A very particular moment in time
I keep returning to one question: when we look back five years from now, which decision about AI agents will we wish we had made earlier?
The useful answer is neither panic nor blind enthusiasm. It is to understand what is changing early enough to shape it. Adoption will move at different speeds across industries, but the direction is becoming clear: AI is moving from answering questions to carrying out work.
The copilot is an intermediate step
A copilot waits for a prompt. It helps one person draft, analyse or decide. That can already create real value, but the responsibility for every next step remains with the human.
An agent is designed around a goal. It can use tools, access approved data, remember context and execute several steps in sequence. Multiple agents can divide work, review one another and coordinate a process. Agents can even help build and test further agents.
That shift—from assistance to delegated execution—is why this is not simply the next software upgrade.
What changes when software can act
The important question is no longer only which model produces the best answer. Companies must decide what an agent may see, which systems it may use, which decisions it may prepare or make, and when a human must intervene.
Those are questions about responsibilities, interfaces and control. They touch operations, IT, data protection, information security and leadership at the same time. A collection of promising pilots does not yet amount to an agent strategy.
AI agents are a management responsibility
If agentic transformation is treated as an IT experiment, it will remain an IT experiment. The technology team is essential, but it cannot define business priorities or redesign accountability on its own.
A credible programme needs an accountable business sponsor, the owner of the process being changed, technical leadership, an agent enabler who translates between possibilities and practice, and the relevant governance functions. The aim is not to create a committee. It is to make ownership explicit.
Start with one process—not with a platform
The first lighthouse project should be relevant enough to matter, bounded enough to control and measurable enough to learn from. A process with clear inputs, repeated decisions and visible handovers is usually more useful than a spectacular demo.
Define one accountable leader and one measurable outcome. Measure not only speed and cost, but also quality, error rates, acceptance and the amount of human intervention still required.
A simple management test
Before selecting a platform, a leadership team should be able to answer three questions: Which process outcome should improve? Where is the boundary of the agent’s autonomy? Who remains accountable when the process behaves differently than expected?
If these answers are unclear, the company does not yet have an agent strategy. It has a technology experiment.
The strategic choice
Companies do not have to automate everything, and they should not. But waiting until every uncertainty has disappeared is also a decision—with consequences for learning speed and strategic freedom.
Agentic AI is moving into business practice. The strategic question is whether a company learns to shape that shift deliberately.
In five years, we may not ask why some companies experimented with agents so early. We may ask why so many waited until change was unavoidable.