Everyone works in processes. Far fewer people own them.
That distinction has become harder for me to ignore as our thinking about AI has matured. When a process is mainly carried out by people, unclear ownership can remain hidden for years. Individuals make sensible local decisions, experienced staff repair broken hand-offs, and meetings compensate for the absence of an obvious decision-maker. The work gets done, so the design weakness is tolerated.
Agentic AI removes some of that tolerance. Before an agent can act, someone must decide what it is allowed to do, which information it may access, how its performance will be measured, where it must escalate and what happens when it is wrong. Those are not technology decisions. They are ownership decisions.
I used to think process ownership was mainly a governance requirement. I now see it as a precondition for learning. If nobody owns the process, nobody is accountable for understanding whether it performs well. Improvement becomes project-based rather than continuous. Problems are discussed at the edges, but the end-to-end result remains unmanaged.
Ownership does not mean one person performs every activity or approves every change. It means someone is accountable for the outcome of the process and for the quality of its design. They know the measures, understand the controls, see the upstream and downstream dependencies, and can explain why the work exists.
The difficult part is that organisational charts rarely show processes. They show functions, roles, and reporting lines. Most important outcomes cross those boundaries. That is why process ownership can feel awkward. It asks someone to care about work they do not directly control and to influence people who do not report to them.
My current view is that authority should match the blast radius. A process owner should be able to approve improvements contained within their process and agreed guardrails. If a change affects upstream inputs, downstream outcomes or shared controls, it should move to a value-stream forum where the people with genuine stakes can decide together. If consensus fails, a named value-stream owner must make the call.
AI did not create the need for process ownership. It made the absence of it impossible to ignore. The first question I now ask about any proposed agent is not what model it will use. It is simpler: who owns the work, by name?