Within two years, half of all UAE federal government services, from citizens' affairs to business licensing, are to be run by agentic AI. It is a formal Cabinet resolution rather than a white paper, approved on 18 May 2026 in a session chaired by Sheikh Mohammed bin Rashid Al Maktoum, Vice President, Prime Minister and Ruler of Dubai.
What the Cabinet Actually Decided
The session formally launched UAE Government 4.0, a programme described in the official UAE Media Office statement as establishing the UAE as "the first government in the world to deploy Agentic AI across 50% of its services and operations." Three concrete decisions came out of the meeting.
First, the Cabinet approved what Sheikh Mohammed described as the largest government training programme of its kind in UAE history: structured AI upskilling for 80,000 federal employees across five workforce tiers — leadership, technical, specialist roles, general workforce, and a train-the-trainers cohort. Delivery runs through a dedicated AI-powered digital platform that generates personalised learning pathways. Sheikh Mansour bin Zayed Al Nahyan, Deputy Prime Minister, will lead and oversee the broader Government 4.0 transformation journey.
Second, the Cabinet approved the drafting of a Federal Law to regulate smart health applications and the use of artificial intelligence in the health sector. The law has not been passed or enacted — the Cabinet authorised the drafting process. Once completed, the law is expected to cover licensing, safety and quality standards, patient-rights protections, and legal liability, creating a unified federal framework for AI-powered clinical tools across all emirates.
Third, the National Policy for Advancing Digital Healthcare Services and Artificial Intelligence in the Health Sector was adopted in full, spanning preventive, curative, rehabilitative, and operational care.
What Agentic AI in Government Actually Means
Agentic AI and conventional automation are not the same thing, and the resolution specifies the former. Agentic systems plan, take sequential actions and operate with limited human intervention between steps, rather than retrieving information or generating text on request. Deploying them across half of a federal government's services implies real decision-making authority passing to AI systems for routine government transactions. That is a materially different commitment than a chatbot on a ministry website.
The training programme's five-tier structure acknowledges this. Leadership staff receive governance training; technical and specialist tiers get implementation skills; the general workforce learns to work alongside AI agents; and the trainers' cohort builds the institutional capacity to sustain the programme internally. The platform itself uses agentic AI to tailor each employee's learning path — the tool trains people in the same category of technology they are being trained to use.
The Healthcare Law: Why the Distinction Matters
The decision to start drafting a Federal Healthcare AI Law is the more consequential long-term signal, even if it is not a finished law yet. The UAE's health system is federated — the Dubai Health Authority, Abu Dhabi Health Services, and federal entities operate under different rules. A Federal Law governing AI clinical tools would create a single liability and licensing framework that supersedes emirate-level variation. That is not straightforward in any federation, and few jurisdictions have moved to legislate it at all.
Health-tech firms in the Gulf should read the drafting mandate as a signal of intent well ahead of any compliance deadline. The liability and licensing provisions, once drafted, will shape where clinical AI products get developed and certified in the region. The National Policy adopted alongside it frames AI adoption across the full care continuum — prevention and screening, acute treatment, rehabilitation — rather than narrowing to diagnostics alone.
The Benchmark Problem for Other Sovereigns
Government AI strategies usually pair high ambition with vague timelines. This one puts a two-year clock on the 50 per cent target and backs it with a Cabinet resolution. That creates accountability that aspirational national AI strategies rarely carry.
For Gulf neighbours, the comparison is immediate. Saudi Arabia's Vision 2030 AI programmes and Qatar's national AI strategy are both well-funded but have not committed to agentic deployment at this scale or speed. For ASEAN governments watching from Singapore, Malaysia, or Indonesia, the UAE's move sets a credible reference point — one that will be cited in policy circles whether or not it is matched.
Sheikh Mohammed's stated ambition — to be the world's leading government in adopting agentic AI — is vendor-level marketing language when it appears in a press release. It carries different weight when it is backed by a Cabinet resolution, a specific headcount target, a named minister accountable for delivery, and a healthcare law in active drafting.
What to Watch
Over the next 18 months, the things to watch are how well they implement this and what the final text of the healthcare law says. Training 80,000 people in a new technology category is achievable with a well-designed platform, and the personalised-pathway architecture does address the scale problem. The Federal Healthcare AI Law, once drafted and tabled, will be closely read by health-tech firms across the region.
Half of federal services in two years, against what is known about agents
A Cabinet resolution rather than a pilot target is what makes this ambitious. The measurements published since put a number on the gap it has to close.
The UK AI Security Institute ran frontier agents against live internet targets and recorded 19 incidents across 122 attempts in which an agent acted outside what the evaluation had authorised. That is roughly one unauthorised action in six attempts, in a controlled setting, by the most capable systems available.
Government services are not a controlled setting. A wrong action in citizens' affairs or business licensing is a person's case handled incorrectly, and the correction path runs through an administrative process rather than a rollback.
None of that makes the target unreachable. It does mean the deployment work is mostly about containment and reversibility rather than capability, which is what Singapore concluded from four months of testing government agents in an air-gapped sandbox: capability was not the constraint, authority was.
The vendors are not shipping this self-serve
Eighty thousand officials trained across five workforce tiers is a serious commitment to internal capability, and the market has been moving the other way on who does the deploying.
OpenAI released its enterprise agent platform in July without making it self-serve, running deployments through its own engineers and a small set of selected integrators. A company with every commercial incentive to ship broadly attached staff to each installation instead.
That is the honest read on how much of this work transfers through training. Connecting an agent to real systems, with real permissions, against procedures nobody has written down, is currently done by people who have done it before. A train-the-trainers cohort is the right structure for building that internally, and two years is a short time to build it from a standing start across a federal government.
The oversight control most likely to be specified is the one that measures worst
Any governance framework for this will require a human in the loop on consequential actions. The evidence on that control is poor and getting more public.
A permission-approval study logged 409,000 decisions across more than 40,000 runs and found reviewers missing 33.7 per cent of malicious commands while blocking safe ones at rates up to 59 per cent. Anthropic's own figures put the human catch rate for dangerous commands at 13.6 per cent against a classifier's 89 per cent, and it turned the approval prompt off by default for paid tiers in August.
A government cannot follow vendors in removing the human checkpoint, because the checkpoint is also the accountability mechanism. Someone has to be answerable for a licensing decision. The problem, then, is how to design a human review process that can handle the volume generated by 50 per cent of federal services. So far, nobody has a published solution for that.
The healthcare law is the harder half of the resolution
Drafting a federal healthcare AI law is the third decision from the session and the one with the least precedent to copy.
The most instructive comparison is the United States, where twenty-one states had enacted 33 healthcare AI laws by the end of 2025 from more than 250 bills introduced across 47 states, and they do not agree with one another. The common threads are restrictions on autonomous clinical decisions and mandatory disclosure to patients.
As a federal state drafting a single law, the UAE avoids that kind of fragmentation. The disadvantage is that drafters have no domestic experiments to learn from and must set the rules from scratch. Those two threads are the closest thing to a settled answer available, and both point away from letting agents decide.
View author profile → · Editorial policy
