This is the McKinsey Podcast, where we help you make sense out of our world's toughest business challenges.
Welcome to the show.
I'm Lucia Raheli.
And I'm Roberta Fasaro.
This idea that you could have a digital coworker isn't a science fiction fantasy.
It's something that you can apply.
That's McKinsey's senior partner, Lorena Yee.
She's talking about the rise of agentic AI and the new reality of working side by side with AI agents.
In just a minute you'll hear more from her, along with fellow McKinsey leaders Roger Roberts and Sven Smit, as we dig into McKinsey's annual tech trends report.
Yes, AI is front and center, but there's a lot more in there too.
But before we jump in...
A quick heads up on what's new at McKinsey.com.
We've got a fresh article where seven McKinsey leaders share their advice for CEOs who are facing unpredictable global trade rules.
And speaking of CEOs, our new book, A CEO for All Seasons, drops October 7th.
Lorena, welcome back to the podcast.
Thank you, Lucia.
I am so thrilled to be here.
Lorena, let's start with some basics.
Could you just differentiate for us agentic AI from the kinds of AI that have come before it?
It's been this evolution that we've been through, and I'm not sure everyone understands what the difference is between the different kinds of AI.
We've been on an AI journey for 50, 60 or more years.
What a lot of us have been already using in business is analytical AI.
And that is predictive.
We also saw over the last 5, 10 years a complete bloom in terms of machine learning.
Generative AI is moving from predictive to a probabilistic model.
It's a different type of thing and it's based on an LLM or a large language model.
And this year, what people have spent a lot of time talking about is agentic.
And that's a different thing, because that's actually taking more of an autonomous ability to complete a task or an action.
So that's saying that you could have a machine take an action, something as simple as changing your password to something as complex as actually working with humans across multiple steps.
One of the things I think is really important is oftentimes when we see the biggest unlock for businesses, they're utilizing multiple things.
They're using machine learning.
They're using analytical predictive AI.
They're using these probabilistic models and capabilities, as well as this wonderful UI where you type a question into a generative AI model.
And they're using agents, say, for their call centers to reply to level one, level two questions.
So help us bring that piece of it to life a little bit.
Give us an example of the way you see AI, put it into practice at work, and how it's actually having impact.
Well, sometimes when you're in a candy store, it's hard to pick which sweet to taste.
I think in the headlines and in the actual deeper research we see, agentic AI is on everyone's mind.
So for example, You can actually use agents to rethink your research process.
And you can say, how do I have agents take on different tasks collectively over a process?
Or maybe something even more straightforward like loan processing.
Are there pieces of stages of loan processing, analytics and approval and risk scoring that actually agents could do?
When you think about call centers and customer care, Instead of actually asking and being told what are the best directions to do even something very simple like change my password or link to my credit card, you could actually have an agent fulfill that task.
And I think that really opens up the sort of aperture of how you can apply this to business, which is why you see so many businesses so excited to start to stand up agents and starting to talk about things like what if I actually had?
10% of my function or my team as digital coworkers?
How can I start to reimagine or even just drive some basic efficiency in how work gets done?
And putting science fiction aside, how do you see AI meaningfully changing the relationship between humans and machines or humans and technology, in the workplace or otherwise?
Would you talk about that? humans and agents working side by side.
This idea that you could have a digital co-worker isn't a science fiction fantasy.
It's something that you can apply.
Now this is where human and leadership judgment become.
Really important is we have to say, how do we apply that power in a safe ethical, thoughtful way?
And by the way, you know, it's not software deployment.
It actually is bringing something into the workforce, so it actually looks more like human capital deployment.
So let me give you an example.
If you have an agent or a team of agents, You actually do have to do all the technical pieces of it get it ready to enter the workforce.
But then you need to onboard it.
You need to do training of it.
And you need to give it feedback.
And then you have to figure out how to teach it.
What are the ways of working?
What are the norms?
How to actually be kind of within our team.
