AI Agents in Government: UAE, Singapore, Estonia 2026

Three governments are pushing AI agents into the public sector in 2026. The UAE is converting 50% of federal services to agentic models within two years, Singapore is registering agents for 150,000 public officers, and Estonia plans digital IDs for agents. None has published the rule deciding which of its own decisions an agent may make.

TL;DR

The compute is procured, the training is funded, and the hard part is still unwritten.

• The UAE started a two-year clock on April 23, 2026. Sheikh Mohammed bin Rashid Al Maktoum announced a framework to transform 50% of government sectors and services to agentic AI "for autonomous execution and decision-making."¹
• The Cabinet approved the implementation framework on May 18, 2026, alongside a Phase One service list covering citizens', residents', business sector, and general public services.²
• 80,000 federal employees are being trained across five occupational categories, from ministers to new joiners — described by the Cabinet as the largest training programme in UAE government history.²
• FedAI is the shared platform. The Federal Agentic AI Technology Platform, presented at a workshop reported on August 17, 2026, routes entities through a National AI Gateway to commercial and open-source models hosted on sovereign infrastructure.³
• Singapore went the opposite direction: inventory first. GovTech is building a registry tracking who owns each agent and what it does, with guardrails that block file deletion and external email.⁴⁵
• Estonia is solving identity. On June 17, 2026, an advisory board backed by Prime Minister Kristen Michal agreed to create "AI ID codes" so an agent gets scoped, auditable permissions instead of inheriting its owner's access.⁶
• China already published a decision-authority taxonomy — three tiers of agent decisions, issued May 8, 2026 — but it governs agent developers, not which state functions may be delegated.⁷
• The UAE's task classification frameworks were discussed in August 2026 workshops but have not been published.⁸⁹

What You'll Learn

• What the UAE actually announced, and the difference between the April, May, June, and August milestones
• What FedAI and the National AI Gateway are, architecturally
• Why task classification is the unsolved part of every national agent programme
• How Singapore's registry approach differs from the UAE's transformation approach
• What Estonia's "AI ID codes" are trying to fix
• Which government-scale agent frameworks have actually been published, and what each one covers
• What the two-year clock means if you build agent systems for regulated buyers

What the UAE Actually Announced

The dates matter here, because coverage tends to collapse four separate events into one.

April 23, 2026 — Sheikh Mohammed bin Rashid Al Maktoum chaired the Cabinet meeting that revealed the framework. His quote sets the ambition level: "Our government will be the first government in the world to largely deploy Agentic AI models across its government sectors and operations."¹

He described the target as transforming 50% of UAE Government sectors and services within two years "to Agentic AI for autonomous execution and decision-making."¹

Sheikh Mansour bin Zayed Al Nahyan oversees implementation. A dedicated taskforce is chaired by Mohammad Abdullah Al Gergawi, Minister of Cabinet Affairs.¹

May 18, 2026 — The Cabinet approved the governance framework that assigns roles to each ministry and federal entity, plus the Phase One list of services to transition, covering four categories: citizens' services, residents' services, business sector services, and general public services.²

The same session approved the training programme: 80,000 federal employees across five occupational categories — leadership, technical, specialist, general workforce, and train-the-trainers — delivered through a dedicated platform with personalised learning pathways.²

June 2026 — A Dubai training event drew more than 300 participants from 50 federal institutions, with a stated goal of identifying which services and procedures suit agentic AI within 90 days.⁸ A separate Dubai workshop run by the Presidential Court and the Ministry of Cabinet Affairs drew 600 staff from both bodies.⁸¹⁰

August 10, 2026 — More than 100 federal officials attended a workshop launching the strategic track, organised by the National Committee for the Agentic AI Project. It ran under the principle "human leads, AI enables" and covered seven pillars: strategy and projects, foresight and strategic intelligence, policies, structures and governance, government performance, global competitiveness, and innovation in government work.⁸⁹

Huda Al Hashimi, Deputy Minister of Cabinet Affairs for Strategic Affairs, described the session as the operational start of deploying AI models across government activity.⁸

What FedAI and the National AI Gateway Are

In August 2026, reported by Gulf News on August 17, the Higher Committee for Agentic AI organised the "Agentic AI and Data Technology Track" workshop. That track is chaired by Omar Sultan Al Olama, Minister of State for Artificial Intelligence, Digital Economy and Remote Work Applications.³

That workshop introduced the Federal Agentic AI Technology Platform — FedAI — as common technology infrastructure for federal entities building and adopting agentic solutions.³

The architecture has three parts worth noting.

