Somewhere in the last year, “AI agent” stopped being a phrase for engineers and started appearing in the same sentence as “headcount.” Managers who used to worry about staffing a project now find themselves deciding whether a piece of software gets a name, a Slack handle, and a line on the org chart. If that sounds strange, you’re not imagining it. It’s happening at a scale most leaders haven’t caught up with, and the way they handle it over the next year will shape how their teams actually work, not just how their slide decks describe it.
This isn’t a story about robots taking jobs, though that fear shows up in the data too. It’s a story about a decision every manager is quietly making right now, often without realizing it: are you treating AI agents as tools, or as teammates? Research published this year suggests that choice matters more than almost anyone expected, and getting it wrong can create costs that don’t show up until months later. Managing AI agents well turns less on which vendor you pick and more on this one call.
Why Every Manager Is Suddenly Talking About This
Adoption has moved fast. A PwC survey of 300 senior U.S. executives found that more than a third of companies have adopted AI agents broadly, another 27 percent have adopted them in a limited way, and 17 percent have rolled them out across the organization. That means most leadership teams are now managing some kind of software-based worker, whether they call it that or not.
And many do call it that. A Boston Consulting Group study of more than 1,200 HR and finance professionals across the U.S., Canada, and the EU, led by managing director Matthew Kropp, found that nearly a third of managers already frame AI as a teammate or employee in everyday language. More than one in five have listed AI agents on a company org chart, next to the humans.
That habit feels harmless. It isn’t. Gartner projects that by 2029, half of all knowledge workers will need real skills in working with, governing, or building AI agents, and that by the end of this year, 40 percent of enterprise software will include task-specific agents, up from under 5 percent just a year ago. This isn’t a side project anymore. It’s becoming the baseline for how work gets done, which means managing it well is no longer optional.
Not every company is capturing value at the same rate, either. Research cited alongside the PwC survey found that just 20 percent of companies are capturing 74 percent of all the value AI is producing across their industries. The executives PwC surveyed cited cybersecurity concerns and cost, each identified by 34 percent of respondents, as the top barriers holding the rest back. PwC’s read is blunt: those concerns often function as safe excuses more than true blockers. Gartner’s longer-range numbers suggest why so many leaders still feel pressure to move. The firm projects agentic AI could drive roughly 30 percent of enterprise software revenue by the mid-2030s, more than $450 billion, up from about 2 percent today. Whatever the exact number turns out to be, the direction is clear enough that waiting quietly on the sidelines is no longer a neutral choice.
The Study That Should Make Every Manager Pause
Here’s the part that should change how managers think about this. The BCG team ran a large-scale experiment testing what happens when people are told an AI system is their “employee” rather than simply a tool they use. The results, published in Harvard Business Review in May, were not what many executives expected.
People assigned an AI “employee” caught fewer errors in the work than people using the same AI as a plain tool. Trust in how the AI was deployed dropped by 10 percent. Concern that AI would eventually replace their own role rose by 7 percent. The accountability problem was the sharpest finding of all: workers began blaming the AI for mistakes instead of taking ownership themselves.
Kropp put it plainly: when software is given a role that looks like a person’s role, people start “passing the buck.” He said it creates “churn and excess overhead,” because instead of catching and fixing an error, employees route it to a colleague for review, which multiplies the workload instead of reducing it. Writing in HBR, the research team identified four problems that come from treating agents as employees: less individual accountability, more unnecessary escalation, weaker review quality, and more confusion about who is responsible for what. None of this improved adoption. It just made the organization messier.
This is the heart of good AI agent governance: not whether you use the technology, but whether your team still knows, clearly, who owns the outcome.
What Gets Lost When Software Gets a Name
There’s a reason this matters beyond one study. Anthropomorphizing a tool changes how people relate to it, and that shift has a cost that doesn’t show up on a spreadsheet. When employees request extra review of AI-generated work because they do not fully trust it, that review is not free. Someone senior has to do it, on top of an existing workload, which is the opposite of the efficiency gain that justified bringing in the agent in the first place.