And so we're just seeing the beginning scratches of this.
This is starting to actually think differently about our workforce.
And to your point, it feels like 2024 was the year that Gen AI dominated the headlines.
And now... agentic is in the spotlight.
Are you seeing anything interesting that's starting to illustrate the value creation thesis for agentic AI.
Absolutely.
One of the things is We should consider why something is trendy, rather than that it's just breaking the headlines.
The reason why it's trendy is it is a profoundly exciting capability.
Some simple examples I've heard.
One company shared that they have used it in their sales process, enabling sellers.
They still have the same amount of sellers, but they actually saw 11 increase in their lead gen and conversion.
That's something very concrete.
Another company on the retail side shared with us how they're using it to advise on shopping decisions and personalization.
So it's a mix of different capabilities, kind of all put together in an ability to say and, by the way, we can have this shipped and delivered to you or complete the thought between what you may want to purchase, what your purchasing history looks like and what you have.
So there are all types of early ideas that are getting at helping with the traditional business challenges and objectives that companies have and using these technologies to achieve those.
How are leaders navigating adoption and ensuring that their employees have the skills necessary to get the benefit of AI at scale?
I think people are trying all different things.
I think one thing that's incredibly humbling right now is we're all learning together.
Some of the things that tend to be working is have you really reimagined your business and how you can work AI native?
Reimagining is giving yourself the space to say if I reimagine my business with all these capabilities, how radically could it change?
Where could I see value that I could not have imagined otherwise?
Separate decision is how much of that you implement.
But I think the first thing is to just look further out.
Then I think the next piece is, How do I get experience in using the tools and capabilities?
And so have I given my team access to the technology?
Am I myself using it?
Are we learning as we go?
Am I talking to my peers and sharing the early sets of solutions that I'm putting in place and why they're working, how they're working, how they're not?
And then I think there are kind of certain laws or norms of business that don't go out of fashion.
We need business cases.
We need to prioritize areas that are going to make a difference and make sure they're resourced, make sure that they have all those basics that we know matter, so that we can see them to their full fruition.
And so it's great to experiment, but at some point you do need to place a couple of bets and really invest big against those.
What's your take on how to get a data-informed, really clear-eyed view of what the risks are and, where relevant, to just start trying to manage those risks successfully?
I do think, very pragmatically, that we have to have our eyes and ears open to say what are some of the unintended consequences or some of the risks that we introduce by using the technology in new ways, in sometimes incredibly intimate ways, with people.
Because we're working side by side.
If you said you have a digital colleague that's actually entering a workspace.
That's very different from a back end tool, so to speak.
And so I do think we need the courage to look at what could go wrong.
But I don't think that takes us all the way to a dystopia of the world.
The world will break down.
So I tend to think that we can be quite pragmatic about this.
I think your question also raises the real uncertainty card here, which is time.
Is this six months, 12 months time? five years, 10 years?
This is actually an incredibly hard question.
We do not know.
And so we're constantly, at least at McKinsey, we're looking at the signals for how companies are implementing this to have some kind of indication.
Any advice for leaders who are tasked with embarking on this transformation that is filled with so much uncertainty?
How should leaders get started thinking about this change?
I think it's a gift if you're given that.
I think as leaders, we have to ask more questions and listen more than state and assert.
We do not have... great certainty and that's okay.
Actually, that becomes part of what we do, which is how do we walk through that?
How do we find the value?
If I don't leave space for the uncertainty and ask the questions, I might miss something.
And so I think that that becomes the way we walk forward.
And I don't think it needs to be a bad thing.
I think it actually is a great thing.
We have to remind ourselves that we decide how fast we go.
Because this is all about us as humans.
And so if we don't want the technology to take 70% of the tasks, you don't have to.
The technology is not asking you to.
The technology can be used however you want.
Lorena, thanks so much for joining us today.
Thank you, Lucia.
This has been so great.
Sven, welcome back.
