In corporate America, a moment has been going on in the background of enterprise software dashboards and HR onboarding decks. There is a new name on the team’s list. It comes with a name, clear duties, and sometimes, really, a place on the organization chart. It doesn’t have a desk. It doesn’t eat lunch. It could have a name like Alex, Aria, or anything else the product team thought sounded friendly. Also, it’s really a piece of software.
Happy to have you in the agentic office. A lot more progress has been made than most people think.
A business professor at Boston University named Emma Wiles recently looked into what happens when managers are told that their AI tool is a “coworker” instead of software. The managers found 18% fewer mistakes in the AI’s work, which should make any business leader stop what they’re doing in the middle of a meeting. Also, they were almost twice as likely to send questionable work to someone else to fix it instead of doing it themselves. This got rid of most of the efficiency gains that the AI was supposed to bring in the first place. It seems like Silicon Valley’s efforts to brand itself with terms like “digital colleagues,” “AI employees,” and “synthetic teammates” are making the people who work nearby worse at their jobs.
That’s no longer a small issue. It was almost a third of the 1,261 managers Wiles polled who said their companies already treat AI agents like employees. One-third of them put them on org charts. It is known that Jensen Huang of Nvidia has talked about workplaces full of “digital humans.” In the past few months, Microsoft, OpenAI, Anthropic, and Google have all released products that help managers of teams of AI agents. Many of these products are specifically marketed as cognitive collaborators rather than tools.
The truth about what these systems do is already very impressive. Agentic AI, on the other hand, doesn’t just sit there until it’s told to do something. You can give it a goal and have it work toward it in a loop, scheduling, creating, reviewing, and suggesting things until the job is done or something goes wrong. It has been amazingly fast to adopt. It took about eight years for traditional enterprise AI to reach 70% adoption. Agentic systems reached 35% in just two years, and most experts think that number will keep going up sharply for the rest of this decade. The infrastructure showed up before anyone had a chance to think about what it would mean.
The tech itself isn’t the issue. It’s how it’s framed and what that frames does to accountability. It’s surprisingly easy for different people to be blamed when something goes wrong in an office where AI is called an employee. What Wiles found was that people felt less responsible for AI output when it was said to come from a “coworker.” In the marketing department, that kind of behavior is unsettling. It’s really dangerous when decisions affect people’s lives, like in healthcare, education, or other fields. “The AI did it” is the easy, available explanation that is already being used by some, and it doesn’t always hold up to scrutiny.
It has been pretty clear that this is what the MIT economist who won the Nobel Prize in 2024 means. He has said that the pitch that AI agents can replace people is fundamentally wrong. The more useful question is whether these systems can make people better at their jobs. Unfortunately, there is evidence that the opposite is happening right now, at least in some ways that can be measured.
Stanford researchers did an interesting parallel experiment not long ago. They showed 1,500 workers in 104 different jobs a list of tasks that AI could theoretically do, and then they asked the workers which tasks they’d actually want automated. The results didn’t always match what tech experts thought they would be. Tracking the progress of a case was helpful for law clerks. Even though it seemed like a natural fit for automation, sales reps wouldn’t give up credit checks. There is a difference between what engineers think workers want automated and what workers actually see as friction. This difference needs to be fixed if offices are to work better, not just differently.
As I watch this happen, I can’t help but notice something that the adoption curve doesn’t show: the strategy hasn’t kept up with the deployment. It’s taking longer for executives to redesign the processes around autonomous systems, make it clear who’s responsible for mistakes, and figure out how to manage something that doesn’t sleep, forget, or have a bad week. This makes offices run faster but make people think less clearly in some ways.
Alex, or whatever the software is called, is not a coworker. It’s not more powerful just because you call it a “one.” People near it are just a little less careful. And being careful is exactly what we need right now.
