United Arab Emirates

The state takes the first step. A person confirms the outcome.

An AI-native government is one where the default shape of a public service is: the state notices what a citizen needs, does the work it can already do on their behalf, prepares the outcome, and a person confirms it — while the citizen says what they want in their own words rather than navigating to a form.

A light-filled civic atrium where a public service review is connected by purposeful violet and coral routes.

The distinction

Most governments use AI. Almost none is AI-native.

If you could remove the model and the service would still work as designed, the service is AI-enabled. If removing the model would make the service impossible, the service is AI-native.

Most governments now use AI. There are assistants on portals, transcription in meetings, summarisation in the back office, translation at the counter. This is AI in government. It leaves the service itself unchanged: the citizen still works out which service they need, still gathers the documents, still applies, still waits.

Most governments use AI. Almost none is AI-native.

Today

AI-native

Who starts

The citizen applies

The state offers and the citizen confirms

How they ask

Find the service, then the form

Say what they need, in their own words

Who does the legwork

The citizen collects and submits

The state does what it already can

What the rule is

Guidance a person interprets

Code that executes, tests and explains

What the person does

Assesses from scratch

Confirms or corrects — and is accountable

Who can act

The citizen, through a browser

The citizen, or an agent they authorised

By that test, very little in production anywhere is yet AI-native. That is not a criticism. It is the size of the opportunity.

What it is not

  • Not a chatbot on the portal. Roughly two in five European government portals already offer automated live support. That is the previous generation, at scale.

  • Not a count of deployed use cases. Inventories are growing quickly; most entries are pilots, internal productivity tools, or assistive features on unchanged services.

  • Not a finished state anywhere yet. The gap between the narrative and the volume is universal. Closing it, service by service, is the programme.

The six elements

Six elements constitute it. Governance is one of six, not the whole.

Intent replaces navigation · the state never asks twice · rules execute rather than only publish · the state takes the first step · every service is reachable by a machine · accountability attaches to each decision.

01

Intent replaces navigation

The citizen says what they need in ordinary words, and the state works out which service, rule and process applies.

Requires

a complete, current, machine-readable description of every service the government offers — which most governments do not yet have.

02

The state never asks twice

Anything the state already holds is never requested from the citizen again.

Requires

registers that are right at the granularity the decision needs, and a legal basis for internal data use.

03

Rules execute, not only publish

Entitlement exists as code that can be run against a case, re-tested when policy changes, and explained in ordinary language — not only as guidance a trained official interprets.

Requires

policy and engineering working as one function. That is an organisational change, not a technical one.

04

The state takes the first step

Where the facts are already known and the entitlement is clear, the state offers and the citizen confirms — rather than the citizen applying and the state assessing.

Requires

registers accurate at the right granularity and period, and an explicit rule that adverse determinations are always visible and disputable before they take effect.

05

Every service is reachable by a machine

Services are exposed as callable functions with defined inputs, outputs and authorisation — not only as pages a human clicks through.

Requires

a way to prove that a specific person authorised a specific agent to take a specific action, and to revoke it.

06

Accountability attaches to each decision

For any determination a machine contributed to, the government can say who is responsible, on what basis it was made, and how a citizen challenges it.

Where the UAE stands

A government that has already done this three times.

Each time, the commitment was announced before the capability existed — and each time it was met.

  • 2013–2015 — mGovernment. Two-year deadline set in May 2013. By May 2015: 96.3% transformation, 337 priority services across 41 federal entities.

  • 2018–2021 — Paperless Dubai. 43 entities, 1,800+ services. 100% paperless by December 2021 — 331 million sheets, AED 1.3bn and 14 million man-hours.

  • 2024–2025 — Zero Government Bureaucracy. 4,000+ procedures eliminated, 1,600 redundant requirements removed, delivery time down 70%+, via 690 teams across 30 entities.

  • 2026–2028 — Agentic government. “Within two years, 50% of government sectors, services and operations will run on Agentic AI” — announced 23 April 2026.

