There is a question that comes up more and more often in conversations about AI in companies, and it generates very revealing answers: how many AI agents do you have running today? Not in a pilot, not in a demo, not in a test environment. In production, doing something real, autonomously, every day.
The usual answer is silence, or some version of "we're exploring." That is not a problem in itself, but it does say something important about where the company actually stands.
What's interesting is that the answer varies a lot depending on which platform the organization uses. Companies running Microsoft 365 Copilot or Google Workspace enterprise suites tend to have something closer to a centralized inventory: agents run within the ecosystem, there is at least a visibility layer showing what was enabled and who has access to what. Something similar applies to enterprise plans for Claude or ChatGPT, which include administration controls and usage visibility. But the reality for many organizations is different: teams that adopted AI tools individually, with personal or small-team plans, without centralized policies, and where no one has a consolidated view of what is running and with what access. That is shadow AI in its most concrete form, and it is more common than anyone admits out loud.
This week, xAI did something that illustrates well the direction the industry is heading: it launched Grok Bot Galaxy, a live stream where they show how to run an entire company's processes using agents. It is not an edited video or a slide presentation — it is live, with agents working in real time. The message that sends is not "AI can do everything"; it is that the major players are already normalizing the idea that a company can operate with autonomous agents making decisions and executing tasks continuously. This is not a lab experiment. It is a public demonstration of what is coming.
The number matters less than the ability to answer with precision.
If the answer is zero, at least it is honest and useful: the company knows it has nothing in production yet. If the answer is "one, in the client onboarding process, running for three months," that also says something concrete and valuable. The problem appears when the answer is "it depends on how you define it" or "we have several things with AI but I'm not sure if that counts as an agent." That answer reveals a lack of clarity about what is actually happening, and that lack of clarity has practical consequences when something goes wrong or when someone on the board asks about the return on AI initiatives.
Having one agent running well, even just one, changes the conversation inside an organization in a way no course or conference can replicate. Because when something real is in production, questions arise that never come up in theory: who reviews the cases the agent could not resolve? What happens when the external system it connects to changes? How do we know if the agent is making good decisions or just fast ones? Those questions build organizational capability. And that capability is what separates organizations that are genuinely maturing in AI from those still in perpetual exploration mode.
This is not an argument for deploying agents at any cost. A poorly designed agent, with broad access to critical systems and no adequate oversight, is worse than having none at all. What is urgent, especially as the industry increasingly normalizes the use of autonomous agents, is that governance and adoption lead the technology rather than trail it. Knowing how many agents the organization has running, with what scope, with what owners, and with what oversight criteria is not a bureaucratic exercise. It is the minimum foundation for operating with judgment in an environment where models advance week after week.
The question "how many agents do you have running today?" is not pressure to have more. It is an invitation to know exactly where the company stands. Because AI maturity is not measured by how many tools were enabled or how many workshops were held. It is measured by how much of what is said to be happening is actually in production, with clear owners, real oversight, and measurable results.
That is the difference between talking about agents and having them.