Great to see you as always.
Hi, Lucia.
It's great to be back.
So let's talk about robotics and the rise of autonomous systems, which is just so interesting and futuristic sci-fi.
Let's start with some context.
Why is robotics a top trend in the research this year?
For many, many reasons.
I personally play a little game, which is Spot the Robot.
And I think many of us have seen cleaning robots.
Very few of us have been in a logistics center and seen robots climb to pick up stuff.
And at the moment some of the logistics centers are hiring or installing more robots than they're hiring people.
And so you see this tipping point.
We then also see the massive amounts of video clips of humanoid robots which are coming, which will have their own version of full self-driving, which is full self-walking, and they're learning stuff and so on.
But if I had to take one that has fascinated me personally, on Spot the Robot in the airport where I go very often Schiphol and Amsterdam there's now a wheelchair.
It actually has you sit down and it drives you to the gate.
So they are coming.
And the question is just how much, how many, when.
The research also showed adoption of robotics accelerating in sectors that I found somewhat unexpected, like energy and utilities.
What does that look like in practice?
And what does it mean in terms of the direction in which we're heading?
In energy, it's around installing equipment, which will require humanoid robots, for example.
If you take almost every bit of manual labor that has some repeat to, it could over time be robotized one way or the other.
I think the complicated stuff of refurbishing an old house electricity wiring or something in energy will take still some time.
But there are standardized jobs in these sectors as well.
That you would find to be the early automation process.
How much of robotics adoption is about replacing human capital, as you just described?
And how much of it will be about augmenting or changing what it is that we do?
I would say that's almost a choice whether robots are going to replace the work or they're going to add to work.
What you know is if the robots are cheap enough, which some of the estimates suggest is that they will make certain things cheaper and therefore the demand to it will go up.
Classically, this has been framed as the stonemason problem.
We could replace maybe half the stonemasons by robotic stonemasons and then you replace half the workforce.
Or you say well no no, no.
The construction of housing and offices will get so much cheaper that, besides every current stonemason, there will be a humanoid robot or some other form of robotics that helps them build bigger, better and more houses for more people.
And affordable housing is a huge problem.
It looks like the robots are going to be cheap enough to make certain jobs more affordable.
And so instead of replacing half the stonemasons, you get every stonemason its own robot.
In an ideal world, I would even say every stonemason owns his own robot.
And basically you get a twofer or a threefer because the robot works 24-7 or something.
But, you know, business models will be changing by the day.
People say full self-driving is a robot too.
Does it replace all the Uber drivers, taxi drivers, and so on?
Maybe.
But it could also bring in a whole new cast of jobs, which is running 10-car fleets, keeping them clean, making sure they're in the right position, monitoring the behavior of people and maybe doing logistics jobs, package delivery and so on.
All of a sudden, the demand goes up.
So I think we always underestimate the demand equation when things get automated and become a lot cheaper.
And in the stonemason example, is this what we mean by cobots?
What are cobots and how should we expect those to begin to appear in the workplace?
Yeah, I think cobots, the words and language will develop over time.
But The positive frame is basically all automation, whether it's robots or AI, by the way is a multiplication of human creativity, ingenuity and production capacity, which I think it largely could be and has always been the case.
If you think about farming, which went, of course, from 100 of the workforce to 2, basically The farmers went more productive based on what was the automation of the day, which was a tractor and a couple of other things.
I think I just see it more like that.
And then what is no longer needed in that profession might go do something else again.
So Sven, what is the first priority leaders should focus on now when it comes to robotics at a high level, acknowledging that the response there is obviously sector dependent.
I think in general in AI. and with that also robotics, the fastest learner will win.
The idea that you know how it exactly will work is, I think, arrogant.
But if you don't participate, you don't know where it's going.
And you also don't know when to scale, when to go deep versus still shallow and learn.
So to me, as companies are going process by process, what can be done by AI?
You need to do the same thing process by process.
