Fifty Percent by 2028: Inside the UAE's Assistive AI Mandate

A federal Cabinet mandate now requires half of UAE government operations onto agentic AI within two years. We've spent the past several months building toward it alongside a federal entity, and this is the first in a series on what that actually takes. This one is the setup: what the mandate says, what it asks of an entity in practice, and a first look at where we are.

On 23 April 2026, the UAE Cabinet put a number and a date on something most governments only talk about in aspirational terms. Fifty percent of federal government sectors, operations and services move onto agentic AI within two years, which puts the working deadline at roughly April 2028. His Highness Sheikh Mohammed bin Rashid Al Maktoum, Vice President, Prime Minister and Ruler of Dubai, revealed the framework under the directives of the President. Implementation is overseen by His Highness Sheikh Mansour bin Zayed Al Nahyan, with Mohammad Abdullah Al Gergawi, Minister of Cabinet Affairs, chairing the taskforce that follows execution entity by entity.

Here is the part that tends to surprise people outside government: the Cabinet has said explicitly that ministers, directors general and entities will be evaluated on their ability to adopt this transformation, their speed against the new standards, and their mastery of the tools. That is not a soft aspiration sitting in a strategy deck. It is a stated basis for how leadership performance gets assessed. When a mandate carries that kind of weight, the conversation inside an entity changes fast, and we have watched that happen in real time.

A new Federal Authority for Artificial Intelligence and Data was stood up on 14 June 2026, chaired by Omar Sultan Al Olama, to hold the national picture together. Roughly 80,000 federal employees are being trained across five tracks, from leadership down to general workforce. A first cohort of federal agents, in procurement, tax auditing, customer happiness and technical support, was already live by May 2026. This is not a pilot programme still finding its feet. It is running, on a public clock, with public accountability attached to it.

What the framework actually asks of an entity

The Cabinet's target sits inside an operating framework issued by the Ministry of Cabinet Affairs, and it is considerably more specific than the headline number suggests. It organises the work into five tracks: operations and institutional support, strategic government work, government services, capacity building and training, and the AI technology and data foundation underneath all of it. Every federal entity has to inventory what it does, decide what is genuinely fit to move onto agentic AI using a defined set of criteria (transaction volume, complexity, how ready the underlying data and processes actually are, and expected impact), and commit to a two year plan against that inventory.

The delivery calendar is where the pressure shows up. After an initial assessment phase, entities move through six launch batches, each running ninety days, each one requiring at least one new AI enabled service or operation to go live. Miss a batch and there is no quiet extension. The final phase, running to May 2028, is expansion: taking what worked and pushing it further across the entity.

What stands out once you read the framework closely, rather than the headline, is how much of it is a compliance and governance exercise wearing an AI label. Entities are ranked against defined criteria: customer satisfaction, error rates, delivery time, cost reduction, and whether the line between what the AI does and what a human decides has actually been respected. Results get published. The best performers get named. That is a very different kind of pressure than "go build some AI features."

A standard hiding inside the mandate

There is a second document most people miss entirely, and it matters just as much as the headline deadline. Alongside the transformation framework, the UAE has a citizen experience standard that sets out what an AI assisted government service is actually supposed to feel like from the other side of the screen. The short version: the citizen states the outcome they want, not the form they need to fill. The assistant carries the load of pulling data and coordinating between systems, rather than the citizen doing it by hand. Conversation is the front door, not a fallback after the buttons fail. The citizen stays visibly in control of what gets shared and can stop at any point. And the experience has to survive a citizen closing the tab and coming back three days later, picking up exactly where they left off.

That is a considerably higher bar than adding a chatbot to an existing form, and it is the subject of its own post later in this series, because it deserves the space.

What this looks like from where we're sitting

For the past several months we have been inside this exact mandate with a federal entity, not reading about it from the outside. We picked one heavily used citizen facing service as the first target, inside a ninety day cycle, because a public facing service under a tight deadline forces honesty about what actually works and what does not.

Picking a model sounded like it would be the hard part going in. It genuinely was not, not in the end. What took the real time was proving, in a way that would hold up to a regulator's questions, exactly what a model was and was not allowed to touch. Data residency here is not a checkbox. It is a design constraint that runs through every layer, from where the model physically sits to how a conversation gets stored to what an audit log has to be able to show, on demand, months later.

None of that makes for an exciting demo video. A consent screen and an audit trail will never look impressive on a big screen at a conference. This far into the build, they are, honestly, most of the actual product. The model does the visible work. The governance layer around it is what makes a federal entity able to say yes.

The first service is close to live now. What used to take a citizen the better part of a working week to get through now moves in a matter of days, and the officers on the other end are spending their time approving a decision rather than assembling the case for one from scratch. A human still makes the final call on every single one. That part does not change, and it should not.

What's coming next in this series

Four more posts follow this one, each on its own timeline. The citizen experience standard in full, with a real walkthrough of what it changes in a service design. How we actually shortlisted a model under a hard data residency constraint, and what we ruled out along the way. The governance layer itself, classification, consent, the audit trail, because that turned out to be where the real engineering effort went. And a plain, human account of what changes for a citizen and for an officer, before and after, once one of these services goes live.

Frequently asked questions

If you are inside a federal entity working through your own version of this right now, we would genuinely like to compare notes. Follow along for the next post in this series, on the citizen experience standard the mandate is quietly measuring everyone against.

Syed M. Umair is Head of Product and Digital ventures at Microvision, where he is currently working with a federal entity on its path through the UAE's assistive AI mandate. This is the first in a six part series on what that work actually involves.  Connect with Syed on LinkedIn