Joe Fuller has spent much of his career trying to help more people find pathways into good jobs. He opened our conversation by telling me that he has changed his mind about the problem he is trying to solve.
For years, Joe believed employers had attractive jobs to fill but too few workers qualified for them. The answer was to prepare more people for the opportunities already available.
He no longer believes that.
Joe, a professor at Harvard Business School and co-leader of its Managing the Future of Work project, now thinks America has a demand problem. Our economy creates many low-paying jobs with little room for advancement and many high-paying jobs with daunting entry requirements. It does not create enough of the jobs that once connected the two—jobs that allowed people to support a household, develop expertise, and move up.
AI could deepen that divide.
Much of the public debate focuses on whether AI will create more jobs than it destroys. But aggregate job counts can obscure what happens to economic mobility. An economy can add jobs while making it harder to build a career.
AI is particularly good at the routine, rules-based work historically assigned to junior employees. The work gaining value requires judgment, adaptability, and what Joe calls “contextual intelligence.” Those capabilities usually come from experience. Yet companies are beginning to automate the entry-level tasks through which people once acquired that experience.
Joe points to a 15 percent decline in entry-level, white-collar job postings. He also sees AI-native startups operating with fewer junior employees and managers, while giving more responsibility to a smaller number of highly capable workers.
He expects that pressure to continue. Whenever AI encounters a valuable capability that appears uniquely human, that threshold becomes a signal to entrepreneurs and technologists. A protected category of work is also an unsolved, potentially lucrative technical problem.
This is why the two sides of the jobs debate may be talking past each other. AI could generate new work and new industries while removing the rungs people once used to reach them.
The question I carried out of our conversation was not simply how many jobs AI will create. It was whether the economy will create enough ways for ordinary people to grow into the jobs that remain.
- Allison
ALLISON: What’s something you’ve changed your mind about recently?
JOE: I’ve changed my mind about what I call the supply and demand problem. In studying this area for many years, I’d adopted the conventional view that there’s a strong supply of good jobs out there begging for applicants, and that labor’s focus should be developing pathways into those jobs. I believed that if we did that well, we could elevate incomes and let more people form economically secure households.
But increasingly, I’m convinced we have a demand problem instead. This economy is good at creating jobs with high skill thresholds—jobs that are difficult to get—and jobs with low skill thresholds that aren’t terribly productive, like retail clerks or fast food employees. Those jobs as currently defined will never pay well, nor do they build the skills people need to advance. Over time, we’ve exported or automated away many good-paying jobs, and lower-productivity, low-skill jobs have filled the gap.
ALLISON: We essentially have a barbell: low skill threshold jobs on one end, higher skill, higher capacity roles on the other. And in between, few jobs exist to support a pathway from one to the other.
JOE: Exactly. Look at income growth since 1980: purchasing power has risen between roughly the 10th and 60th income percentiles—the curve looks like an elephant’s hump. Then, like a trunk, it falls quickly from the 65th percentile to the 85th or 90th. Folks around the 85th percentile have seen negative growth in purchasing power. Meanwhile, it takes off like a rocket for the top 10% of earners.
There’s a reason for that: many of the good, middle income jobs were routinized manufacturing jobs, easy to automate or export. And it’s unlikely that those jobs will return, despite all the talk about reshoring.
ALLISON: What’s your view on where this trend is heading over the next five to ten years?
JOE: The first thing to acknowledge about these prognostications is that nobody—not me, not the US Chamber of Commerce, not the economists at Anthropic or OpenAI—knows for certain how this will play out. I tend to be bearish on workforce disruption, for two reasons.
One, this technology addresses a different part of the income spectrum than prior innovation cycles, as you’ve pointed out. Our barbell-shaped economy is sustained by the purchasing power of that top quartile, which is now susceptible to displacement by AI. When their purchasing power falls, they’ll provide less support for the other end of the barbell.
Two, there’s the issue of what I call market signal. As soon as you find a job insulated from AI, that’s a signal to entrepreneurs and technologists of an urgent, expensive problem. Solve it, and you too could have a billion dollars.
Many fine economists see no real erosion in employment levels, and thus conclude that AI is a big nothing burger for workforce participation. But you can see a 15% erosion in job postings for entry level, white collar workers. That’s not nothing—I’d call it a leading indicator. And the unemployment rate for Master’s degree holders under the age of 35 is 50% higher than the three-year average. That’s not huge, but again, if you’re walking in the dark (as we all are) and hear crunching under your feet, you may be walking on the unfortunate corpses of canaries in the coal mine.
