question of what is your the system you build around change? Because just like AI isn't a bolt on, change isn't a bolt on either, it's actually a build in. So everything you do, actually thinking about how you build change into that reimagination, will become even more important for organizations going forward.
Super. And I can't let this AI focused discussion go by without asking for your perspective on what this means for our cognitive capabilities as humans, as we rely more and more on this new technology.
Sean, I I think, uh, there is so much embedded in your questions. One, you're gonna need to be role model of having a change mindset, you're gonna need to be willing to experiment and learn. Second, uh, Professor Agarwal from the University of Toronto, in his, uh, most recent book, made a very simple, um, representation of the fact that the machine right now is really, really good at prediction, very cheaply. Uh, and when that happens, the complementary goods, you know, increase in value, and there are two in particular relevant.
One is data. If you have proprietary data, the more valuable, but judgment, um, and so I think that, um, we as humans, will be facing a larger number of judgment related, um, uh, actions, uh, which relates back to your question on cognitive load. And then to credit one of my old professor at MIT, Professor Rigobon, you know, he started studying. What are the type of skill set that actually will be harder for the machine to replace?
And where he would encourage your society, uh, to focus on, uh, he has a framework, uh, called epoch about it. But a lot of it relates to empathy, networking, visioning, et cetera. And so what I, what I think will happen is the same way. We were addressing an audience of farmers in the nineteen twenties and we tell them your jobs are going to be automated, people would freak out. But the farmers of 1920 are the bankers of today.
Um, so I'll come back to. I think the skill set that are unique to humans remain, and I think the jobs of the future will be those that take more advantage of what makes us uniquely human.
I think humans remain humans, right? So we, we wanna be healthy, we wanna have a sense of purpose, we wanna have connection and belonging. And I think, uh, we'll see a rebound, right? So I think at some point we'll get to a place where there's a premium on human skills, there's a premium on a strong culture. So, you know, I think we need to get through this transition phase because of the uncertainty. But then I think a lot of this, uh, what stays the same for humans will still be true.
Thank you. And coming back to this concept of maintaining the three lenses, the technology, strategy and organization, as we conclude, I'd love it if each of you could very briefly share your one piece of advice for a CEO who's trying to make sure that this integrated approach is actually happening in their organization. Maybe Sandra, we'll start with you.
Don't wait on your no regrets and your learning moves.
Thank you, Sandra. That was indeed very brief. And Antoine, over to you.
Yes, there's this concept in strategy about being ambidextrous, being able to manage both kind of the uh, exploitation of the present and the exploration of the future. I think this is what it is about. It's being able to manage different timescale and all all of those three lenses are fundamental to shaping the future. You have no choice but to consider all three, uh, at the same time.
Uh, thank you, Antoine and Lari.
So I would have three reflections. One reflection is that when you think of your organization, all of your organization today is people. In the future, it might be one third people, two thirds of the cost base on these agents that the people use. I think it's very important, and the most strategic decision you can do as a CEO and as an executive, is to design what that technology, architecture, selections will look like. It's not a CIO topic, it's your topic as a CEO, because you're putting your organization very dependent on that technology.
Uh, I think then second to the discussion on change and how to change, you know, nobody can give definite answers at the moment. I think the most important thing is to do bold experiments. So, and I like to think of it in two lenses. One is that enable everybody, like Sandra said, like everybody tools, enable them, invest in their change. But at the same time, do very bold, almost like breakaway experiments, which we call skunk works.
And I think if you do that, you'll be successful.
Awesome, thank you, Lari and Tanguy. You get the last word.
I think it's bring the team together. Today, it's hard for me to differentiate the role of the strategist and the role of the technologist. It's hard to imagine tomorrow differentiating the role of the C H R O and the head of technology, because that workforce will be hybrid. So bring your team to learn together, that's how you get multiple lens in the room.
Awesome. Sandra, Lari, Antoine, Tanguy, thank you so much for taking the time with us today. This was a great conversation. Really enjoyed it.
Thank you, Sean.
And thank you to all of our listeners for joining us today on our three hundredth episode. We hope you enjoyed the conversation and we welcome your feedback and ideas for future podcasts. Just email us at itsr at mckinsey dot com, which stands for inside the strategy room. You can also share your ratings and reviews on any podcast player, with many thanks to all who've already done so. We really appreciate the comments and feedback we receive every week and encourage you to keep them coming.
And if you enjoyed this episode and you'd like to subscribe, you can easily follow our weekly series on any podcast player, where you can also access our entire library of 300 episodes. We also offer an inside the strategy room podcast collection page available at mckinsey dot com slash itsr, where you can easily browse our prior podcasts across nine major themes and access written transcripts of all of those conversations. And finally, if you'd like to automatically receive our latest publications and insights, we encourage you to sign up for email alerts on our insights page at mckinsey dot com slash scf, or connect with us on LinkedIn at the McKinsey Strategy and Corporate Finance Practice page.
Thanks again for listening. We look forward to having you join us again next week inside the strategy room.