Artificial intelligence is entering a new phase. For years, businesses have largely treated AI as a tool for prediction, recommendation and decision support. But a new generation of AI systems can do more: pursue objectives over time, take actions, learn from feedback and coordinate with people and other AI systems.
In Decision Science in the Agentic Era: Research Frontiers on Agentic AI and Human–Agent Collaboration, published in Decision Sciences, Sun Tianshu, Dean’s Distinguished Chair Professor of Information Systems at CKGSB, and Yingjie Zhang of Peking University’s Guanghua School of Management, argue that this shift marks the beginning of an “agentic era” for decision-making. Their central argument is that when AI moves from supporting decisions to participating in them, businesses need to rethink how decision-making itself is designed.
Agentic AI is not defined simply by the use of a large language model. Sun and Zhang focus instead on the role AI plays within a decision system. An agentic system can initiate actions, adapt through experience and participate continuously in decisions, rather than simply responding to a prompt or producing a one-off recommendation.
This distinction matters for companies. AI agents can already handle administrative processes, interact with customers, coordinate marketing activities and support complex decisions in areas such as finance and healthcare. As their autonomy grows, the question shifts from “How can AI help employees make better decisions?” to “How should decisions be organized when both humans and AI can act?”
One of the paper’s main contributions is a framework for understanding these emerging systems. Sun and Zhang break them down into three connected layers: atomic structures, decision architectures and field-to-model mapping.
At the foundation are actors, flows and moderators. Actors may be employees, customers, managers or AI agents. Flows describe how tasks, information and control move between them. Moderators include incentives, organizational norms, accountability requirements and risk thresholds.
This framework helps explain why organizations using similar AI technology can produce very different outcomes. Who has access to information? When should an AI agent hand a decision to a person? Who can override whom? What encourages employees to trust, question or ignore an agent’s recommendation? These choices can matter as much as the underlying technology.
The paper compares customer service and healthcare to illustrate the point. Both may use AI to process information before a human reviews the output. Yet human oversight in customer service may be driven mainly by efficiency, while healthcare requires greater attention to risk, uncertainty and accountability. Similar technologies can therefore require very different decision architectures.
The rise of autonomous agents does not remove humans from decision-making. Instead, it changes where human judgment adds value. People may increasingly define objectives, provide contextual knowledge, set constraints, resolve trade-offs and oversee outcomes rather than execute every individual decision.
This also makes accountability more complex. In systems involving multiple adaptive agents, outcomes can emerge from interactions among humans, AI, incentives and feedback loops. Governance therefore becomes a design problem. Organizations need clear escalation rules, override rights and monitoring mechanisms that determine when people should intervene and who remains responsible for outcomes.
Sun and Zhang also set out a broader research agenda for decision science. They call for more work on AI agents as decision-makers in their own right, including how agents learn, coordinate and generate system-level outcomes. Human–agent collaboration is another major frontier: researchers need to understand not only how humans influence AI, but how adaptive agents influence human judgment, trust and behavior over time.
The paper also points to a new role for AI in research itself. Agents could help generate hypotheses, explore large design spaces and run simulations, while human researchers continue to define meaningful questions, interpret results and ground findings in theory and values.
For companies, the key implication is that adopting agentic AI is not simply an IT decision. It is an organizational design decision.
As AI systems gain the ability to act and adapt, managers will need to reconsider decision rights: what should remain human, what can be delegated, when an agent should escalate a case and who is accountable when decisions emerge from a chain of human and machine actions.
The central question is therefore no longer simply what AI can do. It is how organizations should be designed when intelligent agents become active participants in the decisions that create value.
View the full paper here.
