The UAE is not asking whether AI agents should run government services. It has already decided, and set a clock. Under a directive from the National Committee for the Agentic AI Project — chaired by Mohammad Al Gergawi, Minister of Cabinet Affairs, with oversight from Sheikh Mansour bin Zayed Al Nahyan — federal entities are converting 50% of government operations, procedures and services to agentic AI models within two years. Some, including the Ministry of Cabinet Affairs itself, have set an internal target of 75% (Trade Arabia).
In August 2026, the committee moved from ambition to plumbing. More than 100 officials from ministries and federal entities attended a workshop that launched a unified design guide — the Assistant AI Government Experience Design Guide — for building services around AI assistants (Khaleej Times; u.ae). The same session began the harder work: building the frameworks that will classify which government tasks an AI agent can complete on its own and which it may only recommend for a human to approve.
What the guide actually commits to
The design guide itself is a service-experience document, not a governance charter, and it reads that way. Its principles are consumer-grade: don't make a citizen re-supply information the government already has, keep them informed at every step of a request, and — per u.ae's own description — let AI "ask for your approval when needed" while providing "human support when needed." Khaleej Times reports a mechanism for exceptional cases that hands a request to a specialized employee "without requiring the customer to explain their request again," plus a stated ability for users to object or request corrections.
That is a reasonable customer-experience floor. It is not an accountability framework. Nothing in the guide, nor in any of the committee's public materials to date, specifies which categories of government decisions — a benefits denial, a visa rejection, a fine — an agent may finalize outright versus merely draft for a human sign-off. As one industry analysis of the rollout put it, that classification exercise "is the whole ballgame," because no government anywhere has yet published a defensible method for drawing the line between a task an autonomous system may complete and one it may only recommend (AI News).
The steelman: speed is the point, and guardrails already exist on paper
The strongest case for the UAE's sequencing is that publishing a rigid decision taxonomy before you know which tasks agents actually handle well would freeze the programme in committee. Government IT modernization efforts elsewhere have died exactly that way — years of governance drafting before a single service shipped. The UAE's approach — build the guide, run pilots, let 50 federal entities discover in practice which processes are safe to hand off — generates the operational data that a good classification framework needs. And it isn't proceeding on a blank slate: the non-binding UAE Charter for the Development and Use of Artificial Intelligence (2024) already commits the country to human oversight, transparency and "governance and accountability" as design principles for any AI system, government or private (u.ae), and the UAE's Personal Data Protection Law (Federal Decree-Law No. 45 of 2021) already gives individuals a right to object to solely-automated decisions that produce legal or otherwise significant effects on them. Those are real hooks a liability and appeals regime can be built on quickly, not from scratch.
Why the gap still matters
The problem is sequencing risk, not principle. A right to object under the PDPL is only as useful as the appeal channel behind it, and no appeal channel specific to agentic government decisions has been published. Nor has anything establishing where liability sits when an agent gets a case wrong — whether that falls on the deploying ministry, the AI vendor, or nowhere at all pending litigation. As the two-year clock runs and agents move from pilots into a majority of federal transactions, that gap compounds: the number of citizens who could plausibly be affected by a wrong autonomous decision rises every month the classification and liability frameworks stay unpublished, while the guide's own promise — "human leads, AI enables" — remains an aspiration without an enforcement mechanism attached.
This is not a UAE-specific failure. No government has published a workable answer to what happens when an autonomous system gets a citizen's case wrong, because no government has run agentic AI at this scale before. But the UAE has chosen to be first, and being first on deployment while being last on the accountability architecture is a specific, avoidable choice. The fix does not require slowing the 50% target — it requires the committee to publish the classification methodology and a liability/appeal mechanism on a schedule that runs alongside deployment, not behind it. The Charter's principles and the PDPL's objection right already give it the raw material. What's missing is the wiring between them and an agent's actual output.