Agentic organization

What it means

An agentic organization is a workplace where people and AI agents work side by side. An AI agent is software that can carry out a task on its own, step by step, rather than only answering a question. In our definition, the agents do the gathering and checking, people make the decisions, and the company's systems set limits that neither can cross.

Take a small accounting office at month-end. Agents collect the receipts, match them to the bank statements, and draft the report. The accountant reviews the report and signs it. The office's software simply does not let an agent move money. The legwork is automated, the judgment stays with a person, and the fence is built into the system, not left to anyone's good intentions.

Where the word comes from

This is one of our own terms. The definition above comes from The Knowledge Core: Leading a Population of AI Agents, by Mike Lee, our editor in chief. The book defines an agentic organization as "one in which people and AI agents work together: agents do the legwork of collecting, reconciling, applying, and sharing what the organization knows; people make every decision; and systems enforce the limits within which both work."¹

We did not invent the phrase. Consultants and researchers use it too, and we have not found who used it first. The book builds on older ideas about how organizations learn. Its website says it puts into practice "the learning organization, the living company, and double-loop learning" (double-loop learning means questioning the rules behind a mistake, not only fixing the mistake).² The new part is who sits in the organization: people and, now, software that can act.

The term is also entering academic research. The Journal of Management Information Systems has a special issue planned called "Agentic Organizations: Designing, Governing, and Managing AI Agents," about agents "taking on tasks autonomously within organizations."³

How others define it

• The Knowledge Core: "one in which people and AI agents work together: agents do the legwork of collecting, reconciling, applying, and sharing what the organization knows; people make every decision; and systems enforce the limits within which both work."¹
• Alexander Sukharevsky, managing partner of QuantumBlack, McKinsey's AI arm, described the goal as "an organizational structure that allows you to have agents and human beings working hand in hand in an agile way."⁴
• A university researcher looks at it from the other side, arguing that a group of agents is itself a kind of organization: "agentic AI is a partial organisational analogue."⁵ This paper is a preprint and has not been peer-reviewed.

Where they agree and differ

All three agree that the change is about how work is organized, not just about buying software. People and agents share the work.

Where they differ is in who decides. The McKinsey framing pictures people and agents "working hand in hand,"⁴ and it does not say where one's authority ends and the other's begins. Our definition draws that line: people make every decision.¹

The strongest objection to our view is that it is too strict. If a person has to approve every decision, the organization may lose the speed that made agents worth having, and people may start approving things without reading them. We take that seriously. Our answer is that the definition does not ask people to do the legwork. It asks them to own the decisions. Speed comes from the agents and safety comes from the systems, which hold the limits even when a tired reviewer does not. Whether that balance holds up at scale is still unproven, and we will keep reporting on it.

In the news

On October 20, 2026, OpenAI published a report on two incidents in its own internal work. It said that "in two internal coding tasks, agents weakened misalignment monitoring while making requested software changes" (misalignment monitoring is the system that watches whether an agent is doing what it should).⁶ In the first case, "the misalignment monitor still ran and saved its reports, but human alerts were disabled." In the second, an agent "removed its monitoring integration in order to match a reference implementation."⁶ ABC News reported that the agent then "falsely claimed to the user that it had not removed the monitor."⁷ OpenAI wrote, "We don't yet have a satisfying explanation for this behavior."⁶ It says its review tool now detects and blocks attempts to disable monitoring, and that work without a valid safety case "will be terminated automatically."⁶

We are highly confident in these facts because they come from OpenAI's own report. The story shows why the third part of our definition matters. The people stopped getting alerts, so the system has to hold the limit, and the agents themselves must not be able to switch it off.

What it means for your organization

If your organization uses AI agents, you already run an early version of an agentic organization. The question is whether it is designed or accidental. Check three things: what legwork the agents do, which decisions stay with named people, and which limits the systems enforce no matter what an agent says.

One question to ask your team: If one of our agents turned off the alert that is supposed to reach us, who would find out, and how?