A single access point. Through the National AI Gateway, entities reach a range of AI platforms via one secure entry for managing applications and services.³

Model plurality on sovereign infrastructure. Entities can use commercial models as well as open-source models hosted within sovereign infrastructure, with new models addable later.³

A self-assessment classification mechanism. Officials described a mechanism based on technical and operational criteria to determine whether a system meets the requirements of agentic AI.³

Al Olama said the project represents the world's first comprehensive government-wide initiative of its kind to adopt agentic AI technologies across government.³

That is the government's own framing, and it is worth reading precisely: the claim is about comprehensiveness and government-wide scope, not about being first to deploy any agent anywhere.

One structural detail is worth flagging because coverage is inconsistent about it. Gulf News reports the project running across five tracks — operations and institutional support, strategic government work, government services, technology and data, and capacity building and training.³

Other reporting on the August strategic-track workshop describes seven pillars.⁸⁹ The most consistent reading is that the seven pillars belong to the strategic government track rather than the project as a whole, but no official structural document confirming that has been published.

The Classification Problem

Here is the part that is genuinely unsolved, and not only in the UAE.

Deploying an agent requires answering a question procurement processes were never built to answer: which tasks may an autonomous system complete, and which may it only recommend?

The August strategic-track workshop explicitly covered task classification frameworks alongside timelines and anti-duplication mechanisms.⁹ Covered, not published.

As of August 21, 2026, we could not find a published UAE task-classification framework in public sources — reporting describes it as in development.⁸⁹

One government has published a decision-authority taxonomy, and it is worth naming precisely. China's May 2026 agent rules require developers to sort decisions into three tiers: those the user makes, those requiring user authorisation, and those the agent takes alone.⁷

That is a real answer to the question — but it is aimed at whoever builds and ships an agent product, as we covered in China's AI agent regulations. It does not tell a ministry which of its own statutory functions may be executed without a civil servant in the loop.

There is also a gap between two framings of the same programme. The founding language is about autonomous execution and decision-making without human intervention.¹ The implementation language is "human leads, AI enables."⁸

Both can hold across a portfolio of thousands of tasks. The classification framework is what closes the distance between them, which is exactly why it is the document to watch.

This is the same structural question we covered in agent incident reporting under the SAFE draft — when an autonomous action goes wrong, the framework has to already say who was allowed to authorise it.

Singapore Took the Registry Route

Singapore's approach starts from visibility rather than transformation targets.

GovTech is building a registry of AI agents covering roughly 150,000 public officers, sitting inside its AI Assistant Desk suite. Chief Executive Goh Wei Boon has described the goal as giving every public officer a secure personal digital assistant.⁴⁵

The registry tracks the owners and activities of AI agents, across work from drafting reports to managing schedules to writing code.⁴⁵

The guardrails are concrete rather than aspirational. Reported controls include blocking agents from deleting files, blocking email to external recipients, capping recipient counts to limit spam, and automated checks for offensive language entering or leaving the system.⁴⁵

GovTech has been piloting the suite with some public officers ahead of a broader rollout later in 2026.⁵

Note what this is not: it is not a statement about which government decisions an agent may make. It is an inventory-and-permissions layer for officers using agents in their own work.

Estonia Took the Identity Route

Estonia's contribution is the narrowest and, for builders, possibly the most reusable.

On June 17, 2026, at its second meeting, the Eesti.ai advisory board — established on the initiative of Prime Minister Kristen Michal — agreed that Estonia will move forward with digital identities for AI agents, described as "AI ID codes."⁶

Michal backed the proposal to build a solution letting AI act on behalf of people, companies, or organisations within clearly defined limits, in a way that is verifiable and auditable.⁶

The problem it targets is familiar to anyone who has wired up an agent: today the agent typically inherits the full credentials of whoever runs it.

A scoped agent identity instead makes it possible, in the Estonian government's wording, "to specify whether an AI agent may only view data, prepare a document, draw up a payment or act solely within a specific financial limit."⁶

The Estonian government says this builds on existing digital-government infrastructure including X-Road, digital signatures, and digital footprints.⁶

That is the same design pressure behind Web Bot Auth and cryptographic agent verification: if an agent acts, something has to prove which agent it was.

Three Approaches, Compared

The row that matters is the fourth one. Each government is starting from a different question, and the questions are not interchangeable.

The UAE is optimising for conversion rate. Singapore is optimising for observability. Estonia is optimising for accountability.

A complete national programme eventually needs all three, and none of the three has all three yet.

