Is the AI productivity story at a turning point?

‘Is my generation doomed?’ That’s what one Stanford graduating senior asked Professor Erik Brynjolfsson, economist and director of the Stanford Digital Economy Lab. The professor had seen the data, so he didn’t dismiss the question. But as productivity gains begin to emerge, Brynjolfsson argues that the biggest opportunity lies not in replacing workers but in redesigning work. In this episode of McKinsey Talks Talent, Brynjolfsson joins McKinsey leaders Brooke Weddle and Bryan Hancock, as well as Global Editorial Director Lucia Rahilly, to discuss what’s changing, who’s most susceptible to displacement, and what leaders need to know to create value while managing new risks.

To watch the full-length video of this interview, visit McKinsey’s YouTube channel.

The following transcript has been edited for clarity and length.

Early signs of a changing trajectory

Lucia Rahilly: Let’s start with some context. At McKinsey, we often discuss the “AI paradox”—the gap between the massive investment that organizations are making into AI and meaningful, at-scale returns. Your research suggests we may finally be turning a corner on AI and broader productivity gains. Talk to us about where we stand.

Erik Brynjolfsson: To realize the full benefits of these powerful technologies, organizations need process changes, workforce reskilling, and sometimes new products and services. All that takes time. The hard work isn’t just deploying an LLM [large language model]—it’s making those intangible investments. They don’t instantly translate into output, but once they’re in place, gains begin to accelerate.

We look at productivity in terms of the J-curve, and we’re seeing early signs of productivity returns in national productivity statistics. It’s still early, but we’re seeing this change at the micro level for specific companies. When AI is rolled out effectively, the impact can be significant. It’s not widespread across the economy yet, but I’m sure it will be in time.

Brooke Weddle: Seventy to 80 percent of our work with organizations isn’t on what should be automated or the tech enablement part of their workflow. It’s defining new roles, behaviors, and required skills. We also examine the underlying mindsets that enable success. Is that what you mean?

Erik Brynjolfsson: Absolutely. You have to get granular. One way to get more visibility into a specific role is to break it down into tasks. Every occupation is a bundle of tasks, and AI can typically help with some, like writing a memo or analyzing documents, but not others, like lifting a box. Every occupation we looked at has some parts where AI is going to be more effective than others, which means managers need to reconfigure, reorganize, and rebundle those tasks. That requires a lot of creativity.

When cost cutting is the wrong reflex

Brooke Weddle: You’ve mentioned elsewhere that when we think about the agenda for AI and AI innovation, we should really try to hone AI’s capabilities for tasks humans are not good at. As you think about the work you’ve done on automation potential, how do you consider that point?

Erik Brynjolfsson: Many technologists are obsessed with making AI that imitates humans. Alan Turing said the ultimate test is making an AI that is indistinguishable from a human. That’s a cool concept but a terrible business strategy. Instead, AI should focus on what it does best. That makes AI a complement rather than a substitute.

The first instinct of most managers I talk to is to have AI replace the workforce. Not long ago, I was talking to a CFO who said, “We really need to measure the ROI of AI better.” I agreed wholeheartedly. She followed with, “And therefore, we’re going to go through each department and see how much head count reduction they’re getting from AI.”

I reminded her that this is only one, very narrow measure. But the cost mentality is pervasive because, let’s face it, that’s the easiest thing to measure, and people already think of AI as replacing people. The bigger upside is to get AI to allow those people to execute new things outside their typical job tasks.

Bryan Hancock: Where are you seeing that kind of augmentation in practice?

Erik Brynjolfsson: Coding is a strong example. AI can both substitute and augment, often within the same role. We see AI agents doing some routine tasks, but senior developers’ roles are amplified. They’re able to do so much more. They’re working with fleets of agents, not just one or two. They define problems, orchestrate agents, and evaluate outputs—a three-part model: define, execute, evaluate. There’s a reconfiguration or rebalancing of what people do. Ultimately, it can lead to creating similar code, but the really exciting thing is when it enables you to do new things you’ve never done before.

New metrics for a new context

Brooke Weddle: Many of the companies I’m working with are trying to figure out how to approach performance management in this new context. What do they really want to assess? Are there new behaviors? New skills?