  • 2026 — The machinery, so far. The Ministerial Development Council became the Ministerial Council for Artificial Intelligence and Development on 28 April, chaired by Sheikh Mansour bin Zayed, with Mohammad Al Gergawi chairing the taskforce. The first four federal agents — procurement, tax audit, customer satisfaction, technical support — launched on 20 May. Ajman renewed a trade licence through an agent on 23 July, a UAE first. The National Committee for Agentic AI issued a unified design guide for government services on 21 August. Phase two — FedAI, a national technical ecosystem for federal entities to develop, adopt and operate agentic solutions, plus a customer-experience laboratory — was reviewed on 9 September. Dubai Future Foundation opened an accelerator for 33 Dubai entities and 300+ public-facing use cases on 5 October.

  • And Abu Dhabi has already claimed the title. Its Digital Strategy 2025–2027 states the goal as “the world's first fully AI-native government across all digital services by 2027”, and TAMM AutoGov — “the world's first AI public servant” — has run renewals and payments in the background since October 2025. The test above gives a way to measure progress towards it, service by service.

12m

UAE PASS users, July 2026

15,000+ services and 350+ entities (TDRA, 2025)

11th

globally on the UN EGDI 2024

1st in the region

3.8m

TAMM users in 2025

1,150+ services, 92.5% satisfaction

The question is no longer whether the UAE can set a deadline and meet it. It is which services the next one lands in.

What changes · 01

Conversation opens the service. A structured object completes it.

Conversational access will become normal. It will not replace transaction. The distinction is distributional and it decides where money is spent:

The head of the distribution — renew, pay, check status, upload — is high volume, well understood by the citizen, and already well served by forms. A dialogue turn here adds time and cost and nothing else.

The tail — ambiguous, cross-departmental, “which of these applies to me” — is where navigation has always failed and where conversation is transformative.

The business case is not conversation versus forms. It is conversation versus the phone call that a failed digital journey generates. Rank services by call volume relative to transaction volume and start at the top.

What changes · 02

What we mean by “the card”.

A definition, because the term is used loosely and the distinction matters.

An action card is a structured, interactive component the service returns into the conversation: typed fields that carry their own validation, pre-filled with what the state already holds, an explicit statement of what the citizen is agreeing to, and one button that commits it.

Five bands, in order

01

Identity and context

Who this is for and what it concerns — resolved, not retyped.

02

Pre-filled, from the register

What the state already holds, shown so it can be corrected.

03

Deltas only, validated

Just what has changed. Each field carries its own rule and its own error message.

04

The consequence, computed

Fee, entitlement, date, outcome — shown before commitment, not after.

05

The declaration, then one action

What the citizen attests to, and a single committing button.

Not a bespoke build — the formats already exist

Adaptive Cards

Microsoft’s declarative JSON card format, rendered natively by the host. Input types are Text, Number, Date, Time, Toggle and ChoiceSet; actions are Submit, OpenUrl, ShowCard and Execute.

MCP Apps (SEP-1865)

Published 21 November 2025, authored by OpenAI and Anthropic maintainers with the MCP-UI creators. Templates are pre-declared so a host can review them before they run.

Messaging equivalents

WhatsApp Flows, RCS interactive messages and Apple Messages for Business all render multi-field forms inside the conversation. Governments already reach citizens on these channels.

Schema version 1.3 of Adaptive Cards added required-field marking, validation patterns, error messages and labels together — the standard itself concluded a card without them was not fit for purpose. WCAG 3.3.4 says the same thing in law-adjacent terms: for legal or financial submissions, the action must be reversible, checked, or confirmed — and the last two require a reviewable object the citizen can see.

And there is a government-specific reason the object matters. The objection is that a transcript is a perfectly good record. It is — of what was said. But when a citizen later asks why a decision went the way it did, or simply wants to check what they agreed to, the useful record is what they were shown: which options were on the screen, which fields were validated, and what sat above the button they pressed. A transcript cannot reproduce that. A structured submission can, and the citizen can be shown it again at any point.

And one architectural rule follows: the card is rendered by the trusted system, not written by the model. The model chooses which card and which fields; the registry fills them. This keeps personal data out of the prompt, stops a compromised model fabricating a convincing but false declaration, and keeps the record of what the citizen saw under the state's control. Privacy, security and legal control in one decision.

Three cards, field by field

Real services in the UAE. The fields, the validation and the declaration are the point — figures are illustrative.

Illustration only, not a working form

Pre-fill is the whole value

Birth certificate

Ministry of Health and Prevention

Child
Amina Al Suwaidi
✓
Date of birth
14 March 2026
✓
Place of birth
Registered hospital, Dubai
✓
Copies required
1
Delivery
Home · 3 working days
▾

Fee AED 65 · delivery 3 working days

“I confirm I am the parent named on this record.”

