Donna Morris, Walmart’s chief human resources officer, recently predicted that office roles will become broader and more generalist because of AI. Rather than hiring specialists for every area of human resources, for example, you might hire HR generalists who can use AI to fill in their knowledge gaps and work across specialties.
An important new working paper by researchers at Stanford and Harvard Business School explores a related idea by examining the extent to which genAI enables workers to perform tasks that are either within, near, or far away from their area of expertise.
If genAI allows people to stretch into other professional areas, the implications would extend beyond job specialization to workforce mobility, training, and hiring. “For example, data scientists could transition into another function and occupation within the same organization (e.g. marketing analyst, financial analyst) with significantly less retraining than would otherwise be needed without genAI,” the researchers write.
So what did they actually find? Workers could successfully perform adjacent tasks with AI assistance, but they hit a “genAI wall” when trying to perform certain tasks that are too distant from their expertise.
This suggests genAI could make jobs less specialized but only up to a certain point. You still need to have foundational knowledge in the domain you’re working in to be able to perform at an expert level, even with genAI support. The research supports the idea that domain expertise—such as you get from studying marketing, finance, law, or chemistry—still has value even as genAI takes on tasks that used to be performed by humans.
“We feel that genAI is giving us a superpower, but the reality is that if it’s an area that we do not understand and we don’t know much about, we are just performing at the baseline of the model,” explains Iavor Bojinov, one of the paper’s authors and an associate professor of business administration at Harvard Business School. “If there’s some blatant things wrong, maybe we can catch that. But in general, we really can’t improve it.”
The experiment
The researchers ran an experiment with UK-based global trading company IG, where they had three groups of employees write web articles for the company’s website:
- Web analysts, who already do this as part of their job.
- Marketing specialists, who perform different, but related activities.
- Technology specialists, who are mostly software engineers and data scientists without relevant experience.
The act of writing the article was divided into two tasks: creating a template with the article’s ingredients—like its keywords and headings—and actually writing the article.
The researchers found genAI to be a performance equalizer for building the article template. Without genAI, web analysts, on average, did this task significantly better than their marketing and technology peers, as assessed by graders. With genAI, they all performed comparably.
But when it came to actually writing the article, the effects were more mixed. Marketing specialists wrote web articles that were on par with the web analysts’ articles with the help of genAI. The technology specialists, however, lagged significantly behind the marketing specialists.
The “genAI wall”
Why the different effects for building the template versus writing the article?
The researchers explain that making the template is a conceptual task that requires knowledge that’s more explicit. Based on a topic, you generate relevant headings, keywords, etc., following a clear template. GenAI tools like ChatGPT excel at this type of task.
Actually writing the article, on the other hand, is much less structured, and it requires the type of knowledge and judgment you gain through experience. Based on a template, you have to write an article with coherent prose and marketing best practices. If you don’t have relevant knowledge, you may lack the judgment to improve what genAI gives you. For example, one of the data scientists said he cut certain buzzwords from the article the AI tool produced because he “prefer[s] articles that are clear and direct.”
This is where the researchers’ concept of the “genAI wall” comes into play. If you have expertise in the relevant area, genAI can help you perform new tasks. This is why the marketing specialists could use genAI to perform as well as their web analyst peers.
But as you get farther away from your area of expertise, there are tasks for which you will eventually hit a “genAI wall”—a point at which genAI won’t close the gap between you and an expert. This finding echoes what economist David Autor has written about AI and expertise: “AI can extend the reach of expertise by building stories atop a good foundation and sound structure. Absent this footing, it is a structural hazard.”
Specialists vs. generalists
We’ll need more research to know how much the results of this study are generally applicable. The experiment had a very small sample size, and it looked at a specific set of marketing tasks. For example, would the results look the same if you had software engineers, product managers, and marketers each build a new product feature with the help of AI?
Still, the study does suggest that, at least in some areas, jobs could become less specialized. For example, in the future, you may just need people who are experts in marketing who can use AI well, rather than people who specialize in search engine optimization (SEO) or email campaigns. This naturally leads to more questions, like what’s the right level of specialization for workers to pursue, and what does all of this mean for organizations?
Bojinov tells us he personally thinks domain knowledge will be “super, super vital,” but within each domain, you may not need to specialize as much. “I don’t really think you’re going to need people to make a career out of…SEO anymore. I think they would be better off being more general.”
That would allow organizations a new degree of flexibility because you could move people around based on the needs of the business. “If you have one team that is super focused on SEO optimization, maybe there is some time where you don’t need to write that many articles, but you actually have a big conference coming up and you need people to flex and go and work over there,” explains Bojinov. In that case, you’d want a team of expert marketers, who can use genAI to tackle whatever marketing problem the business is facing at that moment.
There will likely be areas of the economy where deep specialization continues to pay dividends, like research where you’re pushing the frontier of knowledge forward or in domains where AI isn’t as helpful. But for many workers, the path forward may be to focus on becoming a deep expert in your domain—such as HR, finance, operations, or software engineering, for example—and learn how to use AI to stretch into whatever tasks come up.