Secondarily, it’s worth pointing out that 80% of venture capital invested in the United States in the last two and a half years is somehow related to agentic AI. Very few of those companies have scalable products in the market. These products are small language model-based, highly focused on specific industries, markets, or processes—and they’ll likely outperform frontier LLMs on things like performance management for financial services, for example, and do it more efficiently at lower cost. Those solutions are coming.
ALLISON: I understand why people hesitate to predict whether AI will increase or decrease the total number of entry-level jobs. But it seems much easier to predict that those jobs will be fundamentally different: AI will absorb much of the rote, routine, rules-based work that has traditionally defined them.
Why aren’t we more confident in that conclusion, even if the ultimate volume of entry-level jobs remains uncertain?
JOE: You’re right to put your finger on it, and you’re right that it’s also being ignored. I hear so many leaders across AI, hardware, and software repeat the refrain: “Don’t worry. The number of jobs and tasks being automated away is far less than predicted.” What’s really happening is that the rate of augmentation in many jobs is much higher than predicted. That means AI is making employees more productive—which means fewer of them are needed.
This is the first technology that gets better by itself, and it touches essentially all forms of work to some degree. I recently worked with Accenture Research on a model of skills displacement using ONET (Occupational Information Network)—the code the Department of Labor has followed for about 100 years (publication forthcoming). The average skills displacement that model generates is 41%. If AI is going to do 41% of the work people do now, you can’t look me in the eye and say that won’t lead to significant displacement of workers. Folks won’t be leaving their jobs to run pottery studios or write their Great American Novel, and also survive economically.
ALLISON: Let’s talk about organizational design—one of the most glaring omissions in the AI and future of work conversation.
People are quick to describe how AI can automate or augment a collection of tasks an individual or organization may do. But AI-native organizations are going a step further, redesigning entire teams with fewer junior workers and managers, broader individual responsibilities, and AI embedded throughout the operating model. They’re running second brains, making contextual intelligence accessible across the org. To what extent are operating models like these indicating the future of corporate work? How should we incorporate organizational redesign into this story?
JOE: To start at the end—yes, I think they’re indicative of how work will change long-term. My colleague, Rem Koenig, just put out a paper on how AI-native startups operate, and reached exactly your conclusion: compared to non-AI-native startups, these companies run smaller teams with more engineers and fewer entry-level and managerial staff. Yet, their valuations are comparable, meaning they get more value from fewer employees. The authors attribute that value to AI in both process and product.
The incumbents aren’t doing this, at least right now, because they’ve misidentified the task. Sixty percent of companies still describe their evaluation of AI as an issue of technology adoption. That’s the wrong way to think about it. The best incumbents I work with instead began by isolating main sequence processes—those absolutely core to the strategy and success of the company—and began experimenting with AI in parallel. They ran A/B tests, refined how they used AI in these key processes, and gradually weaned the company off the old process, adjusting head counts and job descriptions accordingly. This, by the way, is how you get the so-called J curve effect—when you’re running parallel processes, you’re not getting early efficiencies, but you’re also not leaving the positive economies of automation on the table.
Too many companies approach AI like any other SaaS tool, trying to figure out how to bolt it onto the process as currently configured. That’s one reason you may get disappointing results with your AI experiment. But poor data is also a reason for that disappointment, as is employee training. Our data shows people are 50% more likely to use AI at home than at work, and they’re likely using free models. Without more training, it’s like asking someone who drives a five-year-old Prius to suddenly get behind the wheel of a Ferrari in a Formula One race.
ALLISON: Right. I published an article last Monday with Zapier’s Chief People Officer, Brandon Summat, on their AI fluency framework. In my opinion, it’s the best anywhere in the world. They practice your advice closely—entry-level employees are expected to use AI to support individual work. Adoptive employees use AI to orchestrate the team’s work. Zapier increasingly expects its leaders to operate at the transformative level, using AI to re-engineer how work happens. It’s a fundamentally different, relatively scarce skill set, which is why we don’t see more organizations doing it.
But I also think part of the challenge is imagination—seeing a new way of doing a hard thing well. Can you spark some imagination on how you’ve seen AI transform work?
JOE: Sure. Let’s take consumer goods marketing as an example. I can start using AI to create and test social media collateral for a campaign in a fraction of the time it used to take. But what’s more transformative is how I can now use AI to create a large synthetic data set, based on my actual data, to run higher-fidelity experiments on that collateral. With a data set that size, you’ll find new ways of segmenting customers or served populations, and begin to pinpoint new preferences—ones that will radically alter your traditional segments.