Request certificate

Illustration only, not a working form

The state answers before you ask

Trade licence renewal

Dubai Department of Economy and Tourism

Establishment
Najma Trading LLC
✓
Licence
CN-1104872 · expires 30 Nov
✓
Details unchanged
Yes
●
Activities
3 — unchanged
✓
Ejari tenancy
Valid to 12 Mar 2027
✓

Renewal fee AED 1,020 · one condition outstanding: municipality fee unpaid

“I confirm the information shown is accurate and complete.”

Pay and renew

Illustration only, not a working form

Show the outcome before commitment

Social support

Ministry of Community Empowerment

Household
4 people · 2 dependants
✓
Income
From salary record
✓
Housing
Registered tenancy · 18 months remaining
✓
Circumstances changed
No
●
Bank account
•••• 4417
✓

Estimated award AED 3,200 per month from 1 October

“I confirm my circumstances are as shown and will report changes within 30 days.”

Submit application

✓ Pre-filled from the register — shown so it can be corrected

▾ ● The only fields the citizen touches

The consequence, computed and shown before commitment

What changes · 03

The state takes the first step, for a definable class of service.

The class is definable, and a government should define it explicitly before a programme starts rather than case by case during one. A service qualifies when all four are true:

A service qualifies when all four are true

01

The facts that determine entitlement are already held by the state, in an authoritative register.

02

Those facts are correct at the granularity and for the period the determination requires.

03

The determination is an award, not a liability or a reduction.

04

If it is wrong, the error is visible and recoverable.

An offer the citizen confirms is safe and lawful. An adverse determination applied automatically on inferred data is the failure mode behind the best-known failures in this field. That is a governance boundary, not a technical one — write it down before the programme starts.

Which services this actually means

The state takes the first step, for a definable class of service.

Trigger

Services

What it looks like in practice

Mode

Birth

Family allowance, child benefit, registration bundle

Abu Dhabi's 2022 bootcamps built a “starting a family” bundle across entities. The automatic award exists abroad: Austria pays on a registered birth since 2015; Estonia sends an offer to confirm.

A · B

Age

Pension award, concessionary travel, school enrolment

The state knows the date of birth exactly. An application adds no information.

A

Income

Pre-filled tax returns, social tariffs, means-tested support

Abroad, Portugal has awarded its social energy tariff automatically from a tax match since 2016.

A · B

Expiry

Licence, permit and document renewal

ICP's Taqdeer package (December 2025) renews, prints and delivers passports and Emirates IDs for enrolled senior citizens and people of determination before expiry — steps cut “from three to zero”.

B · A

Death

Tell-us-once, survivor benefits, estate notification

Abu Dhabi's bereavement bundle (2022): one notification propagated, replacing the same conversation repeated at the worst time.

C · B

Business

Licence renewal, pre-filled returns, inspections

Ajman (July 2026): a pre-expiry notification opens the renewal, the agent checks lease validity with the Municipality and coordinates the departments, the customer confirms.

B

A = awarded automatically · B = pre-filled offer the citizen confirms · C = notified once and propagated.

A test worth running on your own service catalogue. Name a benefit whose eligibility turns only on age. The state knows every citizen's date of birth exactly. If that benefit still requires an application, the barrier is not data — and it is falling hardest on the people least able to clear it.

A note on the word “predictive.” Proactive services are often described as predicting when a citizen will need something. In practice, every working example triggers on a fact that has been registered — a birth recorded, an income filed — not on a forecast of future need. Detect events; do not forecast needs.

What changes · 04

The assistant arrives, and the question is whose.

On current trajectory, a large and growing share of citizens in digitally mature states will reach government through a conversational agent within this period. The strategic question is not whether that happens. It is whose agent it is, and most governments are answering it by default rather than by decision. There are four options, and they are not exclusive:

The assistant arrives, and the question is whose.