What can be done by robots?
When? and when is the right time to experiment and when is the right time to scale.
But at the end, I think if you're not learning in this space, you will be late to catch.
Now let's turn to another of this year's trends, the energy transition.
Demand for energy is obviously rising with AI a big driver there.
The need to power data centers has been much in the news.
Tell us, why is the energy transition a top tech trend this year?
I think we were in a frame which was we need to replace the old energy with new energy, and that was the transition.
Now we're in a frame we actually need more energy.
And that's because of AI, but maybe also because of additional wealth for the people and so on.
So if you're in a world of more energy, you almost need energy beyond oil and gas, just because we need more.
And then we want it to be reliable, which is what will be the mix.
That's also why you see the discussion about nuclear rising, because that's the clean form of additional energy.
You were early on that.
The fundamentals, I think, dictate that it will be part of the mix.
But if you are in a frame of, let's say, we need two times energy, we already need one time non-oil and gas energy just to get the addition which will have to come from something else solar wind fusion, etc.
Fission and so on.
The shortage of electricity is even faster.
So everybody's looking for what's the next source of at-scale, stable and reliable electricity, which will be a mix of new and old technologies that need to scale very fast.
Which elements of the energy transition are the closest, in your view, to being poised for rapid commercial growth?
I would say it's all driven by cost.
And so we see that the fastest single growth category at the moment is solar.
And wind comes on that side of renewable second.
I think the discussion on nuclear is accelerating.
People are taking it more serious.
Then it goes to what are the storage technologies and so on, the battery discussion.
I think that would be sort of the sequence.
While we're still saying today's delivery also will have to be the oil and gas that we have.
And the research points to scaling challenges beyond the technology itself, including supply chain, including infrastructure needs.
Where do you see the most significant obstacles to scale?
It's a multitude.
If you look at the build rate that's currently happening in China, we're talking mega installations every week, almost.
And we're not at that rate outside China anymore. almost anywhere.
And so we need to go into this build rate that's much, much faster.
So you see people restarting old plants again.
So that's just on the power station side.
In addition, Because it's electricity, you need to build more grid.
You then have to have more storage or other ways to maintain reliability.
And it's an investment project that requires massive, massive build-outs that looked like really one of the traditional infrastructure build-outs that we've had.
You're kind of a famous optimist, or I think of you that way.
But how optimistic are you that we'll reach a more sustainable, more resilient energy transition in a reasonable time horizon?
I think we have accelerated our learning that energy makes the world go round.
And once humans decide to build, we'll get it.
So, when the race was really only the discussion, replace oil and gas, which is, in a way, a positive frame for climate, but it's a negative frame for the old world.
You get resistance all over the place.
If you say well, we need two times energy and we need a fast and AI tomorrow and electricity tomorrow, you go into a built mindset.
And so my optimism is we are starting to move to the idea that the world needs a major construction project and energy.
Mindset matters a lot.
Yeah, the reframing matters.
Okay, what is the first priority leaders should focus on when it comes to the energy transition?
Again, at a high level.
I think we have two priorities.
One is the build speed, which has a supply chain issue that we just discussed in the energy build-out and transition.
But the second thing that's very, very important is to understand that we don't build it too expensively, because affordable energy is what will drive progress, whether it's AI or just human prosperity and so on.
You can prioritize stuff that's very expensive and you can prioritize stuff that is a little less expensive and has chances to get cheaper.
And I would hope that we're building stuff that maybe philosophically, are in the money in the midterm while maybe they might need some subsidies in the short term to get done but that we don't build stuff that's structurally out of the money versus low and affordable energy prices.
And so, to me, the priority of build-out versus affordability while you seek clean energy is very, very important.
So you need to solve reliability and security on top.
How do you see the current geopolitical volatility affecting speed to scale?
Well, if anything, the geopolitics are driving the energy debate to the fore.
More importantly, people want to have their independent source of energy as much as they have dependent sources of energy.