Separately, McKinsey’s 2025 research found that 88 percent of organizations already use AI in at least one business function, while 51 percent reported at least one negative AI incident in the past year, ranging from inaccuracy to compliance failures to privacy breaches. Put those numbers together and the picture is clear: adoption is nearly universal, but the guardrails, and the clarity about who is responsible when something breaks, are not keeping pace.
This is where human-AI collaboration either works or falls apart. Teams that treat AI coworkers like colleagues without building the structure a colleague would need, including performance reviews, clear escalation paths, and defined limits, end up with all of the confusion of a new hire and none of the accountability.
Governance, Not Org Charts, Is the Real Job
Mary-Anne Williams, who directs the AI Lab at UNSW Business School and leads a multi-year Australian Research Council project on agentic AI and strategic management, has pointed to a specific failure pattern across organizations: leaders blend three distinct modes of AI use, drafting, advising, and executing, without telling their teams which mode applies to a given task. An agent that drafts a memo and an agent that executes a financial transaction carry completely different risks, but too many teams treat them the same way.
Williams’s research also flags something counterintuitive: when employees stop overriding AI decisions almost entirely, that is not necessarily a sign the AI has become reliable. It is often a sign that people have stopped paying attention. A near-zero override rate looks like trust from the outside. From the inside, it can signal disengagement.
A separate review of 35 studies on human-AI decision-making, published in AI and Society, found that this cuts both ways. In high-stakes fields such as healthcare, law, and public administration, people show both automation bias, trusting the machine too much, and algorithm aversion, trusting it too little, sometimes in the same organization and sometimes in the same person on different days. Leading hybrid teams well means building a process sturdy enough to survive both failure modes, not hoping employees land on the right level of trust by instinct.
How to Actually Lead a Team That Includes AI Agents
Start by naming, out loud and in writing, which mode each agent is operating in for each task. A tool that drafts language for your team’s approval needs a different level of oversight than one that executes a customer refund on its own. Managers who skip this step are the ones who end up with the org-chart problem, treating every agent the same way regardless of what it is actually allowed to do.
Next, protect accountability the same way you would with a new hire, except more deliberately, because the instinct to blame the system is strong and well documented. When something goes wrong, the question cannot be “what did the AI do?” It has to be “who approved this, and what was the review process?” If you cannot answer that quickly, you have a structure problem, not a technology problem.
Watch your override rate the way you would watch any other quality metric. If nobody on your team is ever pushing back on what an agent produces, that is a signal to investigate, not celebrate. Build in a regular moment, weekly is plenty, where someone checks a sample of agent output against what a careful human would have done, and talk with the team about the gaps.
Resist the pull to give agents names, personalities, or org-chart placement unless there is a real operational reason to do it. The BCG findings are specific on this point: the softer, friendlier framing measurably reduced how carefully people checked the work. If your team needs a nickname to remember which system does what, fine. But do not confuse a mnemonic with a management structure.
Finally, budget real time for review, and say so plainly to your team and to whoever you report to. The productivity case for AI agents often assumes the human side of the workflow shrinks to nearly nothing. The research says otherwise. Review capacity is not overhead you can cut once the agent is in place. It is a permanent part of the job now, and pretending otherwise is how the quiet costs pile up.
Tie review time to an actual measurement instead of a vague sense that things feel fine. Gensler’s 2026 research found that roughly 30 percent of employees now qualify as AI power users, meaning they use these tools regularly and understand their limits. That is the group worth learning from. Ask them, specifically, where the agent’s output needed a real correction versus a cosmetic one, and use that answer to calibrate how much oversight the rest of your team needs. Guessing at the right level of scrutiny wastes time in both directions: either burning hours reviewing work that did not need it or missing errors that did.
None of this requires slowing adoption. Atlassian’s Rovo Agents passed 5 million monthly active users this year and automated 2.4 million workflows in six months, evidence that the technology can work at scale when pointed at the right problems. The gap is not in the software. It is in how clearly leaders define who is responsible when the software is wrong.
Where This Actually Leaves You
The managers who get this right in the next year will not be the ones who adopted the most agents the fastest. They will be the ones who kept a clear line of accountability running through every task an agent touches, treated governance as part of the job rather than a compliance checkbox, and stayed honest with their teams about what the technology can and cannot be trusted with. The tools are already here. What is still being figured out, one team at a time, is who is actually in charge when they are running.