Option

What it means

When it applies

Build

A first-party government assistant

Where the state has the capability to run it; necessary anyway for authenticated transactions

Publish

A machine-reachable service surface any authorised agent can call

Almost always — and it is what makes the first-party assistant work too

Partner

Deliver inside a private platform with population reach

Where reach cannot be achieved otherwise, accepting dependency

Regulate

Conditions on assistants that mediate public services — accuracy disclosure, liability, no unauthorised transactions

Necessary regardless of the other three; we have not found a government that has

The missing infrastructure for all four is delegation: proving that a named person authorised a named agent to take a named action, with a scope, an expiry and a revocation. Versions for human intermediaries work at scale in several countries. We have not found a national digital identity that lets a citizen authorise a software agent to act, with a scope, an expiry and a revocation.

It is worth building for a reason beyond agents. A carer acting for an elderly parent, an accountant acting for a business, an advice worker acting for a claimant and an AI agent acting for a citizen all need the same thing: scope-limited, time-bounded, revocable, auditable authority to act on someone else's behalf. The assisted-access problem and the agentic-government problem are one problem. Solving it serves the most digitally excluded citizens and the most technically advanced use case with the same infrastructure — which is also the version that survives a political conversation.

What changes · 05

Language is an engineering constraint, not a translation task.

Every published accuracy figure for government AI is English-language. The drop is large, and it is specific to the model.

Language is an engineering constraint, not a translation task.

Measured

English

62.8%

Modern Standard Arabic

51.9%

Emirati

49.8%

Egyptian

48.9%

Saudi

48.2%

Moroccan

45.0%

19 models, identical translated questions. Gulf dialects perform best — Emirati is the strongest of the five tested.

Tool-calling degrades in Arabic — even when the tools are described in English. Accuracy drops 5–10% when the citizen writes in Arabic, regardless of whether the tool descriptions are Arabic or English. Look up a record, check eligibility, file a form: that is what an agentic service does. This cannot be engineered away by keeping the backend in English.

And one question to settle early. Article 7 of the Constitution makes Arabic the official language. The UAE has an Arabic language law in preparation. Which language version of an automated determination governs is a question worth settling in policy before the first determination, not after the first appeal.

  • Citizens do not write in MSA. They write in dialect, code-switch and transliterate. A system evaluated on Modern Standard Arabic meets a distribution it was never tested against.

  • Arabic-specific is not automatically better. Published Arabic benchmarks consistently show large general models matching or beating smaller Arabic-specific ones. Choose on measured performance, not on origin.

  • Right-to-left is a platform matter. W3C's Arabic Script Gap Analysis (September 2026) records that browsers still cannot fully justify Arabic text. If the card is the record, RTL correctness is part of it.

What changes · 06

The rules that will shape this, and where they come from.

Two regimes matter to a UAE entity: the UAE's own, and the EU's — which reaches further than most assume.

In the UAE

  • UAE. Federal entities are governed through standards, procurement and the AI Ethics Principles; Federal Decree-Law 45 of 2021 (PDPL) applies outside government. That puts the weight on how a programme is specified and bought.

  • June 2026. The Federal Authority for Artificial Intelligence and Data is established, merging the AI Office, TDRA's digital government sector and the Emirates Data Office. It is the natural owner of what comes next.

  • DIFC. Regulation 10 (September 2023) is a binding regime for autonomous and semi-autonomous systems — certification or a named Autonomous Systems Officer for high-risk systems. A working model already inside the UAE.

  • The agentic programme's own standard. The National Committee for Agentic AI's Practical Guide for Government Entities on Designing Government Services with Agentic AI (August 2026, Arabic) sets the design rules for federal services: the customer states the need in their own words and the system understands it before any procedure starts; information already provided is never requested again; the customer is kept informed at every step; privacy and security hold; the customer wants a result, not a sequence of procedures; and exceptions hand over to a competent employee with the full details of the request. Elements 1, 2, 5 and 6 of this page are, in substance, now federal design policy. Elements 3 and 4 — rules as code, and the boundary on adverse determinations — are where each entity's own programme builds on it.

  • The readiness assessment. Federal services powered by assistive AI are assessed against a ten-item readiness template before launch (uaemodel.egsep.ae). The table below maps the six elements to it.

And the European timetable

The EU AI Act binds any provider or deployer whose system output is used in the Union. For a UAE entity serving EU residents, or procuring from a supplier that does, these dates already apply — and they are becoming the reference standard governments are measured against.

The rules that will shape this, and where they come from.

Date

What

Detail

2 Aug 2026

Transparency

A person must be told they are interacting with an AI system, from the start of the first interaction. Article 50 — already in force, and not deferred.