I also think it’s helpful for folks to realize that AI transformation is most advantageous to the company with the most market share. They have the most to lose if competitors with smaller shares fully lean into AI, but they also, however, have the most to win from their own experimentation. Economics favors incumbents. If they start using AI more effectively than the company with 30% market share—because they have more and better data—that advantage becomes insurmountable. It’s just math.
ALLISON: Judgment, curiosity, adaptability, problem solving—we’ve long understood their importance at work. How do you expect hiring to change—or not change—as employers like Zapier continue to expect these skills on day one?
JOE: Recruiting is just like any other process: if we use AI to augment what we’re currently doing, we’re missing the point.
AI is now accessible to both sides of the recruitment transaction. AI is reading more resumes and screening more candidates, but it’s also being used to create phony portfolios and apply en masse to jobs that barely fit candidates’ qualifications. The average number of applicants for a decent-paying job went from 72 in 2024 to 300 this year—I don’t think that’s a coincidence.
In response, I think we’ll see companies doing more live interviews, focused on proving out skills in real time. This is how many hiring processes have often worked—academia does this now; many creatives get jobs through real-time skills tests; and consulting firms have long used case studies. But the experience will be quite different: I’m imagining an employee working through five different simulations, using any model they want, for which they won’t be able to prepare beforehand. I ran a consulting firm before I was a professor—that’s the kind of case study setup that could reliably indicate higher rates of advancement and lower turnover in the role.
ALLISON: Are you suggesting that because of new assessment modalities like simulations and experiential projects, we may actually start anchoring hiring on durable metacognitive skills?
JOE: I think it will be more plausible. But education needs to pivot dramatically too: we need to teach students contextual intelligence—how to understand concepts in action—and they need more experiential learning, since that’s where that intelligence is developed.
So what should kids study in college—and what should the new core curriculum look like? Personally, I think this looks like fewer highly specialized courses, even things like BC Calculus, and more courses that teach students how to think, reason, problem-solve, and work with data.
Statistics. Logic. Rhetoric. Theories of mind. College needs to teach people how to decompose a problem and theorize and test multiple solutions—that imperative requires experiential learning and competency-based assessment. Developing those skills also benefits from a breadth of study. Philosophy teaches us how to think. We learn empathy by reading novels like Catcher in the Rye and To Killing a Mockingbird. The advent of AI and its impact on work is more likely to be the salvation of humanities and social studies than their ruination.
We should also acknowledge that significant reforms are required across the K-12 system. Many of its deficiencies are well known and reflected in America’s declining importance relative to other countries in math, writing and technical problem solving. But others—like the lack of focus on cultivating students’ social skills—are ignored. Social skills will be essential in the workplace of tomorrow. They are largely formed in childhood and adolescence. Changing pedagogical methods in K-12 to include more experiential learning and team activities would help deepen the development of social skills, as would far more exercises requiring spontaneous written and oral communications.
ALLISON: What’s your call to action for educators on the importance of experiential learning?
JOE: I’d simply say that higher education needs to heed what employers are asking for, as a matter of survival. For forty years, the cost of college has grown faster than the rate of inflation—that can’t continue forever. Enrollment declines are putting the entire sector on the defense. Meanwhile, we’re experiencing a ubiquitous crisis in standards that may make preparing a future workforce in these skills almost insurmountable. American students are entering college performing at the worst rate in the last half century. AI could help bridge that gap at the K-12 level, if designed right—but not without substantial reform across K-12. Grade inflation is making the value of a degree almost worthless. And, most unsettling, students are loath to be held responsible for their learning, because they haven’t been before. All of these problems will compound the workforce problem.
The US does always have an incredible ace in the hole—we can turn on the spigot and import talent educated in a more rigorous system. We have been squandering that advantage in recent year in ways that may diminish America’s long term attractiveness to skilled workers. But that’s not a solution that will make the lives of American citizens objectively better.
In short, Allison: I don’t see the level of alarm in higher ed that would actually drive change at the extent needed. So much of higher education’s capacity over the next decade will go toward managing cash flow, rather than innovation. And we are seeing quite a lot of data suggesting that higher ed, in its current form, is at risk of passing its sell-by date.
Allison: So, you’re a glass-half-full guy?
JOE: Ha.
ALLISON: Sometimes I leave these weekly conversations thinking, “The world is about to get crazy, but I’m optimistic we can navigate it.” Other times I leave thinking, “The world’s about to get crazy, and every one of our institutions is ill-equipped to support people through it.” I’m feeling less optimistic today. How do you feel? What’s your level of hope, optimism, fear?