So you see a massive push for independent source build-out while at the same time securing some of the sources.
So the geopolitics plays into the security of energy supply, but they also play into the affordability of energy supply.
Because you don't have Affordable energy, you don't win the AI war, which doesn't make you win the security war.
Sven, thanks so much for joining us.
Thank you.
It was great to be with you.
Roger, welcome to the podcast.
Well, thanks.
It's always great to be here.
Let's talk about one of this year's trends that may seem, on its face, a little less hot than, for example, agentic AI, but that undergirds much of what AI makes possible, and that is application specific semiconductors.
And, to be clear, application specific semiconductors are purpose-built chips that are optimized to perform specialized tasks and that also offer great speed, energy efficiency and performance.
I would love your best quick and dirty on why these chips matter and why they're growing in importance for business.
The growth of graphics processing units, or GPUs, started out as simply the graphics accelerators for those PCs that helped you play those real-time games.
And now it turned out that they were just the right thing, just the right processing architecture to allow for AI computing at very high rates of speed.
That has led to a world where, as AI computing for both training of models and inference, meaning the use of that model to generate outputs has created incredible demand for semiconductors.
Now we're seeing a ongoing level of innovation in this space that moves from thinking about these GPUs as only quote general purpose way of delivering AI computing to more application specifics.
So that can mean hey, I'm going to have particular types of semiconductors that might accelerate my muddling of biological processes or the exploration of molecular performance inside a cell.
And what that kind of simulation can do unlocks lots of new possibilities for therapies.
And so when we start to think about application-specific chips and look ahead, we get excited about the possibilities here that can come from accelerating these very particular uses or workloads.
Interesting.
And how do you see developments like data center expansion affecting chip development and specifically disruptive innovations that might reduce the power intensiveness that AI requires?
Well, there's several things going on there.
One is the chips themselves are maybe a thousand X better over the past several years at pushing more computing through a piece of chip real estate, let's say.
Right.
Very dense modules and racks and those racks placed inside a data center.
That means I'm getting more output per unit of square feet or capacity or volume inside that data center.
And what that does is it generates a ton of heat.
And so we start to get down to the physics of how do I move that heat off that chip quickly?
And that means that we're starting to see people really try to innovate in cooling technology moving from air cooling just moving the heat away with airflow to water-based or even more esoteric fluids that are used to circulate around the chip and move that heat away.
And what that can do is allow us to get denser and denser.
So Then, are innovations at the model layer that allow the model to use the chip capacity more efficiently.
And we might see another thousand X there in terms of the software efficiency and just using chips, the right data at the right time, or exercising parts of a model's architecture, not all of it, in order to deliver a great result.
And so all of those things taken together, I think, are in fact bending the curve on power demand.
Roger, are there any consequences we should be aware of given this power demand?
The vast demand for processing power.
More and more tokens, more and more generative outputs from AI usage will fight against that and drive our power consumption up.
And so fundamentally we are going to see significant growth in power demand, significant pressure on both electrical generation capacity and our ability to move that through transmission to data centers.
What's an example of how these chips might be used to improve performance or cost efficiency?
Well, one example could be in robotics, or sometimes people now talk about it as embodied AI agents.
And so that can take many different forms, but ultimately, the Integration of multiple kinds of AI computing on that robot means that I'm in some sense needing to bring all the human senses together.
I've got elements that are helping me with language understanding perhaps, but also with vision and navigation and being able to control myself inside my physical environment in an effective way.
And what that then does means is that you might have different application-specific AI chips inside that robot that allow for these various functions to come together and be integrated as a whole.
What's the outlook vis-a-vis talent in the semiconductor industry?
What does the supply and demand ratio look like as far as talent goes?
Well, semiconductors are constrained by many things, right?
They're constrained by the capacity of fabrication, our ability to sometimes to move them around the world in a supply chain.