2 Dec 2027

High-risk duties

Full obligations where AI is used to determine eligibility for essential public services: impact assessment, human oversight, logging, registration. Annex III, point 5.

2 Aug 2030

No grandfathering for public bodies

Legacy AI already in use by a public authority must comply. Private-sector legacy systems are exempt absent significant change. Public ones are not. Article 111(2).

These are the dates in the EU AI Act — Regulation (EU) 2024/1689 — as amended by the Digital Omnibus, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026. The high-risk duties were deferred by roughly sixteen months. They were not weakened.

How the six elements map to the federal readiness assessment

The federal programme assesses each agentic service against ten readiness items before launch. Read against the six elements:

The rules that will shape this, and where they come from.

Readiness item

Element in this page

Outcome-first intent · natural language

1 — Intent replaces navigation

Uses the customer's context and data · gathers government data and coordinates across entities

2 — The state never asks twice

The path adapts to eligibility

3 — Rules execute, not only publish

Follow-through on steps that are not immediate

4 — The state takes the first step

Actually executes rather than linking · consent and revocation

5 — Every service is reachable by a machine

Exceptions and handover to a person, with context

6 — Accountability attaches to each decision

Pause and resume

The card: a structured submission can be resumed; a transcript cannot

Three items — actually executing, handing over with context, and following through on non-immediate steps — are where the difference between an AI-enabled and an AI-native service shows. They are the committee's version of the test at the top of this page.

How to start

Two tracks, always running. Neither waits for the other.

No government can stop shipping for three years while it fixes its registers. It does not have to, and the evidence says it should not try.

Two tracks, always running. Neither waits for the other.

Track one · the service track

Track two · the foundation track

Visibility

Visible

Invisible

Tempo

Weeks to months

Quarters to years

Content

Services shipped into the foundation you already have. Requirements removed. Journeys consolidated. Conversational entry added in front of what works.

Registers corrected. Rules turned into code. Services exposed as callable functions. Delegation and assurance built.

What it buys

The credibility and the budget for the other track

What the first track will be able to ship next year

The foundation does not decide whether you ship. It decides which services you can ship into — and each quarter of foundation work opens the next tier.

So the sequencing question becomes a portfolio question

Ship now — services whose facts already sit in a register you trust. Renewals, confirmations, status, pre-filled returns.

Ship next — services needing one register corrected or one rule written as code. Scope the foundation work from the service you want, not the other way round.

Ship later — services needing cross-entity integration or a legal change. Start that work now; do not hold the first two tiers behind it.

How the UAE has actually done it

McKinsey's assessment of the UAE model: the country “pursued concurrent initiatives rather than sequential phases”, with foundational work running in parallel with service delivery.

  • Concurrent, not sequential. This is not a theory — it is the observable operating model.

  • Abu Dhabi runs three declared tracks at once. Services (TAMM to 3.8m users), platform (sovereign cloud migration, unified ERP, a unified government data centre) and human capital (95% of staff AI-trained) — the plumbing built underneath a running, growing service estate.

  • Dubai solved it organisationally. Law 22 of 2023 makes Digital Dubai Government Establishment (delivery) and Dubai Data & Statistics Establishment (data) co-equal siblings under one authority. Neither gates the other.

  • Zero Bureaucracy inverted the order deliberately. Phase 1 attacked visible procedural burden — deleting a procedure needs no register. Phase 2 attacked the digital plumbing. Speed came from subtraction and from 690 teams working in parallel.

  • And the method is documented. Abu Dhabi's experience programme runs two concurrent streams on one 100-day clock: “minilabs” for single services, “bootcamps” for life-event bundles that need cross-entity integration.

  • And the 2026 programme is running the same way. Phase one of the federal Agentic AI Project — building capability, surveying services and operations, setting priorities — ran while the first four federal agents, Ajman's renewal and Abu Dhabi's AutoGov were already live. Phase two adds the shared foundation, FedAI, underneath services that are already shipping.

Priorities

Six priorities. Not six phases.

Not six phases. Each produces something you can hold at the end of it.

Six priorities. Not six phases.

Priority

What it produces

01

Remove the requirement before automating it

A list of requirements for one service line, each with an owner and a date by which it is either justified or gone.

02

Make the registers right for the services you are shipping into

A one-page readiness note per candidate service: what the register covers, how current it is, and whether it records the fact at the resolution the decision needs.

