[DigitalToday reporter Hyunwoo Choo] Companies are putting AI agents to work, but on the ground the management burden is growing faster than productivity. Business Insider reported on Oct. 6 that workers complained the agents require constant checks, retraining and review of results rather than simply taking over tasks.

Sumaiya Noor, chief product officer at a British social impact investment platform, said a customer support AI agent increasingly gave answers outside its assigned scope over time. The agent was trained to hand off tasks it could not handle to humans, but it gradually tried to solve them on its own and confidently produced inaccurate answers, increasing customer complaints. Noor said she sets aside time every few months to retrain the agent’s role and add exception cases.

The pace of AI agent adoption is accelerating. A report released by Boston Consulting Group in June found 30 percent of organisations had integrated AI agents into their workflow. That was up from 13 percent the previous year. Still, handing over work did not mean employees had the same amount of free time.

Cisco engineering director Sergio Freitas said AI agents did not extend working hours, but they did not make the day shorter either. Coding used to be immersive work, but coordinating agents requires more frequent task-switching and supervision, increasing fatigue, he said.

The problem is there is no guarantee results improve even if output rises. Boston Consulting Group described the phenomenon as “AI brain fry” based on a recent survey of 1,488 full-time U.S. workers. It said excessive monitoring of AI agents can build mental fatigue and lead to reduced focus, delayed judgment and headaches. Eliza Wu, an associate professor at the University of Washington’s Foster School of Business, pointed out that people are wearing themselves out trying to keep up with agents that run 24 hours a day.

Accountability issues also remain. Newsletter writer Dan Lewis said the time spent checking results has increased more than the time spent doing something. Sebastian Gierlinger, vice president of AI and IT at Storyblocks, said developers felt a heavy burden when AI tools were first introduced because of the amount of code they had to review.

At the adoption stage, the compliance burden is also heavy. Nicholas Cohen, an intern at Headwater Energy, said preparing data to fit agent-based workflows is the biggest challenge. Platforms such as AWS Bedrock that keep data within a company’s own environment did not completely eliminate compliance risks. Training in-house models on internal data can increase control, but it requires large computing resources and costs, he explained.

Over the long term, performance degradation also repeated. Salesforce described a phenomenon it calls “context rot,” in which new context keeps coming in, pushing out existing guidelines and letting the model drift beyond operating constraints. Ivan Burazin, chief executive of Daytona, said teams give sufficient instructions early in a project, but prompts become flimsy over time and results worsen accordingly.

Even so, companies are starting to see AI literacy as a core capability. In work that requires creativity, however, there was still a role for people. Brad Sams of Stardog Software said he manages an average of 30 agents a day, but outputs that people read, such as marketing materials, must be written by a person or at least reviewed by a person.

The case shows that adopting AI agents does not end with expanding automation. Managing performance degradation, review burdens and compliance problems has emerged as a task for companies if the tools are to take hold in real work.