Whose Jobs Does AI Really Replace?
Matt Sigelman, Shad Ahmed, Joe Fuller, Michael Horn, and I on the past, present, and future of AI and work.
Last month, I sat on a panel at ASU/GSV called “Whose Jobs Does AI Really Replace?”
Michael Horn, who has a rare talent for making a ballroom feel like an actual conversation, guided us through the question from three angles: what history tells us, what the labor market is showing now, and what each of us thinks may come next.
The group was unusually well-matched to the topic. Matt Sigelman, president of the Burning Glass Institute, brought the labor-market evidence: what job postings and skills data can tell us. Shad Ahmed of Anthropic brought a view from inside one of the companies building the models, with data on how AI is spreading across tasks and occupations. Joe Fuller of Harvard Business School brought the employer lens: how companies actually redesign work, and why most of them are still treating AI too much like software and not enough like a new operating model.
The conversation kept moving past the simple replacement question. AI may change work less by deleting whole occupations and more by changing the shape of jobs and organizations. Fewer managers. Smaller teams. More leverage for people with judgment, or what Joe called “contextual intelligence.” Less room for routine cognitive work. And maybe most importantly, a much shakier bridge from school into a good first job.
The full conversation can be listened to here. Edited transcript below.
Enjoy,
Allison
Past - What We Know From History
MICHAEL: Let’s start with the past. Matt, I’ll start with you. My mentor — and Joe’s former colleague — Clay Christensen used to say that the really inconvenient thing about data is that God only created it to describe the distant past. I figured we’d ground ourselves there, because AI has been around for several years and we’re starting to know some things. Talk to us about tech revolutions in general, and how what we’ve seen in the early innings about what AI is doing in the labor market.
MATT: In the spirit of the Ghost of Christmas Past, I’ll rattle my chains and take advantage of my age here. Let’s rewind the tape to 1998, for those of us who can remember that far back, when the Internet was a new thing. It was the greatest mania since the Dutch Tulip craze — you’d look at the predictions being made at the time and be right to be skeptical. But if you fast-forwarded to today, you’d find the hype was understated. We’ve found that about 90% of workers are in jobs that make significant use of the Internet. That’s probably not what most of us would have expected in 1998.
However, you’d also find that those changes took much longer than anticipated. Some jobs are only now going away — meter readers started disappearing in the last five to seven years as connected devices became more prevalent. Parking attendants? There’s an app for that.
You’d also find the changes were less predictable than expected. There were 65,000 job openings last year for social media managers — a category that didn’t emerge until twelve years after 1998. And most of us back then, except maybe the thwarted CEO of Webvan, wouldn’t have anticipated that the biggest category of Internet-driven job creation would be delivery drivers and warehouse workers.
Past tech revolutions did carve away some jobs, and there are reasons AI may be different — I’m sure we’ll get to those. But — with apologies to my former travel agent — most of what we’ve seen is that past tech revolutions changed how we work rather than whether we work. That’s what we’re seeing so far; very few jobs are showing signs of going away.
That doesn’t mean work won’t change. We are seeing meaningful shifts in job definitions and the skills employers seek, in ways that align closely to the interplay between automation and augmentation as they rewrite work.
MICHAEL: Let’s have Shad weigh in too; Anthropic recently put out a interesting report on historical tech revolutions and AI’s current coverage versus its theoretical capabilities. Where do you agree with Matt, and where do you disagree? Joe and Allison, jump in as well.
SHAD: We agree that jobs are a package of tasks, and those tasks are reformulating as we speak. Talking about job disappearance is challenging. We’re monitoring how tasks are being refashioned as AI use increases. We’ll see some jobs shift and change, some new ones emerge, and some may disappear — but that’s a process we all have to watch closely, especially in education, where we’re trying to monitor what’s happening in the labor market.
Where we might diverge — and you didn’t say this explicitly — is on speed. Based on our data, AI diffusion is moving much faster. We’ll probably reach 80% national AI use in the US within seven to eight years. With the Internet, it took almost twice as long. So we’re seeing a rapid uptake.
Secondly, our research is showing that AI is being broadly applied across many jobs, but not yet deeply applied across most jobs. About 50% of jobs are using AI on roughly 25% of tasks. We see depth only in very few areas — about 4% of jobs, like software engineering, where AI is being used across 75% of tasks. We are in the very early stages of understanding how automation will play out.