JOE: I do think AI will have significant positive effects on society. But I feel less optimistic about this moment, for reasons beyond AI. The pillars of our modern world have been knocked over by the people who created them. We’re reckoning with our institutions right now, from higher education to government, from corporations to the media. That disorder in the broader context makes it less likely we’ll respond well to the changes AI brings.
ALLISON: It certainly reduces our capacity for change. When you can’t see a clear path forward, you need to make sense of the ground shifting beneath your feet, on one hand, and metabolize what you’ve learned into society, on the other. You have to get better and faster at both. I think AI radically accelerates our ability to do sense making. But the second piece is so much harder because of what you’re describing; we’ve lost the institutional loops that help metabolize learning into action, with any accountability or speed. That’s where my pessimism lies.
JOE: That’s well said, and I endorse it. It also invokes a theme we discussed earlier: the general lack of self-awareness among decision makers. The influential voices in this space are at best whistling past the graveyard, at worst deliberately obfuscating what’s happening to advance their self-interest. And there’s no institutional governance ushering in this technology, as there was with the internet. For good or ill, DARPA, the Department of Defense, and the US intelligence community were deeply involved in gating and launching the internet; 17 universities sat as the major nodes. ICANN spun out of that.
At this point, AI is driven exclusively by capitalism. There’s a reason we regulate capitalist activity—it’s dangerous to society to let something live purely by the interests of individuals and the allocation of rents. We’ll continue to see political movements like the Democratic Socialists of America, a group largely driven by downwardly mobile educated people. Revolutions are caused by downward mobility and a significant number of young males with no reasonable prospect of household formation. Both forces are active in the world today.
ALLISON: What’s one small signal to which we should be paying more attention?
Joe: HB8 reforms—not such a small signal, to be frank. That’s the state legislation dictating that state-supported postsecondary institutions will now be evaluated at a program level based on the economic outcomes that accrue to graduates. Funding will be cut or eliminated for programs that consistently issue credentials without labor market value.
We’re beginning to see a massive shift to non-credit programs in community colleges, because for-credit programs are being shut down, and because earnings data is now accessible to more learners.
Listen, people: Higher education needs to accept that the vast majority of students don’t go to college to get educated. They go to get a credential that will get them a job and an independent lifestyle. Educators don’t want to be held accountable for that outcome. They say: “we are here to hone the minds of the future, not fill a corporation of worker bees,” but that’s exactly what we’re for. We need to claim it, and be accountable to it.
One more signal comes to mind: People should assume that AI will eventually score a 95 on every test—so you’ll need some skill on which you can score a 98. And we’ll also need the contextual intelligence to put these tools to productive use.
Here’s why: Augmentation will still lead to pretty severe labor market disruption—one of my forthcoming papers, out next month, overlays AI’s capabilities onto the job descriptions of six industries, and reveals a few things we haven’t seen before.
One: org structures become much more fluid, because AI changes what’s sought from workers, and what they need from superiors and subordinates. The point of org structures is to create a durable operating model that survives uncertainty and disruption; AI puts the disruption inside the four walls of the organization for the first time.
Two: there’s a significant redistribution of talent across the hierarchy, including at the entry level. That’s why we’re seeing comparable unemployment for call center reps and software engineers under 30 in those two quite different roles—both are highly addressable by generative AI. Companies are keeping the workers with deep contextual intelligence, and not hiring the people who lack it.
ALLISON: We’ve recently built what people sometimes call a “second brain” at Humanist: a shared, searchable memory of the company. It brings together the knowledge we generate across documents, applications, Slack messages, voice notes, and call recordings, so both people and AI can draw on the full context of our work. What happens to the value of organizational knowledge when AI can access, connect, and build on that context over time?
JOE: Augmentation will creep for people whose contextual intelligence sustains their employment. This illustrates the market signal issue I described: every time AI hits a “uniquely human” ceiling, that’s a signal to the market that there is remaining value to unlock.
ALLISON: Second brains are a great example of how AI will transform work unpredictably. They weren’t on our radar in April, when I first heard you talk about contextual intelligence—three months later, we’ve rolled one out in our firm, and I’m astonished by how quickly it’s changing how we work. This technology can redesign how work gets done. But that’s an incredibly hard thing to predict before the technology is invented.
JOE: Entirely right. And it’s well beyond the ability of most companies to manage the breadth of the change that is suggested.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.