Will your occupation endure AI?

Recently, several high-profile executives from major corporations, including Ford and J.P. Morgan Chase, have been making predictions that AI will lead to substantial losses of white-collar positions.

Some leaders in technology, like those at Amazon, OpenAI, and Meta, have acknowledged that the current wave of AI, known as agentic AI, is significantly closer to drastically altering the workplace than they had initially expected.

Dario Amodei, CEO of AI company Anthropic, commented that nearly half of entry-level white-collar roles in sectors like tech, finance, law, and consulting could be supplanted or entirely removed by AI.

Christopher Stanton, Marvin Bower Associate Professor of Business Administration at Harvard Business School, researches AI in the workplace and teaches an MBA course titled “Managing the Future of Work.” In this edited dialogue, Stanton discusses why this new breed of AI is advancing so swiftly and how it could disrupt white-collar labor.

---

Many leading executives are now forecasting that AI will significantly reduce white-collar employment much sooner than anticipated. Is that a fair assessment?

I believe it’s premature to determine. If you hold a pessimistic view regarding labor market disruptions and the devaluation of skills and human capital, examining the tasks performed by white-collar workers versus the capabilities of AI shows that approximately 35 percent of those tasks overlap according to labor market data.

“My personal view — this isn’t solely grounded on a thorough analytical model — is that policymakers will find it extremely challenging to address this issue unless it’s via subsidies or tax incentives.”

The optimistic perspective suggests that while machines might automate certain tasks, the work they do could enable people to focus on other aspects of their roles. For instance, you might observe that 20 to 30 percent of tasks a professor might undertake could be managed by AI, while the remaining 70 to 80 percent would be complementary to what AI produces. Those are the two ends of the spectrum.

In reality, it’s likely still too soon to gauge how this will unfold, but we’ve identified at least three or four factors that might lead to the assumption that AI could have a more disruptive impact on the labor market.

One indication is that computer science and STEM graduates are encountering more difficulty securing positions now compared to prior years, which aligns with the notion that AI is taking over work traditionally performed by software engineers.

Looking at reports from entities like Y Combinator or other tech-focused organizations reveals that a significant amount of the code for nascent startups is now generated by AI. Four or five years ago, that wasn’t the case at all. Thus, we observe the adoption of these tools aligns with the expectations set forth by these CEOs. That’s a key insight.

Another point is that even if you don’t view it as displacement, one could argue that AI will influence wages.

There are two contrasting viewpoints regarding the future trajectory. Initial evidence examining AI implementations in contact centers and frontline jobs suggests that AI could diminish inequality among employees by boosting the performance of lower-tier workers.

Notable studies on this topic focus on the randomized application of conversational AI tools or chatbots in call center operations, revealing that lower-performing employees or those situated at the bottom of the productivity spectrum disproportionately gain advantages from these AI tools. If these workers lack certain knowledge, the AIs can effectively bridge those gaps.

What factors are propelling the rapid evolution and application of this generation of AI within organizations?

Several elements contribute to this phenomenon. I co-authored a study with researchers at Microsoft examining AI adoption in workplaces and its effects. Our tentative findings suggest that significant coordination is required to effectively realize some productivity benefits from AI, although it had an immediate influence on specific tasks like email management.

“Our preliminary conclusion indicated that while considerable coordination was necessary to notice productivity benefits from AI, it had an immediate effect on isolated tasks like email.”

One of the messages from that study that hasn’t been widely shared is that this is potentially the fastest-spreading technology we’ve seen.

In our sample, half of the participants who gained access to this Microsoft tool were utilizing it. Thus, the uptake has been substantial.

I suspect that one reason executives … didn’t anticipate this development is due to the extraordinarily rapid diffusion of this technology. We are witnessing various individuals across different teams conducting their own trials to understand its application, and some of these experiments will yield unforeseen insights.

The second factor that has enhanced the efficacy of these models is known as a chain-of-thought framework. The earliest iterations of generative AI tools often generated inaccurate responses. The chain-of-thought reasoning aims to provide real-time error correction.

Thus, instead of delivering answers that could be flawed or misleading, the model itself
“““html

Could you confirm that? Please verify again.

Consequently, you’re witnessing a significant influx of early-stage startups utilizing natural language inquiries for coding, often referred to as “vibe coding” today. These vibe-coding platforms incorporate some inherent error correction, enabling you to produce functional code thanks to these feedback systems developed by model creators.

The third factor propelling major uptake, particularly in the technology sector, is that model developers have created tools for code deployment. Anthropic offers a tool that lets users generate code purely through queries or natural language, which can then be deployed using Anthropic’s tools.

Other platforms like Cursor or Replika will ultimately empower users to direct a machine to create segments of technical software with minimal technical expertise. There’s no absolute necessity for specialized technical tools, which has significantly simplified the deployment process.

This ties back to what I mentioned earlier about the numerous experiments and notable diffusion taking place. A primary reason for this widespread diffusion is the availability of these tools and models that enable individuals without specialized knowledge to create and discover what they can develop and how they can accomplish it.

Which categories of work are most likely to transform first, and how? You indicated coding, but are there additional areas?

I haven’t encountered immediate data indicating job losses, but one could easily theorize that any knowledge-based work could experience some employment impacts.

In reality, if we reflect on previous forecasts concerning AI and job displacement, making such predictions is exceptionally challenging.

We had extensive discussions in 2017, 2018, and 2019 regarding whether we should halt the training of radiologists. However, radiologists are just as active as ever, and we didn’t cease their training. They’re doing even more work, partly because imaging costs have decreased, and many of them are equipped with AI tools.

Thus, in some respects, these tools may take over certain tasks previously performed by humans, while also reducing the expenses associated with new activities. The overall impact is difficult to anticipate because augmenting tasks that complement what humans in these fields are responsible for may ultimately necessitate more humans engaging in slightly altered roles.

Consequently, it’s premature to declare that we will inevitably observe a net displacement in any single industry or in general.

If AI were to suddenly displace a significant number of middle-class Americans or diminish the value of their education and skills, the repercussions could be disastrous for the U.S. economy, governance, and overall quality of life. Are there any policy measures lawmakers should consider today to proactively address this impending shift?

My personal view — not necessarily backed by rigorous analytical modeling — is that policymakers will face considerable limitations in addressing this issue unless they resort to subsidies or tax incentives. Any effort to bolster employment could result in competitors who are more agile and cost-effective without the same traditional labor overhead dynamically outpacing others.

It’s not entirely clear that policy intervention is justified when the technology remains somewhat ambiguous. I suspect that the solutions favored by policymakers will be reactive rather than proactive. I believe improved safety-net policies and enhanced retraining initiatives will be essential rather than attempting to hinder the adoption of the technology.