What Has Actually Been Published

It is worth separating what exists from what is promised, because several real documents already cover adjacent ground.

IMDA's Model AI Governance Framework for Agentic AI launched January 22, 2026, announced by Minister for Digital Development and Information Josephine Teo at the World Economic Forum in Davos.¹¹ IMDA describes it as "first in the world to include a comprehensive guide for enterprises to deploy Agentic AI responsibly."¹¹

It covers four dimensions: assessing and bounding risks upfront, making humans meaningfully accountable via defined approval checkpoints, implementing technical controls across the agent lifecycle, and enabling end-user responsibility.¹¹

Compliance is voluntary, and it targets organisations deploying agents — not the question of which government functions may be delegated.¹¹ IMDA updated it on May 20, 2026 with industry feedback and case studies covering multi-agent systems and third-party agents.¹²

The World Economic Forum's readiness framework for government agentic AI was published on April 24, 2026, produced with the Global Government Technology Centre Berlin and Capgemini. It maps 70 core government functions against two axes: the public value agentic AI could add, and the complexity of deploying it responsibly.¹³

Half of those 70 functions land in the high or medium readiness bands, with high-volume rule-based work — appointment management, document validation, public information provision — scoring best.¹³

The report does invoke "bounded autonomy," being deliberate about what agents are allowed to do.¹³ But the instrument itself answers where should you start, not what may the agent finish in any specific function.

China's Implementation Opinions went furthest on authority. The Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly released the Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents on May 8, 2026.⁷

It requires developers to "clarify the reasonable boundaries and required authority for various decision-making methods," splitting decisions into three tiers: limited to the user, requiring user authorisation, and autonomous.⁷

For "sensitive areas and key industries," it directs regulators to apply filing, testing, and problem-product recall measures. Healthcare, transport, media and public safety are named separately in the document, as fields where mandatory standards should be developed.⁷ Much secondary coverage merges those two provisions into one; the text keeps them apart.

It also proposes exploring an agent registration platform offering digital identity management, discovery, and capability declaration.⁷ So the registry idea and the agent-identity idea are not unique to Singapore and Estonia — three governments reached for them independently inside six months.

One caveat is worth keeping straight, because most coverage does not. "Opinions" (意见) sits below binding "Measures" (办法) in Chinese administrative practice — a guiding policy instrument rather than an enforcement statute with fixed penalties.

The published text carries no effective-date clause, though some legal trackers describe it as taking effect on July 15, 2026.⁷ The rule that did take binding effect that day governs companion bots, not work agents. We untangled the two Chinese rules people keep merging previously.

So the public shelf holds a deployment guide for enterprises, a prioritisation guide for governments, and a product-side decision-tier rule. What is still missing is the version written from the state's side of the table: which government functions may be executed autonomously, and who answers when one goes wrong. The UAE is on a clock that forces it to write that one.

What This Means If You Build Agents

Three practical implications, independent of whether you sell into any of these three governments.

Sovereign hosting is becoming a procurement default, not a premium tier. The FedAI design explicitly accommodates commercial and open-weight models on sovereign infrastructure.³ Our read: if your agent platform assumes a single hosted API endpoint with no self-hosted path, expect that to become a scoring problem in public-sector procurement.

Agent inventory is becoming a requirement. Singapore is building ownership tracking into the assistant layer from the start rather than retrofitting it.⁴ Expect "list every agent, its owner, and its permissions" to appear in RFPs.

Scoped, non-inherited credentials are the direction of travel. Estonia's AI ID proposal and Singapore's per-agent guardrails both point that way, and the UAE is at least building a mechanism to classify what counts as an agentic system in the first place.³⁴⁶ Agents that simply run as their operator will age badly. This is the same containment logic behind restricting agent capability at the deployment boundary.

The Bottom Line

The UAE has done something few governments have: attached a percentage and a deadline to agentic AI in the public sector, and tied the performance assessment of ministers, directors general, and entities to hitting it.¹

That converts an open research question into a delivery deadline. Within two years there will be public evidence about what agents actually do to service delivery in a national government, and everyone else gets to read it without running the experiment.

But the procurement, the platform, and the 80,000-person training programme are the tractable parts. The constitutional part is a document that has not been published: the one that draws a line between a government task a machine may complete and one it may only recommend.

Singapore answered a smaller version of that question for enterprises in January.¹¹ China answered the product-side version in May.⁷ Estonia is answering the identity half.⁶ The UAE is the one that has to answer it from inside the state, at scale, on a clock.

Watch for the classification framework. Everything before it is procurement.

Sources