.
Image: Freepik
If you listen to the CEOs of elite AI companies or take even a passing glance at the U.S. economy, it’s abundantly obvious that AI excitement is everywhere.
America’s biggest tech companies have spent over $100 billion on AI so far this year, and Deutsche Bank reports that AI spending is the only thing keeping the United States out of a recession.
Yet if you look at the average non-tech company, AI is nowhere to be found. Goldman Sachs reports that only 14% of large companies have deployed AI in a meaningful way.
What gives? If AI is really such a big deal, why is there a multi-billion-dollar mismatch between excitement over AI and the tech’s actual boots-on-the-ground impact?
A new study from Stanford University provides a clear answer. The study reveals that there’s a right and wrong way to use AI at work. And a distressing number of companies are doing it all wrong.
The study, conducted by Stanford’s Institute for Human-Centered AI and Digital Economy Lab and currently available as a pre-print, looks at the daily habits of 1,500 American workers across 104 different professions.
Specifically, it analyzes the individual things that workers actually spend their time doing. The study is surprisingly comprehensive, looking at jobs ranging from computer engineers to cafeteria cooks.
The researchers essentially asked workers what tasks they’d like AI to take off their plates, and which ones they’d rather do themselves. Simultaneously, the researchers analyzed which tasks AI can actually do, and which remain out of the technology’s reach.
With these two datasets, the researchers then created a ranking system. They labeled tasks as Green Light Zone if workers wanted them automated and AI was up to the job, Red Light Zone if AI could do the work but people would rather do it themselves, and Yellow Light (technically R&D Opportunity Zone, but I’m calling it Yellow Light because the metaphor deserves extending) if people wanted the task automated but AI isn’t there yet.
They also created what’s essentially a No Light zone for tasks that AI is bad at, and that people don’t want it to do anyway.
The results are striking. Workers overwhelmingly want AI to automate away the boring bits of their jobs.
Stanford’s study finds that 69.4% of workers want AI to “free up time for higher value work” and 46.6% would like it to take over repetitive tasks.
Checking records for errors, making appointments with clients, and doing data entry were some of the tasks workers considered most ripe for AI’s help.
Importantly, most workers say they wanted to collaborate with AI, not have it fully automate their work. While 45.2% want “an equal partnership between workers and AI,” a further 35.6% want AI to work primarily on its own, but still seek “human oversight at critical junctures.”
Basically, workers want AI to take away the boring bits of their jobs, while leaving the interesting or compelling tasks to them.
A chef, for example, would probably love for AI to help with coordinating deliveries from their suppliers or messaging diners to remind them of an upcoming reservation.
When it comes to actually cooking food, though, they’d want to be the one pounding the piccata or piping the pastry cream.
So far, nothing about the study’s conclusions feel especially surprising. Of course workers would like a computer to do their drudge work for them!
The study’s most interesting conclusion, though, isn’t about workers’ preferences—it’s about how companies are actually meeting (or more accurately, failing to meet) those preferences today.
Armed with their zones and information on how workers want to use AI, the researchers set about analyzing the AI-powered tools that emerging companies are bringing to market today, using a dataset from Y Combinator, a storied Silicon Valley tech accelerator.
In essence, they found that AI companies are using AI all wrong.
Fully 41% of AI tools, the researchers found, focus on either Red Light or No Light zone tasks—the ones that workers want to do themselves, or simply don’t care much about in the first place.
Lots more tools try to solve problems in the Yellow Light Zone—things like preparing departmental budgets or prototyping new product designs—that workers would like to hand off to AI, but that AI still sucks at doing.
Only a small minority of today’s AI products fall into the coveted Green Light zone—tasks that AI is good at doing and that workers actually want done. And while many of today’s leading AI companies are focused on removing humans from the equation, most humans would rather stay at least somewhat involved in their daily toil.
AI companies, in other words, are focusing on the wrong things. They’re either solving problems no one wants solved, or using AI for tasks that it can’t yet do.
It’s no wonder, then, that AI adoption at big companies is so low. The tools available to them are whizzy and neat. But they don’t solve the actual problems their workers face.
For both workers and business leaders, Stanford’s study holds several important lessons about the right way to use AI at work.
Firstly, AI works best when you use it to automate the dull, repetitive, mind-numbing parts of your job.
Sometimes doing this requires a totally new tool. But in many cases, it just requires an attitude shift.
A recent episode of NPR’s Planet Money podcast references a study where two groups of paralegals were given access to the same AI tool. The first group was asked to use the tool to “become more productive,” while the second group was asked to use it to “do the parts of your job that you hate.”
The first group barely adopted the AI tool at all. The second group of paralegals, though, “flourished.” They became dramatically more productive, even taking on work that would previously have required a law degree.
In other words, when it comes to adopting AI, instructions and intentions matter.
If you try to use AI to replace your entire job, you’ll probably fail. But if you instead focus specifically on using AI to automate away the “parts of your job that you hate” (basically, the Green Light tasks in the Stanford researchers’ rubric), you’ll thrive and find yourself using AI for way more things.
In the same vein, the Stanford study reveals that most workers would rather collaborate with an AI than hand off work entirely.
That’s telling. Lots of today’s AI startups are focusing on “agents” that perform work autonomously. The Stanford research suggests that this may be the wrong approach.
Rather than trying to achieve full autonomy, the researchers suggest we should focus on partnering with AI and using it to enhance our work, perhaps accepting that a human will always need to be in the loop.
In many ways, that’s freeing. AI is already good enough to perform many complex tasks with human oversight. If we accept that humans will need to stay involved, we can start using AI for complex things today, rather than waiting for artificial general intelligence (AGI) or some imagined, perfect future technology to arrive.
Finally, the study suggests that there are huge opportunities for AI companies to solve real-world problems and make a fortune doing it, provided that they focus on the right problems.
Diagnosing medical conditions with AI, for example, is cool. Building a tool to do this will probably get you heaps of VC money.
But doctors may not want—and more pointedly, may never use—an AI that performs diagnostic work.
Instead, Stanford’s study suggests they’d be more likely to use AI that does mundane things—transcribing their patient notes, summarising medical records, checking their prescriptions for medicine interactions, scheduling follow-up visits, and the like.
“Automate the boring stuff” is hardly a compelling rallying cry for today’s elite AI startups. But it’s the approach that’s most likely to make them boatloads of money in the long term.
Overall, then, the Stanford study is extremely encouraging. On the one hand, the mismatch between AI investment and AI adoption is disheartening. Is it all just hype? Are we in the middle of the mother of all bubbles?
Stanford’s study suggests the answer is “no.” The lack of AI adoption is an opportunity, not a structural flaw of the tech.
AI indeed has massive potential to genuinely improve the quality of work, turbocharge productivity, and make workers happier. It’s not that the tech is overhyped—we’ve just been using it wrong.
ABOUT THE AUTHOR
Thomas Smith is a Johns Hopkins-trained artificial intelligence expert and journalist with 15 years of experience. Smith was hailed as a "veteran programmer" by the New York Times for his work with human-in-the-loop AI, served as an Open AI Beta tester, and has led the AI-driven photography agency Gado Images as Co-Founder/CEO for 12 years.
FAST COMPANY