And they are also constrained by the talent of people, not just to design them, but also to transition them to at-scale manufacturing.
And so there's only a few companies in the world that can design the equipment or equipment who have the experience in really scaling up manufacturing to the levels and quality and performance that are required.
Now, a lot of those people today are concentrated in a few companies and in a few countries.
And so that does definitely become a very precious resource.
Yeah, particularly if more chip manufacturing moves to the United States.
There will have to be some kind of capability building funnel, presumably.
For sure.
In the US as we lost some of our chip manufacturing to overseas markets and providers, we've also had an erosion in that pragmatic practical, hands-on talent right.
We have wonderful design capabilities, but the scaled manufacturing requires really trained people.
And so, as manufacturers transition operations to the US and build more and more world-class fabs here, there's going to need to be a lot of upskilling of the human capital that supports that, from people working right there on the front lines to those who are designing and shaping the processes and practices inside those fabs.
Now let's turn to another of these trends, digital trust and cybersecurity.
Give us the high level on this trend and what's at stake as we advance it to the AI economy.
How does it affect, for example, customer choice or stakeholder support, etc.?
?
So fundamentally, we see trust as critical to accelerating the path to AI adoption and impact.
And not something that you do just because a regulator tells you to do it, but you do it because you want to create the most impact, the most adoption possible and get your innovations out there into the hands of the world.
Is the rise in geopolitical volatility significant as a risk factor here, or do you view that as a relatively constant risk?
I think all the players in the world that are interested in AI adoption are going to care about trust and they're going to have to.
We don't have to necessarily have trust among AI in order for companies to care about trust.
So I'll start there.
Now, would it be better if we had more transnational collaboration and cooperation so that companies could see a more consistent set of standards around the world, a consistent set of expectations?
Sure that would allow for companies to operate with a common and simpler set of rules everywhere.
But we haven't seen that happen in other domains.
And yet we've made a lot of digital progress.
So I am still optimistic that economic forces will help push us in the right direction here.
And, more broadly, where are businesses in terms of adoption on the relevant technologies necessary to enable digital trust, and what kinds of challenges might they face if they're not as far along as they should be in this area?
Well, there's so much to do because the landscape keeps evolving.
So, if we think about the world of agentic AI and particularly agentic commerce, if I'm running a website, Digital entities showing up on my site attempting to inspect my products and prices used to be referred to as malware bots.
Right now, they could be someone whose agent is simply seeking to make a potential purchase.
So we're going to have to be able to differentiate between good traffic and bad traffic, good bots and bad bots.
And that is going to be.
That means those bots are going to have to bring authentication forms of tokens that allow for that site to know.
Yeah, I'm going to let you in.
I'm going to let you behave like a human on my site.
I'm going to let you act on behalf of a human on my site.
And the human has to trust that agent's behavior to operate on that site within the intentions and guardrails that they've set for that agent's behavior.
Interesting.
And is talent an issue here as well?
What does the talent picture look like in this area?
Well, talent in this area is definitely going to be a challenge and a constraint.
We're going to have to train up lots of next generation e-commerce talent to understand what's it take to create talent, a good agent experience and not just a great CX on my site.
If I'm welcoming commerce agents in, I've got to actually help them traverse and navigate the site.
Now the good news is that they can do that in human-like ways today and doesn't require an entire revamp of your architecture.
On the other hand, it doesn't mean that what's best...
For today's model of human browsing will be the best for agentic commerce and driving sales through a sales funnel and closing transactions.
Fascinating.
Roger, thanks so much for joining us.
Thanks for having me, Lucia.
Thanks so much for listening to the McKinsey Podcast.
I'm Lucia Rahilly.
And I'm Roberta Fasaro.
Find us on McKinsey.com.
We'll have a transcript of this episode up shortly.
And download the McKinsey Insights app, where you can find this podcast and other helpful content updated daily.
If you enjoyed the show we'd love for you to leave a rating and a review.
We'll see you in two weeks.
