The United Arab Emirates plans to hand over half of its public services to autonomous AI systems within two years, forcing a reckoning Europe cannot ignore. The shift from machines that recommend to machines that decide marks a fundamental rupture in how states wield power.
For decades, public debate about artificial intelligence has revolved around what algorithms can predict, generate, or suggest. That conversation has outlived its usefulness. The next frontier concerns authority itself: who commands, who escalates, and who bears responsibility when automated systems act.
The UAE’s federal roadmap commits to migrating at least 50% of state operations to agentic AI, backed by sweeping civil service retraining. Yet the real significance lies not in the technology but in its bureaucratic placement. Algorithmic intelligence now sits precisely where analysis converts into execution.
European regulators have built sophisticated frameworks through the AI Act, covering risk mitigation, fundamental rights, and civil liberties. However, agentic AI poses an operational problem that legislation alone cannot solve. Once an automated system gains technical authority to trigger multi-stage administrative actions across official registries, institutions must verify not only legal compliance but also sound engineering of competencies and boundaries.
Consider a citizen entitlement claim. The system queries databases, evaluates statutory criteria, requests missing documents, and adjudicates the file without human intervention at each step. Efficiency gains prove undeniable. But when contradictory datasets or atypical circumstances emerge, processing speed becomes irrelevant. The crucial question: should the agent execute, suspend, or escalate?
Operational experience in critical infrastructure demonstrates that data awareness and command authority constitute distinct competencies. A system may possess comprehensive situational awareness while lacking legal mandate to decide. High-reliability operations reconcile these dimensions through explicit escalation protocols. Public governance requires identical discipline.
Europe’s digital public sector must embed clear command boundaries into service architecture, not hide them in source code. Agents need unambiguous awareness of which transactions they can finalise autonomously, which require human validation, and which demand immediate transfer to accountable officials.
This safeguard matters most across the Mediterranean, where shared workflows span migration data, WTO-monitored logistics, energy interconnectors, and customs processing. Autonomous systems will eventually interface across sovereign borders. An algorithmic error could cascade through international networks with unprecedented speed.
The answer lies not in halting modernisation but in building technical reversibility, procedural auditability, and hierarchical escalation into administrative systems from the start. The states that lead the next phase will not be those deploying the most autonomous models. They will be those answering one elementary question with absolute precision: when the agent encounters uncertainty, who assumes command?
