Will your job survive AI?

In recent weeks, several prominent executives at big employers such as Ford and J.P. Morgan Chase have been offering predictions that AI will result in large white-collar job losses.

Some tech leaders, including those at Amazon, OpenAI, and Meta have acknowledged that the latest wave of AI, called agentic AI, is much closer to radically transforming the workplace than even they had previously anticipated.

Dario Amodei, chief executive of AI firm Anthropic, said nearly half of all entry-level white-collar jobs in tech, finance, law, and consulting could be replaced or eliminated by AI.

Christopher Stanton, Marvin Bower Associate Professor of Business Administration at Harvard Business School, studies AI in the workplace and teaches an MBA course, “Managing the Future of Work.” In this edited conversation, Stanton explains why the latest generation of AI is evolving so rapidly and how it may shake up white-collar work.

---

Several top executives are now predicting AI will eliminate large numbers of white-collar jobs far sooner than previously expected. Does that sound accurate?

I think it’s too early to tell. If you were pessimistic in the sense that you’re worried about labor market disruption and skill and human capital depreciation, if you look at the tasks that workers in white-collar work can do and what we think AI is capable of, that overlap impacts about 35 percent of the tasks that we see in labor market data.

“My personal inclination — this is not necessarily based on a deep analytical model — is that policymakers will have a very limited ability to do anything here unless it’s through subsidies or tax policy.”

The optimistic case is that if you think a machine can do some tasks but not all, the tasks the machine can automate or do will free up people to concentrate on different aspects of a job. It might be that you would see 20 percent or 30 percent of the tasks that a professor could do being done by AI, but the other 80 percent or 70 percent are things that might be complementary to what an AI might produce. Those are the two extremes.

In practice, it’s probably still too early to tell how this is going to shake out, but we’ve seen at least three or four things that might lead you to suspect that the view that AI is going to have a more disruptive effect on the labor market might be reasonable.

One of those is that computer-science graduates and STEM graduates in general are having more trouble finding jobs today than in the past, which might be consistent with the view that AI is doing a lot of work that, say, software engineers used to do.

If you look at reports out of, say, Y Combinator or if you look at reports out of other tech sector-focused places, it looks like a lot of the code for early-stage startups is now being written by AI. Four or five years ago, that wouldn’t have been true at all. So, we are starting to see the uptake of these tools consistent with the narrative from these CEOs. So that’s one piece of it.

The second piece is that even if you don’t necessarily think of displacement, you can potentially think that AI is going to have an impact on wages.

There are two competing ways of thinking about where this is going to go. Some of the early evidence that looks at AI rollouts and contact centers and frontline work and the like suggests that AI reduces inequality between people by lifting the lower tail of performers.

Some of the best papers on this look at the randomized rollout of conversational AI tools or chatbots and frontline call-center work and show that lower-performing workers or workers who are at the bottom of the productivity distribution disproportionately benefit from that AI rollout tool. If these workers have knowledge gaps, the AIs fill in for the knowledge gaps.

What’s driving the accelerated speed at which this generation of AI is evolving and being used by businesses?

There are a couple of things. I have a paper with some researchers at Microsoft that looks at AI adoption in the workplace and the effects of AI rollout. Our tentative conclusion was that it took a lot of coordination to really see some of the productivity effects of AI, but it had an immediate impact on individual tasks like email.

“Our tentative conclusion was that it took a lot of coordination to really see some of the productivity effects of AI, but it had an immediate impact on individual tasks like email.”

One of the messages in that paper that has not necessarily been widely diffused is that this is probably some of the fastest-diffusing technology around.

In our sample, half of the participants who got access to this tool from Microsoft were using it. And so, the take-up has been tremendous.

My guess is that one of the reasons why the executives … didn’t forecast this is that this is an extraordinarily fast-diffusing technology. You’re seeing different people in different teams running their own experiments to figure out how to use it, and some of those experiments are going to generate insights that weren’t anticipated.

The second thing that has accelerated the usefulness of these models is a type of model called a chain-of-thought model. The earliest versions of generative AI tools were prone to hallucinate and to provide answers that were inaccurate. The chain-of-thought type of reasoning is meant to do error correction on the fly.

And so, rather than provide an answer that could be subject to error or hallucinations, the model itself will provide a prompt to say, “Are you sure about that? Double check.” Models with chain-of-thought reasoning are much, much more accurate and less subject to hallucinations, especially for quantitative tasks or tasks that involve programming.

As a result, you are seeing quite a lot of penetration with early stage startups who are doing coding using natural-language queries or what they call “vibe coding” today. These vibe-coding tools have some built-in error correction where you can actually write usable code as a result of these feedback mechanisms that model designers have built in.

The third thing driving major adoption, especially in the tech world, is that model providers have built tools to deploy code. Anthropic has a tool that will allow you to write code just based on queries or natural language, and then you can deploy that with Anthropic tools.

There are other tools like Cursor or Replika where you will ultimately be able to instruct a machine to write pieces of technical software with limited technical background. You don’t necessarily need specific technical tools, and it’s made deployment much, much easier.

This feeds back into the thing that I was telling you earlier, which is that you’ve seen lots of experiments and you’ve seen enormous diffusion. And one of the reasons that you’ve seen enormous diffusion is that you now have these tools and these models that allow people without domain expertise to build things and figure out what they can build and how they can do it.

Which types of work are most likely to see change first, and in what way? You mentioned writing code, but are there others?

I have not seen any of the immediate data that suggests employment losses, but you could easily imagine that in any knowledge work you might see some employment effects, at least in theory.

In practice, if you look back at the history of predictions about AI and job loss, making those predictions is extraordinarily hard.

We had lots of discussion in 2017, 2018, 2019, around whether we should stop training radiologists. But radiologists are as busy as ever and we didn’t stop training them. They’re doing more and one of the reasons is that the cost of imaging has fallen. And at least some of them have some AI tools at their fingertips.

And so, in some sense, these tools are going to potentially take some tasks that humans were doing but also lower the cost of doing new things. And so, the net-net of that is very hard to predict, because if you do something that augments something that is complementary to what humans in those occupations are doing, you may need more humans doing slightly different tasks.

And so, I think it’s too early to say that we’re going to necessarily see a net displacement in any one industry or overall.

If AI suddenly puts a large portion of middle-class Americans out of work or makes their education and skills far less valuable, that could have catastrophic effects on the U.S. economy, on politics, and on quality of life generally. Are there any policy solutions lawmakers should be thinking about today to get ahead of this sea change?

My personal inclination — this is not necessarily based on a deep analytical model — is that policymakers will have a very limited ability to do anything here unless it’s through subsidies or tax policy. Anything that you would do to prop up employment, you’ll see a competitor who is more nimble and with a lower cost who doesn’t have that same legacy labor stack probably out-compete people dynamically.

It’s not so clear that there should be any policy intervention when we don’t necessarily understand the technology at this point. My guess is that the policymakers’ remedy is going to be an ex-post one rather than an ex-ante one. My suspicion is better safety-net policies and better retraining policies will be the tools at play rather than trying to prevent the adoption of the technology.