Finally, productivity gains are starting to accrue among those who use AI most. Our last report showed that six months of continuous AI use is producing outsized ROI at the individual level, and we’re beginning to see it at the firm level as well. We’re early, and as models have rapidly improved, those gains are starting to hit firms and individuals. We’ll see how they play out.
JOE: I’d like to build on that. It’s very dangerous to extrapolate from the current data for a couple of reasons.
One: The vast majority of companies are doing nothing that remotely resembles organized AI training. Our data — published quarterly at Harvard in partnership with Vanderbilt — shows that the average person is almost twice as likely to use AI at home as at work. And if you’re using it at home, which version are you using? The free one. The advanced subscription tools are Ferraris compared to the ten-year-old Priuses of the free versions. We don’t yet know what people will produce using sophisticated tools with adequate training.
Two: Bizarrely, 80% of white-collar employees expect their supervisor to teach them how to use AI, while 90% of supervisors expect their white-collar subordinates to come to them with AI-driven recommendations and ideas. There’s a football bouncing along the ground and no one’s falling on it.
Some companies are tackling this — I know some of them, and I’m seeing the productivity impacts — not so much jobs going away as jobs getting reformulated, but far more efficiently. Total positions in those emerging roles will likely shrink. New roles may emerge to offset some of that, but I’m seeing the Anthropic spider diagram already playing out in the better companies. They are pushing to the efficient frontier much faster, and that will have a profound impact on their economics.
ALLISON: When you look at how technology has changed work over the last ten or fifteen years, it’s not just at the task and job level — it’s at how organizations work and function. Think about our reliance on email, Slack, shared knowledge bases, the ability to hold live conversations with anyone in the world through Zoom. It’s changed how we design organizations. The modern product and engineering team looks nothing like its equivalent from a decade ago.
One of the biggest waves of change around jobs will be how organizations choose to redesign themselves — not just at the task or job level, but at the organizational level. You can already see this in early-stage companies. How do startups organize themselves? Increasingly, it looks fundamentally different than even a year ago, with teams being smaller, flatter, and individuals taking on much larger spans of control as they leverage AI as specialists.
You can see early change in bigger companies too, with Jack Dorsey at Block and what people have started calling “Dorsey mode.” The basic idea is that AI does not just sit inside the org as a productivity tool. It starts to absorb the context that middle managers used to carry: what teams are doing, what decisions have been made, what tradeoffs matter, what work needs to move next. Once that happens, you can start redesigning the company around fewer coordinators and more builders.
That is a much bigger shift than “AI makes everyone faster.” It changes the shape of the organization. It changes which jobs exist, which jobs disappear, and what it means to be valuable at work. I think this may end up being one of the most lasting effects of AI on jobs, and weirdly, it is still not a big enough part of the public conversation.
MICHAEL: Allison, your point echoes the analogy of electricity — initially, there were no productivity gains when subbing it in for steam. But then factories redesigned their organizational model around it, and manufacturing took a quantum leap forward.
Present - What We See Happening Now
MICHAEL: My wife and I hosted a dinner party recently — six people from very different professions — and every single person had a completely different lived experience with AI. It speaks to your point, Joe, about how some companies are really leaning in and seeing those productivity gains. My sense is that within sectors, different companies are approaching AI in completely different ways. I’d love to get some of that nuance, name some names, get into specific sectors, and unpack why the data varies.
JOE: Even today, over 60% of executives describe deploying AI as a technical task. That’s a complete misread of the situation. AI is about managing work and organizing processes. Someone who bolts it onto existing processes like a SaaS package may get incremental efficiency gains, but they’re missing the point. This is a general-purpose technology. You have to build your processes around what it does — just as Dorsey is doing. (I’ll note I’ve never previously invoked him as a model of management, so there’s a first for everything.) But the vast majority of companies are not prepared for that.
The large companies making good progress have isolated a small number — usually no more than three — of what I’d call main-sequence processes: those fundamental to their competitive position. They’ve effectively followed a model not unlike Apple’s in revolutionary product development: isolate the process where you’ll deploy AI and insulate it from the rest. This is what drives the J-curve dynamic, where companies run parallel processes until confident they have their arms around the new one.
You’ll see companies — JPMorgan Chase, Wells Fargo, Procter & Gamble, Coca-Cola — pushing hard in those main-sequence areas: risk analysis for banks, marketing and promotion for consumer packaged goods. The economics there, which I know well enough to say with confidence, are very compelling.
There’s another issue: if you’re the biggest bank in the United States or the biggest beverage company in the world, and you are standing on the accelerator and learning faster than your competition, you will reach a point where it’s mathematically impossible for others to close the gap. The opportunity for early leaders to gain meaningful competitive advantage in these markets is very real. That raises interesting questions about industrial policy and antitrust law, but that’s a future conversation.
SHAD: To build on Joe’s point — among our customers, who are among the earliest adopters and many of whom are at the frontier — there’s a genuine choice to be made. And this is why we’re still early enough to shape some of those choices. Market forces and quarterly shareholder pressure will drive a certain set of outcomes; there are also alternatives worth exploring.
We are seeing many companies take the productivity gains from AI and invest them back into growth, innovation, and their people. Walmart has been very public about trying to do just that. Others go straight to the bottom line. Those are discrete choices, sometimes within the same industry.
In some industries there’s less room for growth — you’re competing on margin and cost — and you’ll see more efficiency gains and jobs taken out. In others, you can do more with the people you have.
Anthropic is trying to understand what distinguishes companies getting outsized returns. Anecdotally, we see them creating safer spaces for experimentation, running strategic top-down AI projects, and thinking about how the technology can help employees work smarter, make processes faster, and make products better. The organizations outperforming the others tend to be intentional about two of those three things: getting the technology into the hands of their employees, and being deliberate about the specific processes or products to which they’re applying AI.
What you do with the resulting productivity gains is the next question. We see it at Anthropic every day. If you can do more with the people you have, it creates more opportunity. Even our AI-driven organization isn’t yet structured to deal with that massive productivity influx, but there’s a universe of opportunities and backlogs, especially in engineering, that teams are finally getting to that they couldn’t have reached before. There are some interesting effects we’ll start to see play out, particularly in growth industries.
ALLISON: I’ve been trying to be deliberate about sharing where I’m changing my mind, because that’s where the most interesting insights are. Something Shad said made me think about a view I’ve started to revise just in the last month — watching what’s unfolding at Block was part of the inspiration.
My mental model has been this: the companies that win are the ones that use AI for productivity gains, then take the extra people and reinvest their talents in the next horizon of growth, doing things they fundamentally couldn’t have imagined before. The headcount container would stay roughly in equilibrium as we go through the next chapter or two.
I’ve started to change my mind on that. Block’s recent layoff was significant — around 40%, and in some parts of the organization, 70%. Ethan Mollick was quick to argue that’s a losing strategy: take that extra headcount and invest it in doing something different, driving growth, innovation, and creativity — which is roughly where I was until recently.
But working with early-stage companies, I’m seeing something different: you often move slower with more people. There’s a lot of shared context you have to establish with a human team to do something innovative quickly — meetings, communication, alignment. With fewer people and tremendous AI leverage, you can in some cases accomplish far more than a team twice the size.
So I’ve really started to question whether even for companies treating AI as an opportunity for growth — not just productivity — the headcount container will stay stable. It may continue to shrink, and fairly rapidly. I’ve updated my mental model on this just in the last few weeks.
JOE: You’re right.
ALLISON: He says I’m right.
MATT: There’s a lot of attractiveness to the extreme version of what Allison is describing — the mythology around the single-founder company — and we’ve all been hearing about it. But it’s a bit early to write off the advantages of scale, not just in terms of assets and data, but people. Joe pointed specifically to the importance of using this as a moment to rethink not just work, but how we organize work. And that’s what will make this a moment of greater productivity.
Burning Glass does a lot of work measuring which parts of jobs are being automated away and which are being augmented. What we’ve been seeing is that rather than some jobs being characterized by a lot of automation and others by a lot of augmentation, they’re happening in the same jobs. For my fellow data nerds: it’s a 0.87 correlation between the two. This is quite significant. It means that nature abhors a vacuum — when you start to take tasks away, people start doing other, more valuable things.
That’s also why it’s so important — to Joe’s earlier point — to redefine the unit of analysis. The real unit of analysis needs to be use cases: how specifically, sector by sector, role by role, are people using AI? Only when we can define those use cases can we track those changes.
Right now, a relatively small fraction of jobs can articulate a specific use case — not just broadly saying “we’re an AI-first company,” but something precise. When you get to that precision, though, you can see AI adoption growing very rapidly. Then we can also start to understand the broader questions we are all here to find answers to: which jobs go away, and how do jobs change from the inside?
Future - Our Best Guesses for What’s Next
MICHAEL: As I listen, I’m hearing us describe the impact of AI on jobs across a couple of different vectors: is the work growing or contracting? Does having more humans actually create more friction?
Matt, I’d love to start with you. Looking at the task level, where are you seeing future growth? Where are you genuinely uncertain? And where are you seeing clear evidence of a task or set of tasks going away?
MATT: Much of this will hinge on who has the agency to innovate within their existing jobs. Ultimately, AI will do more than create efficiency in jobs where people have the agency to use it creatively. That isn’t necessarily just white-collar work. There are a lot of jobs where people have real agency to do genuinely ingenious things. Take something as simple as a real estate survey: we now use drones to do things that once had to be done manually or simply weren’t done at all.
Productivity is a ratio — output to input, or the value of what you produce relative to the cost. In jobs with agency, that ratio will tilt toward the numerator, toward value. Instead of productivity gains coming from less or lower-cost labor, they’ll come from people’s work actually being worth more.
JOE: People who can do this kind of value-finding with AI have a durable asset in this labor market, something I call contextual intelligence: a deep understanding of the purpose and process of their work, a marketplace, or competitive reality. When they move to that more expansive frontier Matt is describing and exercising more agency — in part because AI is creating a buffer against the constant interruptions Cal Newport has written about, things like email, Slack, and smartphones — they will work with AI to get ever-improving outcomes. That unlocks a lot of productivity.
Jobs that don’t require much contextual intelligence but involve lots of routine cognitive work are going to be imperiled. One thing that gets misreported is the assumption that AI is coming for the recent college graduate with a 3.9 GPA from a selective school. In fact, it’s going to hit a lot of clerical and lower-value-added work, often held by people with some college but no degree — maybe a community college credential and good work experience. Those jobs simply don’t require the contextual intelligence that makes a human a strong partner for AI.
SHAD: We should probably talk about early career, because that’s where the data is moving in a concerning direction.
Empirically, there are fewer entry-level white-collar jobs than there used to be. Among many of our customers and partners, it’s essentially a no-hire, no-fire situation right now — they’re being very cautious.
We’ve studied how our own software development work is changing; today, most of our early engineering talent are managing multiple agents — effectively playing an engineering manager role. In our AI fluency framework, discernment is critical, and it’s a skill you typically develop through your first few years on the job. The question we’re asking is how you accelerate to that level of discernment — the pattern recognition you normally learn over two to three years in a role. Maybe there’s a role for higher education and other talent development programs there.
The other concern, which Joe raised, is the work from Opportunity@Work and Brookings looking at gateway jobs for STARs — workers skilled through alternative routes, without degrees — and how those are under threat. Many of those clerical and entry-level corporate roles are the ones that get you into corporate America and allow you to prosper over time. If those disappear, what the ladders of opportunity look like becomes very unclear.
These are the questions we’re concerned about. We’re trying to do action learning with partners like Code Path: how do you move people up that spectrum quickly within an education program, knowing the jobs themselves are changing every six months and the AI underneath them even faster? If we were training people on specific Anthropic certifications six months ago, that training would already be outdated. So how do we embed that dynamism in our partners so we’re all moving with this transition? Because if we don’t, the bimodal distribution — between those deeply engaged with the technology and those left out — will become extreme.
The access question comes up a lot. If you’re in Silicon Valley, people say this technology is great and everyone will get it eventually. Yes — but we know it takes time for technology to reach specific communities, those communities are often behind, and given the outsized productivity gains we’re seeing, falling behind becomes more costly. Even over the next five years, access will have a significant effect on how people participate in the labor market.
MICHAEL: Allison, I’ll go to you and then everyone can weigh in. You’ve observed that some of this debate depends on the perspective you take around where in the stack you sit relative to AI. Can you expand on that — why are we getting such sharply different takes on where work is going?
