When the AI agent joins the org chart

The AI agent on your team doesn’t have a favorite lunch spot or worry about vacation days. It doesn’t get distracted in meetings or take sick leave. But it does handle assignments, and increasingly, it handles them well enough to force a rethink of what work looks like.

Across industries, these systems are stepping into roles that once belonged to junior staff. They’re summarizing research, drafting reports and parsing financial data. No longer just passive tools, they’ve become embedded contributors that demand oversight, evaluation and strategy.

As AI agents spread through companies, managers face a new challenge: how to lead a team that includes non-human workers. These systems may be tireless and efficient, but they don’t ask for feedback, they don’t explain their reasoning and they don’t raise their hands when things go wrong. The question is no longer whether AI can contribute to real work, but how companies will manage accountability, reliability and performance when it does.

“Managers will need a plan to manage performance and make decisions based on the actions of both AI and humans,” Mindy Shoss, a Psychology Professor at the University of Central Florida, told IBM Think in an interview.

At IBM, the transformation is already well underway. With tools like watsonx Orchestrate and internal research on multi-agent collaboration, AI agents are becoming embedded in core business processes. In human resources, they match resumes to job descriptions and generate shortlists for hiring managers. In procurement, they scan contracts to classify vendors and flag compliance risks. While people still make the final decisions, the manual legwork is increasingly happening in the background.

IBM Research has been piloting teams of agents to tackle complex projects, such as large-scale document analysis. In one experiment, thousands of regulatory policy documents were analyzed to extract clauses and compare jurisdictions. The system delivered a matrix of results for analysts to review, a task that once took weeks, now compressed into hours with human oversight.

“AI agents will help us build faster, work with smaller teams and bring more ideas to life,” Kunal Sawarkar, a Distinguished Engineer at IBM, told IBM Think in an interview. “They take the grunt work off our shoulders so we can focus on what’s valuable. It’s a powerful shift where everyone can become a creator, not just an executor.”

Even with these gains, thorny questions remain about trust, supervision and liability. Mistakes happen. When an agent misreads a prompt or overlooks a variable, someone has to be accountable.

Stephen Casper, a researcher at MIT who focuses on AI reliability, views the current moment as a transitional phase, rather than a settled future. “If AI systems get reliable enough to allow this, that could be a big opportunity, but also a very big liability,” he told IBM Think in an interview. “Making something like this a reality would be a long process of gradual adoption, trial and error.”

Rather than envisioning a future of AI peers, Casper sees a slower, bounded integration. Systems may grow more capable, but they will continue to operate under human instruction. “Some of the most impactful applications of autonomous AI, such as driving or coding, happen in a setup where a human is in control but able to delegate tasks to the AI system,” he said.

Delegation may expand, but discretion will remain human. “Even if it’s possible for AI systems to do 90%  of the work that humans do in some jobs,” he added, “that last 10%  is going to be by far the hardest to automate.”

AI agents are already hard at work. Behind the scenes at logistics firms, AI routing agents are analyzing real-time weather, traffic and fleet data. When a storm forms or a highway shuts down, the agent suggests reroutes and reassignments before the human dispatcher has poured a second cup of coffee. The final call still belongs to a person, but the first alert often comes from code.

Organizational structures may follow suit. “Humans would be evaluated in a similar manner to how leaders at higher levels of the organization's hierarchy supervise leaders at lower levels,” Shoss said. AI agents, while not employees, still generate work that must be managed. “Accountability doesn’t mean the same thing with AI as it does with humans,” she added.

Casper agreed that even in an agent-rich workplace, oversight won’t disappear. “We are much further away from AI coworkers than lots of demos and hype would suggest,” he said.

Managing people was hard enough. Now, with agents in the mix, enterprise leaders are reimagining the entire playbook.

“Leaders should adopt a dual-track mindset: be cautious in implementing key operational changes to avoid disruption, but aggressive in experimentation to rapidly explore the potential of AI agents,” said Tianyi Peng, a Professor of Business at Columbia University, in an interview with IBM Think. “The goal is not mass replacement,” he added. “It’s progressive integration.”

Measured deployment helps avoid what Peng calls the backlash trap. “Long-term trust and performance depend on measured integration and employee engagement,” he said.

As agents proliferate, management practices are evolving. “Be a fast learner, not an expert,” Peng said. Technical know-how isn’t required, but fluency in prompting, interpreting and refining is.

“Managing a hybrid team means orchestrating workflows where humans and agents collaborate,” he said. That means understanding capabilities, routing tasks strategically and knowing when to override or retrain.

Metrics are shifting, too. “Performance includes cost-effectiveness,” Peng noted. That means matching model complexity to the job—using lightweight systems for summaries and reserving more powerful models for high-value decisions.

“Teams should track agent ROI, match model size to task and optimize workflows not only for accuracy, but for scalability and value,” he said. “Even if AI makes decisions, humans design the system and sign off on its outputs.”

Looking ahead, Peng said, companies hoping for a clear playbook will need to accept a more fluid path forward. “Everyone’s waiting for a playbook,” he said. “But in truth, we’re writing it as we go.”