03

Turn one service's rules into code

The eligibility logic of one high-volume service, running in parallel with the existing process and tested against cases already decided — including appeals that were upheld.

04

Retire the legacy that blocks the next tier

A list of the systems standing between you and the services you want next, with funded dates.

05

Stand up assurance before the first incident

A register of automated decisions — not of systems — with a named owner for each.

06

Publish performance, including who it does not reach

One page per service per quarter: completion, first-time-right, abandonment and exclusion by group, appeals and overturns.

Nothing here says “deploy a model.” That is deliberate. The model is the least scarce input in this programme and the easiest to change. Governments that start by procuring a model and work backwards to the service consistently end up with an assistant in front of an unchanged process.

Impact and benefits

Three of the four real benefits are not cost reduction.

What the evidence supports, and how much of it applies here.

Take-up

Proactive delivery closes the gap between who is entitled and who claims. This is the largest single citizen benefit available, and it increases expenditure.

Consistency

Rules as code remove variation between offices and officials, and make the effect of a policy change immediate and testable rather than a year of retraining.

Compressed variance

The best causal evidence available — a field study of 5,172 customer-support agents given a generative AI assistant (Brynjolfsson, Li and Raymond, Generative AI at Work, Quarterly Journal of Economics, 2025), a private-sector setting that must be treated as an analogue — shows AI raising output by about a third for the newest and least skilled workers, with little effect for the most experienced. In benefits assessment, casework and first-line advice, where governments face high turnover and long ramp times, the value is bringing new staff to competence faster, not making the best caseworkers better.

Speed

Application-to-decision latency disappears where the state already holds the facts. Processing measured in hours has fallen to seconds in production proactive services.

Applicability in the UAE: take-up is the weakest fit — high-coverage entitlement programmes leave less unclaimed than European welfare systems. Consistency, compressed variance and speed all apply directly, and speed is where the published UAE evidence is strongest: Zero Bureaucracy cut delivery time by over 70% before AI was the main instrument.

Why time saved is not money saved

This is the gap that undermines most government AI business cases, and there is a plausible mechanism for it. Careful labour-market research — Humlum and Vestergaard's study of some 25,000 Danish workers in eleven exposed occupations (Large Language Models, Small Labor Market Effects, 2025) — finds self-reported time savings among users alongside no measurable effect on earnings or hours at the population level, while around one worker in twelve reported entirely new tasks created by the tools. The hypothesis, and it is a hypothesis rather than an established causal finding, is that saved time is absorbed by verification, review and governance before it ever reaches a budget line. It is testable. A government running an AI programme should measure the new work as deliberately as the saved work.

The business case to make, and the one to avoid

For conversational services, build the case on first-time-right completion, error reduction and diverted contact-centre volume — not on deflection, and not on headcount.

Proactive delivery increases outlay and reduces administrative cost. The outlay increase is the point. Any business case showing a large net saving is almost certainly counting money recovered from citizens.

For proactive services, state the arithmetic at the outset:

That is the earliest available warning of a programme heading for the failure mode described above — visible in the finance annex, before a line of code is written. The honest case is: this costs more, it reaches the people the law intended to reach, and it reduces the cost of serving the people it already reached.

The measurement problem

We could not locate a current government cost-per-transaction-by-channel dataset published anywhere since 2012, any published containment or resolution rate for a named government conversational service, or any controlled comparison of conversational service initiation against conventional forms.

A government that commissions its own baseline in the first quarter of an AI programme will be the only party in the room able to say whether it worked. That is the cheapest high-value item on this entire list.

One metric to refuse to report: containment or deflection. It measures a citizen failing to reach a human, no audit institution publishes it, and it will be optimised at the expense of every other number.

The position

Start where the foundation already reaches.

Then extend it, one tier at a time, while the service track keeps shipping. Bring one high-friction journey and the people accountable for it; leave with the requirements to remove, the register-readiness note, the register of automated decisions and the baseline nobody currently holds.

Discuss a service to redesign

This quarter

  • Service track — Pick services whose facts already sit in a register you trust — renewals, confirmations, status, pre-filled returns. Remove the requirements you cannot justify. Ship.

  • Foundation track — Scope the foundation work backwards from the services you want next year, not from an abstract maturity model.