Erik Brynjolfsson: Totally. The companies that are most successful are the ones that measure well. This is really the frontier that we’re leading right now—to find better ways to measure output. There is an old saying that goes, “You can’t manage what you don’t measure,” and we need a new set of metrics.

Brooke Weddle: What metrics would you prioritize?

Erik Brynjolfsson: We want to boost productivity, and the way economists define productivity is output per input. Too many people focus only on input. You must be creative about developing better measures of output—better quality, better customer service, new products, lower employee turnover. Most companies underestimate how much administrative data they already have. Ultimately, you start being able to identify various fine-grained metrics, and you can see how those metrics move as the company rolls out the technology.

A related part is getting causal inference from these metrics, because you don’t want only correlations. You have to see causal relationships, which means using different techniques—different instrumental variables that economists have been using for a long time. There’s been a credibility revolution happening, and one of my missions is to bring that credibility to businesses and business data.

Bryan Hancock: How have your expectations of clients evolved?

Erik Brynjolfsson: We’re using clients’ own data and doing more of a bottom-up assessment. Most organizations show a power law distribution—top performers deliver 10 to 20 times more. If you can codify and scale what those top performers do, it’s an easy win to level people up.

Brooke Weddle: Many leading organizations encouraged experimentation early on. Now they’re asking how to scale it in a more structured, value-driven way.

Erik Brynjolfsson: Absolutely. That’s such an important point. While it’s great to capture this knowledge, you also want to direct it somewhere. You’ve got to have the bottom-up and the top-down direction established so you understand what’s possible but also make sure it’s aligned with meaningful outcomes.

The rise of tokenomics

Lucia Rahilly: You’re talking about accelerating productivity and progress on the J-curve trajectory. Agentification could obviously make a significant difference there. But you did some interesting research on the costs of AI agents, with the surprising result that most humans are really bad at estimating how token-intensive and costly certain tasks can be relative to others. Talk to us about that.

Erik Brynjolfsson: Yes, there’s a big disconnect there. Just a few months ago, people were “tokenmaxing” without worrying about token cost. Now it’s becoming a first-order concern. Companies are spending tens of millions of dollars on tokens. The problem is you don’t know how much token cost goes into a task. Agents themselves are very bad at estimating.

This leaves a big gap, because how can you start a project without knowing what it’s going to cost? Or suppose you get halfway into a task and the agent doesn’t finish the job? There’s an agenda for doing more work to understand, predict, and manage token and agent costs up front so AI projects don’t stall or overrun budgets halfway through.

‘Is my generation doomed?’

Lucia Rahilly: You and many others, McKinsey included, have called this the most consequential transformation since the Industrial Revolution. Are we doing enough to prepare for a change of this magnitude?

Erik Brynjolfsson: We are not ready. I’m impressed by the capability improvements and how rapidly they’re happening, especially in Silicon Valley. People are spending hundreds of billions of dollars to push those capabilities to be even faster.

At the same time, we do not have a solid grasp on what’s going to happen to employment, productivity, wealth, income, wages, the centralization or decentralization of power, or even meaning. There’s very little being invested in understanding it, and that is the biggest challenge for the next ten years or so. We must figure out how to close that gap between these capabilities and our economic understanding.

I worry that we’re going into a very disruptive period. I mentioned that I’m optimistic about productivity, but I’m worried about inequality and centralization of power. It will probably be very disruptive for a lot of people, and we’re not ready for that.

Bryan Hancock: It’s also one of the things I worry about. When people talk about augmenting, they’re often talking about “front office” tasks—for example, in sales. When they talk about automating, they often mean “back office.” If you start to layer on the questions “Who’s in the front office? Who’s in the back office?” it no longer looks like a wave of automation coming for people with master’s degrees or advanced capabilities. It looks like a different disruption pattern. What are you seeing in the research?

Erik Brynjolfsson: It seems like it’s cutting across lots of different groups now. I was stunned when a student came to my office a month ago. She was a graduating senior from Stanford who didn’t have a job lined up, and neither did her friends. She asked, “Is my generation doomed?”

