For most of its history inside institutions, AI has been an adviser. It surfaced a recommendation, scored a risk, drafted a document, and a human decided what to do with it. The accountability question had an easy answer, because a person always stood between the machine's output and the institution's action. That person decided. That person answered.
Agentic AI removes the person from that position. An agent does not surface a recommendation and wait. It plans, it acts across systems, and it completes the task. The human who would once have stood in the decision moment is now somewhere upstream, having set the agent loose, or downstream, discovering what it did. The seat where accountability used to sit is empty at the moment that matters.
An AI agent does not recommend a decision and wait for approval. It takes the action itself. That single shift, from advice to action, is what turns accountability from a question an institution can answer after the fact into one it must answer in advance.
This is why agents compound accountability debt faster than any tool before them. A recommendation engine that is wrong produces a bad suggestion a human can catch. An agent that is wrong produces a completed action, already taken, on real systems, on the institution's behalf. The window in which a human could have intervened has closed by the time anyone looks.
The familiar answers to "who is responsible" all fail here, each for a different reason. The user is responsible, except the user did not make the decision; the agent did, autonomously, often across data and systems the user never directly touched. The vendor is responsible, except the vendor built a capability, not the specific action, and no regulator holds a software supplier answerable for an institution's operational decision. The model is responsible, except a model is not a person, and accountability is a property of people.
What is left is the work institutions keep deferring. Before an agent acts, three things have to be settled. A named human must own the agent's mandate, answerable for what it is permitted to do. The agent's reach must be bounded, so it can act only on what that owner has authorised and nothing beyond it. And every action the agent takes must leave a record a third party could follow, showing what it did and on what basis. None of this can be assembled after an agent has acted. It is either in place before, or it is missing when it is needed.
This is no longer abstract in the Gulf. Governments here are mandating agentic AI at the scale of entire administrations. The UAE has committed to moving tens of thousands of public-sector workers onto agentic tools and half of government services onto autonomous systems. The agents are arriving by directive, on a clock. The accountability for what they do is not arriving with them, because it cannot be mandated centrally. Each institution has to assign it, agent by agent, before the agent goes to work.
The question an institution should ask of every agent it deploys is not whether it works. It is who answers when it does something no one would have chosen. The institutions that can name that person before the agent acts are governed. The ones that cannot are merely automated, and exposed.