ALLISON: I’m a very visual person, and I came up with what I call the bot sandwich: humans above and below, bots in the middle.
I interview experts on the future of AI and work all the time for our publication, and I kept getting wildly different takes. Some were evangelists: AI is going to change work for the better, make it more purposeful, unleash agency and creativity, shorten the time between idea and execution. Others had the exact opposite view: AI is already being used to surveil, optimize, and dehumanize work.
I couldn’t figure out how incredible experts — so deeply admired, so data-centered — were arriving at such radically different conclusions about the implications for purposeful work. And then I realized: the difference was the roles they were looking at.
Are they looking at roles that sit above the bots — directing them, getting leverage from them, expanding their span of control, doing things they never could have imagined? Or roles that sit below the bots — being managed by them, told what to do, surveilled on every ticket, every task? And increasingly, staying on top is harder, because it requires judgment, discernment, intuition, and the ability to leverage these tools not just in your own work but cross-functionally with peers.
One of my concerns is that as AI gets more powerful, the top piece of bread gets smaller, and more people start falling below the line. Those jobs become more rote, more rules-based, and eventually get automated away. That’s one of the ways I interpret why people are either very excited or very dystopian about the future of work: it usually comes down to which jobs and which people they’re looking at.
MICHAEL: Alright — lightning round to close it out. Hot takes, predictions, things we haven’t explored yet. Matt, you want to start?
MATT: I’m going to end with the opposite of a prediction. There’s no shortage of prognostication about AI — we don’t need more of it, however intelligent the defense of whatever position you’re advancing. What we need is a Doppler radar: something that tells you, as you’re spreading your picnic blanket out on a July afternoon, whether there’s a thunderstorm coming in the next forty-five minutes.
When we have that, it will allow us to see where changes are happening at the leading edge of any occupation, to preskill rather than reskill — because reskilling has historically never worked — and to actually anticipate where new jobs are emerging and get people there.
SHAD: For me it comes back to this idea that we’re early and we have to act with intentionality. I was in Cleveland a few weeks ago running a workshop with small and medium businesses. The things people built in one hour — with no coding knowledge whatsoever — for their own businesses were absolutely astounding. Things they wouldn’t have done otherwise, couldn’t have afforded to buy as software, wouldn’t have found the time to commission. But suddenly they had solutions for inventory management, for tax compliance, for all sorts of specific problems. The cost to build things is coming down, and many more people can build hyper-local solutions that matter to them. SMBs employ about 47% of workers in America. There are effects starting to emerge there that are worth watching closely — what is emerging, as well as what is disappearing. That’s genuinely exciting.
Another example: advanced bioinformatics and genomic research used to be the purview of a handful of R1 institutions. Now, with AI, universities that couldn’t previously get in the game are getting in. They can create more roles for their postdocs; it’s no longer centered around a small elite. Think about what that will do for scientific discovery.
MICHAEL: Joe, Allison — thirty seconds each.
JOE: Two quick observations. First, the logic underpinning most human resource management rules in large companies is no longer fit for purpose, and all of it will have to be overhauled. Second — and we haven’t mentioned this word once — agents. Agentic AI is going to be the next, and I think much more substantial, wave.
The inference and token cost of using LLMs for every task is high; a rigorously trained, smaller language model would handle many of those tasks far more efficiently. Anthropic’s strategy — being the tool provider regardless of whether it’s an LLM or an SLM — is brilliant for exactly that reason. Agentic AI is being funded at a colossal rate. Those preserved jobs we see right now are essentially a market signal for people to develop business plans around the economic problems agents will create. We are not at the end of the beginning yet, but when we really raise the curtain on agentic deployment, we may have an entirely new vector to pay attention to.
ALLISON: My most confident prediction over the next decade is this: the future of good jobs is non-routine, non-rules-based work, because AI is good at the inverse. And that starts, as Shad has pointed out, all the way at the entry level. That has enormous implications for our education systems, because most of those systems have been designed to prepare people for rules-based, routine work. Full stop. We absolutely have to change that. And I think that has to be the number-one provocation for the future of high school and the future of college.
MICHAEL: With that, please join me in thanking Allison, Joe, Shad, and Matt. Thank you.