That’s a hard outlook to have. I see that angst. People are really worried. I tried to give her some hope. I think that, yes, there are a lot of jobs being destroyed—I don’t want to sugarcoat that. But there’s also all this opportunity being created. You need to possess much more agency and aggressiveness to lean into that. No one is going to tell you what your job is. You have to be the creator.

At my final class for my Stanford students, I told them, “When you hear AI, you should not think ‘artificial intelligence.’ You should think ‘amplifying intention.’ If you have intention, AI will allow you to do a lot more than you could have before.” I understand that’s not what the conversation is about, but we need to change the conversation and remind people that this really could amplify whatever intentions they have.

The data on displacement

Lucia Rahilly: Your research also showed that early-career professionals are being disproportionately affected.

Erik Brynjolfsson: Yes, we did find that conclusion. The groups experiencing shrinking employment were early-career, entry-level workers aged 22 to 26 in the most exposed occupations. You could rank all the occupations—about 700 of them—by how exposed they were. If you take the top 150 most exposed occupations, the top quintile, what you find is that the young workers in those categories had about a 13 percent drop in employment when we first did the study.

We’ve been collecting data every month since then. That drop is 16 or 17 percent now. The effect has been growing over time. That’s why when that student approached me about her troubles, I didn’t dismiss it. I saw that in the data.

Lucia Rahilly: Were there any implications for wages?

Erik Brynjolfsson: We worked with our partner, ADP, which is the world’s largest payroll processor, to answer this question. They gave us access to data so we could analyze it carefully. We saw shifts in employment, with some roles shrinking and others growing, depending on age and AI exposure. We did not see a big change in wages compared with employment.

Lucia Rahilly: Is it only a matter of time before more senior-level employees are also susceptible to these kinds of displacements?

Erik Brynjolfsson: That’s a big concern. That’s a possible future, but another is that we lean in and show senior-level workers how to use these technologies to augment what they do, discover new products and services, and expand the market.

One of the things I really stress when talking to senior executives, policymakers, and economists is that this is a design problem. Let’s figure out the right incentives, the right structures, and the right intangible investments so we can end up on the winning side. We should not be passive.

The future of work is a design choice

Brooke Weddle: I want to loop back to understanding how to better prepare for an AI future. What are your hypotheses on a potential solution or set of solutions? Are we solving a collective-action problem? An incentive problem?

Erik Brynjolfsson: Every company will have to reinvent itself. Part of it is a collective-action problem. What we’re seeing is that the pyramid in many companies, where people come in and work their way up, is becoming more of a diamond shape. That flat base isn’t there anymore. That saves some costs in the short run, but it also poses the question of where middle managers or middle-skilled people will come from if there aren’t people in entry-level positions.

That’s a real challenge. Part of the solution is having public investment in education and supporting training. Another part of it—and enlightened companies are already leaning into this—is acknowledging that AI itself can be a great teacher. You can get a lot of personalized education via AI.

McKinsey Talks Talent Podcast

Bryan Hancock, Brooke Weddle, and other talent experts help you navigate a fast-changing landscape and prepare for the future of work by making talent a competitive advantage.

See more episodes

Brooke Weddle: I think a lot of organizations are a little stuck there and want to know immediately how the pyramid should evolve. It’s a choice, just like any operating model or organizational design. There’s no one right answer. You’ve got to think through how value is created across the organization and what that means for what humans do versus what agents do. Once you have a view of that, you can say, “Here’s the talent system to support it.” That’s a process, and many organizations risk moving straight to efficiency versus really thinking it through.

Erik Brynjolfsson: Going straight to efficiency, honestly, is a little lazy. The easy way is to cut costs. It’s not hard to measure. But it’s not as sustainable as finding new sources of value and investing in your workforce. You’re going to get more resistance from your workforce if you tell them, “This is a tool for cutting heads.”

I really think there’s a much bigger opportunity to find ways of creating new sources of value. It’s what happened with earlier technologies. All through history, wages and employment grew even as new technologies came along, and we want to keep that record going.

