When Systems Begin to Act in the United Arab Emirates
From decision support to autonomous government
From answers to actions
Within two years, a significant share of government services in the United Arab Emirates could be executed by systems that do not merely advise, but act. This ambition marks a shift that is easy to underestimate. For years, artificial intelligence has been positioned as a tool—analysing data, generating outputs and supporting human decisions. What is now emerging is something different. Systems are beginning to move beyond recommendation into execution. The change is not only technological. It is structural.
Most AI systems today operate within a familiar boundary. They produce outputs that inform human judgment, while the final decision—and the responsibility that comes with it—remains with the individual or institution using the system.
Agentic AI alters that balance. These systems can interpret goals, break them down into intermediate steps and interact with digital environments to complete tasks. A permit can be processed, a request routed, a decision implemented—without a human intervening at every stage.
What appears as a gradual evolution is, in practice, a categorical shift. We are no longer asking what AI can answer. We are beginning to ask what AI is allowed to do.
When the black box acts
This transition redefines an existing problem.
As explored earlier in this series, many advanced AI systems operate as black boxes. They perform with high accuracy, yet their internal reasoning remains difficult to reconstruct. In advisory contexts, that opacity can be tolerated. When systems begin to act, it becomes harder to defend.
The difference is no longer between a correct and an incorrect answer. It is between a justified and an executed action.
A system that produces the wrong answer can be corrected. A system that takes the wrong action creates consequences that cannot always be reversed.
As autonomy increases, the margin for error narrows—and the cost of opacity rises with it.
Speed as an enabler of action
The rise of agentic systems does not happen in isolation.
As companies such as Mistral AI demonstrate, AI is becoming faster, more efficient and easier to deploy. This acceleration lowers the threshold for implementation, allowing systems to be integrated into workflows at scale and with increasing speed.
Efficiency, in that sense, does more than improve performance. It enables action.
What can be deployed quickly can also act continuously, across multiple domains and with minimal friction. Speed is no longer just a competitive advantage. It becomes a structural condition for autonomy.
Control under pressure
At the same time, the question of control becomes more urgent.
As discussed in relation to Aleph Alpha, explainability offers a pathway toward making AI systems auditable. By exposing how outputs are constructed—through sources, signals and intermediate reasoning steps—these systems allow decisions to be examined and, if necessary, challenged.
Agentic AI raises the stakes. It is one thing to explain a recommendation. It is another to justify an action that has already been executed.
This compresses the window for oversight. Control can no longer sit entirely outside the system. It must be embedded within it.
From tools to actors
What emerges is a shift in the role of artificial intelligence within institutional processes.
AI is no longer confined to supporting decisions. It is beginning to operate as an actor—interacting with systems, triggering processes and executing actions within defined parameters.
This does not eliminate the human role, but it transforms it. The human moves from operator to supervisor, from executor to governor. Responsibility does not disappear. It changes location.
A divergence in approach
The UAE’s strategy reflects a willingness to move quickly, integrating agentic systems at scale within a relatively centralised governance model. The emphasis is on execution, responsiveness and technological leadership.
In Europe, the emphasis has been different. Here, the focus lies on legitimacy, accountability and the conditions under which AI can be integrated into legal and institutional frameworks. The question is not only what is possible, but what can be justified.
This creates a divergence. Not necessarily in ambition, but in sequencing.
One model risks moving too fast, before accountability is fully established. The other risks moving too slowly—and becoming dependent on systems built elsewhere.
The system behind the system
Taken together, the developments across this series form a layered transformation.
The black box exposed the limits of understanding. Speed enabled large-scale deployment. Explainability created the conditions for control.
Agentic systems introduce action. Each layer builds on the previous one. Each increases both capability and complexity.
The question ahead
The next phase of artificial intelligence will not be defined by capability alone, nor by speed or explainability in isolation.
It will be defined by whether systems that act can remain accountable. Because once decisions become actions, the margin for error narrows—and the consequences become immediate.
The question is no longer abstract. Can we build systems that act— without losing the ability to question them?
This article is part of The Black Box Divide, a series exploring how Europe is redefining intelligence, accountability and control in the age of AI.
🎨 Credit
Illustration by Altair Media (AI-generated)
✍️Caption
When systems begin to act, accountability must move with them.
