AI Companies Want to Give You a New Job. Your New Team? A Million AIs.

If AI agents can do your job better and faster, what’s left for you to do?

AI agents are already taking over a range of human tasks. Specialized AI agents are at work right now in customer service, drug discovery, and software development, increasing productivity and speed-to-market by 50% or more, according to one study. 

This may be just the beginning of a drastic overhaul of the nature of work itself. In a not-too-distant future, we may not be doing work at all— at least not the work we’re used to. Instead, we’ll be overseeing AI agents performing familiar tasks for us. 

In a recent interview with podcast host Dwarkesh Patel, Microsoft CEO Satya Nadella laid out such a future — one in which each person effectively manages millions of AI agents. 

“I feel there's a new inbox that's going to get created, which is my millions of agents that I'm working with will have to invoke some exceptions to me, notifications to me, ask for instructions," he said, adding that AI agents will be “like lean manufacturing for knowledge work, reducing waste and increasing value.” 

On Wednesday, Nadella announced a “big day” for Copilot, his company’s effort to make this vision a reality. New “Researcher” and “Analyst” agents, which now serve as Nadella’s “go-to 24/7 experts,” are part of Microsoft’s push to make “everyone” an “agent boss.”

The idea may sound appealing; none of us wants to spend time doing tedious, repetitive tasks. But managing an army of digital agents working on our behalf comes with a host of difficult questions and open problems. 

How will productivity be defined if your job is to steer rather than execute? How will companies restructure in response? Could agents enable us to do so much that we become unfocused, and thus don’t excel at anything? And how far are we from any of this coming to pass?

Doing 100 Jobs

Amjad Masad, the founder and CEO of Replit, a company developing an advanced AI agent for programming, believes agents will fundamentally change the nature of knowledge work. 

“Management will not be as important, because employees — when skill is no longer an issue — will be able to work as generalists and be expected to find the most impactful problem to solve,” he said in an email, adding that agents will give workers so much leverage that each person will be able to do the job of 100 people. 

However, he added, this shift will require companies to completely change how they’re structured, with far less isolation between roles and departments. 

“If you’re in sales, you won’t need a role that is primarily focused on generating leads because the salesperson can spin up agents to do that,” he said. “So roles will start to blur, and eventually that will lead us to fundamentally change how organizations work.” 

Instead of a department with ten financial analysts, a company may just need one or two to review the work done by agents. Companies will become less hierarchical, since employees will have access to more information and each will be managing multiple agents. 

A large chunk of your job will be to track and summarize your agents’ work — potentially yielding more time to collaborate with colleagues and participate in cross-functional work. 

For example, rather than putting days into a quarterly earnings report and then forwarding it to the sales team, financial analysts may review a report created by an agent, then participate in sales meetings or other strategic initiatives themselves.

/odw-inline-subscribe-cta

Is More Work Good Work?

Carl Benedikt Frey, Professor of AI & Work at the Oxford Internet Institute and Director of Future of Work at the Oxford Martin School, isn’t so sure this way of working will bring widespread benefits, particularly when it comes to workplace productivity. 

Since agents will enable employees to manage more workflows, employees may spread themselves too thin, doing a mediocre job on many tasks rather than a thorough job on a few. 

“AI supercharges our ability to do more things,” Frey said in an interview. “But there’s always a tension; how deep do I go into one particular thing? The more emails I write, the more emails I get back. It’s an enabler, but it’s also something of a curse.”

He pointed out that the 1980s and 1990s brought unprecedented improvements in our ability to communicate and exchange ideas. 

“We can do things so much more effectively by email than we could by letter or telephone, yet we’ve only seen a blip in productivity statistics,” he said. “Our ability to do more back office operations seems to have contributed to more back office operations, not more time for productive contemplation.”

This “less is more” principle can already be seen in digital media. AI has facilitated the creation of videos, images, stories, articles, and other media, giving anyone with a computer and internet the tools to generate “original” content. The vast majority of that content, however, is mediocre — or worse, and sifting through it to find high-value items becomes its own task. 

Might a similar Pandora’s Box of AI agents end up bogging knowledge workers down with more responsibilities than they can properly attend to? And as importantly, how far are we from this becoming an issue? 

At present, while agents are being trained and implemented across a broad range of capabilities, their accuracy and reliability still leave much to be desired. 

Alan Chan, a research fellow at the Centre for the Governance of AI, shared one experience: He tried to use an AI agent to help him look for a new flat in London. He gave the agent his criteria and asked it to create a list of options he should look into further. It failed to do so, instead cycling through a loop of clicking on random websites. The agent was, however, able to complete a simpler task for Chan: buying movie tickets.

How Soon?

“An open question is, how quickly are AI agents improving?” Chan said in an interview. 

He brought up a recent report from METR, an organization that evaluates AI systems’ capabilities. The study ranks tasks according to how long it takes humans to complete them, then graphs AI agents’ ability to complete those tasks. 

In 2021, agents could only do tasks that take humans five minutes. But by the end of 2024, they could do tasks that take humans an hour.

“The key takeaway is that there seems to be a doubling time,” Chan said. “Every seven months, agents are able to complete tasks that last twice as long.” 

If this trend continues, seven months from now agents will be able to complete tasks that take humans two hours — and in a few years, tasks that take us a week. However, there’s reason to be skeptical about whether this trend will hold up. 

“The focus so far has been on digital and software engineering tasks,” Chan said. “We haven’t really evaluated agents on more real-world tasks. Not everyone is doing coding in their office jobs.”

He added that a second source of uncertainty is that reliability hasn’t been a focus. METR’s benchmark evaluates 50 percent reliability—meaning agents can complete a two-hour human task correctly about half the time.

“Something being 50 percent reliable isn’t a very strong indicator of whether it’s going to be adopted or not,” Chan said.

The Agent Boss

While some doubt remains, Microsoft’s Satya Nadella and other tech leaders are betting that much of your job is automatable.

Historically, automation has, eventually, opened the door to new forms of work. Job titles like software engineer or IT consultant didn’t exist before the information age, yet they’re now held by tens of millions.

The “agent boss” could define the next era of work. Whether it’s a role worth wanting — or one that can absorb the millions of jobs AI is expected to displace — is unclear.