Bryan Hancock: On design, what I hear you saying is we’ve got to be thoughtful at different levels. There’s the societal level—how we’re investing in education and training and how we’re creating the right signaling mechanisms so that relevant skills gained in one place can be picked up by another employer. But there’s also company-level design.

To use an analogy, electricity is another general-purpose technology. It’s not just about putting electric lights in the factory. It’s about electrifying the entire factory, which then requires you to rethink the way the factory works. That’s a workflow-level design. Making everything work requires societal-, company-, and workflow-level design.

Erik Brynjolfsson: We’ve got work to do at all those different levels. At the level of specific companies, you need to understand the required tasks so that you can restructure. Those two go together. I think of tasks as the atomic unit, and once you understand where AI can affect each of those, you’re in a better position to carry out the required reinvention.

Moving from experimentation to scale

Brooke Weddle: Companies are under pressure to use AI to drive efficiencies. But reimagining a workflow or an enterprise is not something most leaders know how to do. One of the things I’m working on with a lot of organizations is creating a playbook for doing that and doing it consistently. What’s that methodology? I don’t think we know yet.

Erik Brynjolfsson: No. It’s an inherently harder problem. It takes more creativity. You can look at what already exists and say, “How are we going to automate that?” But to imagine something new—that’s harder.

My view is that it’s riskier to not try to implement automation. If you focus on what you’re currently doing and try to hang on, that’s riskier in the long run. Ultimately, no company, no country, and no person has ever succeeded by focusing on the same thing over time. You need to expand your horizons. That’s what’s helped make America successful, and successful companies do that as well.

Bryan Hancock: One area that’s exciting for AI is small businesses, because they’re now able to access talent and capabilities they wouldn’t have been able to access before. For example, in the past, a physician’s office might have had to sell to a private equity practice to have the back office taken care of. Now there are AI tools that can help them keep and grow their business.

Erik Brynjolfsson: This is exactly what we talk about in my masterclass. You’ve all heard about the “one-person unicorn,” and maybe that’s a bit of an exaggeration—the one person with a billion-dollar business. It’s not that much of an exaggeration to have a small group of people who can now use AI to leverage so much more.

Lucia Rahilly: Do you see companies successfully building AI-native ventures within their larger organizations to innovate and test, and then integrating them?

Bryan Hancock: One of my clients described this as engine one and engine two. Engine one is the current business. Engine two is where they invest in disruption—where their new AI-native product or offering will live.

Some companies are saying, “I’m uncomfortable with that setup. What does it imply for people who work in engine one as we’re gearing up engine two? Does everybody in engine one know that they are in engine one, and that it’s on the ‘wrong’ track? Do we need instead to invest behind the lead innovators across the business and then scale up?”

I think the jury’s a bit out among clients on which way they will be able to build up AI, because engine two gets more tangible output faster. However, is that going to be the ultimate answer that gets you to scale? That’s the question our clients are wrestling with.

Brooke Weddle: I have one large global banking client that took an experimental approach, and some of those experiments became famous within the organization for the problems they were solving. They have one use case that is an internal “go and see” for how to execute AI transformation well and use it as a force for augmentation. And they’re using that information as an input into this learning and scaling model that we’ve been working with them on. It’s exciting and a good example of a more nascent approach to implementing AI.

The long view on expertise

Lucia Rahilly: We’ve been talking about augmentation and ways to push novel thinking to accelerate productivity while keeping humans engaged and employed. There is also a lot of discussion about information collapse if AI ultimately provides too much ballast for humans in the workplace. How do you assess that risk?

Erik Brynjolfsson: There is a concern that AI is generating more and more content on the web, in books, and everywhere else. One thing we must do is make sure we have incentives for human creators to continue to create value so that LLMs can continue to become better. Right now, existing copyright law doesn’t necessarily reward content creators.

You want to strike a balance. If you give content creators too much incentive, then the downstream people lack the incentive to use AI tools as effectively. On the other hand, if you don’t give them enough support, the content isn’t created. So moving forward, we need to rebalance the way we reward creation versus the use of different knowledge types to have a thriving ecosystem.