  • Governance — Write down one boundary before anything else: an offer the citizen confirms is safe; an adverse determination applied automatically on inferred data is not.

Candidate journeys in the UAE: trade-licence renewal; a birth-triggered bundle; a social-support determination, where the control boundary matters most.

Research context

Sources, checked 7 October 2026

This page is the web edition of the Cognirise point of view The AI-Native Government (United Arab Emirates edition, version 5.1, 7 October 2026). Figures are attributed to the body that published them so a reader can weigh them. Where the public record does not support a claim commonly made in this market, the page says so rather than filling the gap.

  1. UAE Cabinet, framework to deploy agentic AI across 50% of government sectors and operations within two years (23 April 2026)The commitment, its two-year horizon, and the oversight structure.Official statement of intent; scope covers sectors, services and operations.

  2. UAE Government portal, Agentic AI for Government Services and the practical design guide (updated 30 September 2026)The National Committee for Agentic AI, the 2028 target and the design guide (Arabic).Principles above are as reported in English-language press; the Arabic original governs.

  3. The National, Ajman uses agentic AI to renew trade licence in UAE government first (23 July 2026)The AjmanOne renewal: pre-expiry notification, journey mapping, lease-validity check, cross-department coordination.Press report of a government announcement.

  4. Sharjah24, Ministerial Council for AI reviews second phase of the Agentic AI Project (9 September 2026)FedAI, the customer-experience laboratory and the content of phase one.Programme announcement.

  5. The National, Dubai launches agentic AI accelerator for public-facing government services (5 October 2026)33 entities, 300+ use cases, the partners and the Dubai AI Seal.First phase.

  6. UAE Government Media Office, Mohammed bin Rashid launches phase 2 of Zero Bureaucracy programme (16 June 2025)Procedures and requirements eliminated, delivery-time reduction, 690 teams across 30 entities.Figures as published by the programme.

  7. Department of Government Enablement, Abu Dhabi Government Digital Strategy 2025–2027AED 13bn programme and the AI-native-by-2027 goal.Projections as published.

  8. Department of Government Enablement, TAMM 4.0 and AutoGov at GITEX Global 2025 (14 October 2025)AutoGov’s scope and TAMM’s service and partner counts.Operator-reported.

  9. Gulf News, UAE launches automatic passport and Emirates ID renewals (December 2025)The Taqdeer package and its “three steps to zero” claim.Press report of an ICP announcement.

  10. Digital Dubai, World Economic Forum recognises UAE PASS (2026)12 million registered users as at July 2026.Registration count as published.

  11. u.ae, UAE mGovernment initiativeThe 2013–2015 deadline and the 96.3% result.Self-reported by the delivering government.

  12. Morgan Lewis, UAE establishes Federal Authority for Artificial Intelligence and Data (June 2026)The authority and what it merges.Law-firm summary of the decree.

  13. DIFC, Regulation 10 — personal data processed through autonomous and semi-autonomous systemsThe DIFC regime for high-risk autonomous systems.Applies within the DIFC only.

  14. arXiv 2510.27543, evaluation of 19 models across English, Modern Standard Arabic and four dialects (2025)The English–Arabic accuracy gap on identical translated questions, including Emirati at 49.8%.Academic benchmark; not a government-service evaluation.

  15. arXiv 2601.05101, Arabic prompts with English tools (2025)Tool-calling accuracy drops 5–10% for Arabic prompts regardless of tool-description language.Benchmark adaptation of the Berkeley Function Calling Leaderboard.

  16. Khaleej Times, UAE to introduce Arabic language law by 2027 (29 April 2026)The law’s status and horizon.The law is in drafting.

  17. Government Digital Service, GOV.UK Chat launches (14 May 2026)The 18-month pilot, the question volume and the deliberately narrow launch scope.UK example, used for the evidence pattern only.

  18. OECD Observatory of Public Sector Innovation, Estonia’s proactive family benefitsProcessing time, satisfaction and the share of recipients who had previously applied.Estonian example, used for the evidence pattern only.

  19. The Markup, Mamdani to kill the NYC AI chatbot caught telling businesses to break the law (30 January 2026)The documented reversal: the March 2024 findings, the shutdown and the reported costs.US municipal example; costs as reported in the press.

The source manuscript labels its research as checked 7 October 2026. That date is inherited from the supplied manuscript, not a new verification performed during website integration.

Source check date: