Turns out, managing bots is just as annoying as managing people

Manager holding a giant magnifying glass over humanoid robots working at a desk with computer monitors. — Getty Images; Alyssa Powell/BI

One member of Sumaiya Noor’s customer service team is trying to overachieve — and it’s causing her some problems.

Her AI agent is trained to loop her in when faced with a task it can’t handle. However, she’s noticed that over time, it tries to solve problems outside its scope, providing inaccurate answers with great confidence, which leads to dissatisfied and angry customers.

So every couple of months, Noor said she has to carve out some time from her work to refresh the agent on what it’s supposed to do, and train it on more edge cases.

“Over the period of time, it gets worse rather than getting better,” said Noor, the chief product officer at a UK-based social-impact investment platform who also manages two other AI agents: a junior product manager and an internal “chief of staff.” She said deteriorating performance was a major difference she observed among her three AI agents compared with her seven human reports.

Tech leaders sold an enticing vision of AI agents in the workplace.

Zoom CEO Eric Yuan painted a picture of digital twins taking his meetings while kicking back at the beach on a four-day workweek. Salesforce teased a similar concept: the ultimate corporate sidekick to fire off sales pitches and run your day.

For many workers, that future hasn’t quite arrived. Instead, managing bots can be just as difficult as managing people. Workers shared with Business Insider a laundry list of growing pains that come with managing agentic AI: “rot,” long-running tasks derailed by a single error, and agents duplicating one another’s work.

Agents ≠ less work

Agentic AI is still far from ubiquitous in the workplace — but adoption is rising. In a Boston Consulting Group report published in June, 30% of respondents said their organizations had integrated AI agents into workflows, up from 13% the year before.

Offloading tasks doesn’t necessarily mean employees get more time back — and that might be the biggest misconception embedded in the promise of agents.

Cisco engineering director Sergio Freitas said agents haven’t caused him to work longer hours, but they haven’t shortened his days either. While traditional coding required focus, it felt like an immersive craft that many devs enjoyed, he said. Orchestrating agents demands a different kind of attention: more task switching and more oversight. That can make the work feel more intense, he said.

“The fatigue comes from the ability to do more,” Freitas said.

More output doesn’t mean better work. In fact, it can often indicate the opposite. A recent Boston Consulting Group study of 1,488 full-time US workers described the phenomenon as “AI brain fry,” which is the mental fatigue caused by excessive monitoring of AI agents. Participants reported mental fog, trouble focusing, sluggish decision-making, and headaches — strains that could ultimately drive up error rates and turnover.

Part of the challenge is keeping up with agents who can operate 24/7, said Elijah Wee, an associate professor of management at the University of Washington’s Foster School of Business.

“People burn themselves out signaling their worth,” Wee said, adding that humans can’t compete with agents.

‘We handed ourselves a second job’

Workers aren’t just grappling with the reality of a constantly interrupted workflow — they’re also on the hook if their agent screws up. The result is more time spent babysitting bots and policing legal liability.

“I spend less time doing things and more time checking them,” Dan Lewis, a newsletter author, told Business Insider.

Sebastian Gierlinger, vice president of AI and IT at Storyblok, said developers were “genuinely frustrated” by the volume of code they had to review when the company first introduced AI tools. Review and testing practices now need to keep pace with the speed at which AI enables employees to build, he said.

Even setting up the agents can come with a legal headache. Nicholas Cohen, a 20-year-old intern at solar and energy storage development firm Headwater Energy, said the “biggest challenge” is preparing data for efficient use in agentic workflows.

“You can’t just chuck an agent into a Citadel or a big multi-strat firm, and expect everything to be compliant,” Cohen, who also founded energy AI company Redr Labs, said. “Especially if it’s not trained in-house.”

Even platforms like AWS Bedrock, which allow companies to keep their data within their own environment, can still pose compliance risks, Cohen said. Building models in-house and training them on proprietary data may offer companies greater control, he added, but doing so requires significant computing power and can quickly become expensive.

AI agents sometimes get worse at their jobs

Human employees — for the most part — improve their skills the longer they’re in a workplace. The same can’t be said about AI agent employees.

The phenomenon of AI context rot has been well documented, referring to LLMs returning increasingly poor results the longer they are used. This happens when new context is fed to LLMs, pushing out old context, causing agents to drop specific guidelines and “drift away from established operational constraints,” according to a Salesforce report by Cody Gould, one of the company’s forward-deployed engineers, and Molly Futrell, a technical writer.

Ivan Burazin, the CEO of New York-based AI firm Daytona, said managers bear some responsibility for AI agents progressively getting worse at their jobs. He said that at the start of a new project, when people are giving the agent the task for the first time, they’ll provide a big prompt and tons of context.

“As you keep progressing, you become lazy as a human. Your prompt five days later is like, ‘Make this better,'” Burazin said. “Because your prompts degrade, the results of the agents also degrade.”

Burazin said his engineers run at least two to three AI agents at any time, with his more experimental engineers deploying 15 agents in parallel. Despite their challenges, agents help him keep his actual team lean. Daytona has a 30-person workforce.

Gierlinger said his workaround is to periodically wipe the memory and caches in his AI environment, starting with a clean slate. While the models lose old context, it breaks whatever pattern the model has locked into.

There’s also no buildup of trust and integrity over time between AI agents and managers, said Michael Kruesi, an associate professor at the Singapore Institute of Technology who has researched and taught organizational behavior.

A long way to go before AI agents can supersede human employees

AI literacy and knowing how to work with AI agents are now non-negotiable for many employers. Thomas Laboulle, the founder and CEO of Toku, said that it’s a requirement at his company, “in the same way that you expect someone to be able to use a computer or to use Excel.”

But there’s still a long way to go before human employees become redundant.

When it comes to deciding whether to hire a new employee or spin up another AI agent, Brad Sams, a vice president at Michigan-based Stardock Software, creativity is a big factor.

Sams, despite managing an average of 30 agents daily himself, said they don’t hit the mark for creative projects.

“We believe that if a human is going to read it — marketing material, for instance — we need a human employee at Stardock to have written it, or at least reviewed it,” he said.

As agents become more integrated in the workforce, Michael Housman, the founder and CEO of AI Performance Lab, said that workers need to “double down” on the parts that make people human. That includes the ability to lead teams, read the room, tell stories, and collaborate.