Bryan Hancock: We’re seeing that happening live in professional-services organizations. We can acknowledge that you are an expert. Your expertise is what generates client demand and momentum. If you’re not appropriately incentivized to share that knowledge and put your information into the firm’s systems, everything collapses. We’re seeing a lot of client interest in setting up knowledge management systems to make sure they get the best outcomes.

Erik Brynjolfsson: You’ve got smart clients, because I’ve also seen the opposite approach, and it’s really destructive. In our call-center paper, we found that less skilled workers got the biggest boost from LLMs and began performing almost as well as the most skilled workers.

Some of the folks using these call centers spoke with me and said, “Now we don’t have to hire as many of those most skilled workers anymore, because the less skilled workers are doing almost as well.” That’s a very short-sighted approach.

Exactly to your point, where does that knowledge come from in the first place? It came from the most skilled workers. In some ways, they’re even more valuable now, because they’re answering a question for their own client, as well as replicating that answer throughout the organization.

You want to make sure you invest in highly skilled workers who are creating valuable content. That’s how you upskill the rest of your workforce. You need your company to be sufficiently forward looking to understand that vision. Reward people not just for what they’re doing today but also for how their knowledge creates value in the future.

Know who your mavericks are

Lucia Rahilly: Earlier in the conversation, we talked about your CFO client who reflexively turned to head count reduction. And as Brooke mentioned, not all leaders have the capabilities to reenvision processes end to end and reap the benefits of intangibles. Are you seeing many leaders build those capabilities successfully?

Erik Brynjolfsson: It’s really hard. There aren’t many. We’ve been working with Nasdaq, and I think that’s one of the companies doing a great job. They’ve had great leadership who have been reinventing at the organization level for a long time, even before LLMs came along. They’re continuing that track record. Most companies are struggling with making that change, which is understandable. People underestimate how difficult it is.

Brooke Weddle: Many of these workflows that need to be reenvisioned are cross-disciplinary. And in my experience, when you gather a cross-disciplinary group in a room, everyone becomes a little defensive. They might say, “My part of the workflow works great, so we don’t need to reinvent that.” There’s a lot of learning and openness required to arrive at the reimagined end of a proposed plan. I think leaders can set the conditions for that openness in a way that allows for bold, challenging thinking.

Erik Brynjolfsson: CEOs have to step up and play a role in enacting that change because the rest of their management is not selected for that big transition. They already have something that’s working and want to preserve it, so they need a jolt to the system. You also need the openness that you talked about. Acknowledge that by saying, “You can do this. It’s going to be okay. We don’t expect it to go totally smoothly. Let’s be bold.”

Brooke Weddle: I was with a group of business executives, including CEOs and military leaders, discussing some of the behaviors required for change. One thing that stood out is that you should be looking for the mavericks in your organization. A maverick can exist at multiple levels but is going to challenge, be bold, and not be defensive.

A question that might be worth asking a CEO or a CHRO is, “Do you have a good sense of who your mavericks are? Are you strategically allocating them to drive this transformation?”

Erik Brynjolfsson: I like that. This is a time for mavericks.

Lucia Rahilly: Before we close, do you want to answer the $64,000 question plaguing parents from shore to shore—namely, what skills our kids will need in the future?

Erik Brynjolfsson: I focus less on specific skills and more on a broader framework, which goes back to what we said earlier about having intention and figuring out what you really want to do. I think that’s a skill that can be engendered, even taught. We can teach kids to be open to being the person who directs the project or who comes up with new ideas. If your kids have that positive attitude, they’re going to be much more successful in the future.

I think almost everyone is going to be managing not just one agent, but a fleet of agents. They’ll be the CEO of their own little entity, and they’ll have to possess those leadership skills while directing their agents. The ones who are good at pointing them in the right direction and evaluating them are going to thrive.

Bryan Hancock: So from a policy standpoint, should we invest more in things that enable entrepreneurs?

Erik Brynjolfsson: I think so, yes. One of the things we’ve seen over the past couple of decades is that technologies have advanced, but productivity hasn’t grown all that much. Part of the reason is that there’s less dynamism in the economy. Even though we see a lot of it here in Silicon Valley, there are fewer start-ups overall in the United States today. If we can encourage more entrepreneurship and more dynamism, then I think these technologies will have a bigger beneficial effect.