On 23 April 2026 the UAE set one of the most specific government AI targets anywhere: half of federal government sectors, services and operations running on agentic AI within two years. The announcement came with accountability. Ministers, directors-general and federal entities will be assessed on how quickly they adopt the new tools, a task force under the Minister of Cabinet Affairs will monitor execution, and every federal employee will be trained.
That leaves each entity with a clock and a question: which half, and in what order? This is the sequence we use with public-sector clients.
What “agentic” means for a service
A chatbot answers. An agent acts. In the Cabinet’s own description, these systems “monitor changes, provide analysis, issue recommendations and carry out sequences of actions”. For a licence renewal, that can mean:
- Reading the application and the attached documents.
- Checking them against the rules and the records the entity already holds.
- Asking the applicant for anything missing, on the channel they used.
- Preparing the decision with the evidence attached.
- Routing it to the officer who signs, or issuing it automatically where the rules allow.
The value is in steps 1 to 4. The governance question is step 5.
Months 0–3: inventory and ranking
Start with a list of every service and internal operation, with four numbers for each:
- Volume: transactions per month.
- Rule clarity: how much of the decision is written down.
- Risk: the harm if the agent gets it wrong, and whether the decision has legal effect on a person.
- Data readiness: whether the records the agent needs are digital, reachable and clean.
High volume, clear rules, low risk and ready data move first. A typical first wave is information requests, appointment and status queries, document completeness checks, internal HR and procurement requests, and correspondence drafting.
Two Cabinet decisions from the same meeting help here: the data sharing policy (collect data once, use it securely) and the digital records policy. Both reduce the integration work an agent needs.
Months 3–9: the first wave in production
Build the first agents on one shared platform rather than one tool per department. The platform needs:
- Identity and authority: each agent has a defined scope, like a job description, and cannot act outside it.
- Audit trail: every step the agent took, the sources it used and the rule it applied, stored where an auditor can read it.
- Human handover: a clear point where the case moves to a named officer, with the context attached.
- Arabic and English on every channel citizens already use: web, app, WhatsApp, voice and the counter.
- In-country hosting: models and logs inside the UAE, with evidence of where each inference ran.
Measure from day one. The federal assessment will ask what changed, so record resolution rate, time to decision, cost per transaction and the share of cases handed to a person.
Months 9–18: scale and redesign
Once the first wave is stable, the larger gains come from redesigning the service, not automating the old form. If the agent can pull a record from another entity under the data sharing policy, the citizen no longer needs to upload it. This is where “zero bureaucracy” and agentic AI meet.
This phase also brings in internal operations: finance reconciliation, procurement evaluation support, workforce scheduling and reporting. These rarely face citizens but count toward the 50% of “operations”.
Months 18–24: the hard cases and the evidence
The last services are the discretionary ones. Many will stay human-led, with agents preparing files and checking consistency. That is still agentic adoption, and it keeps accountability where the law puts it.
By month 24 an entity should be able to show, per service: what the agent does, what a person still decides, the volumes handled, the error rate and how errors were found.
Common failure points
- One pilot per department. Twenty pilots on twenty tools do not add up to 50%. Use one platform with shared guardrails.
- No owner. Every agent needs a named business owner, not just an IT owner.
- Measuring usage instead of outcomes. Logins do not matter. Time to decision and cost per case do.
- Ignoring the channel. Many residents prefer WhatsApp or the phone. An agent that lives only in the portal reaches the people who needed it least.
How GridStudio helps
GridStudio is headquartered in Dubai and builds AI services for government and regulated enterprises. We run the service inventory, build agents on a shared orchestration layer with audit and handover built in, host in-country, and operate the system under an SLA if the entity prefers. See our work for government and the public sector and our UAE page.
Questions
On 23 April 2026, Sheikh Mohammed bin Rashid Al Maktoum announced that the UAE will move 50% of government sectors, services and operations to autonomous, agentic AI within two years. Ministers and federal entities will be assessed on how fast they adopt it, and all federal employees will receive AI training.
An agentic system does more than answer questions. It monitors a process, analyses it, recommends an action and can carry out a sequence of steps, such as checking an application, requesting a missing document and routing the case, within limits set by the entity.
It should not, for decisions with legal effect on a person. A sound design gives the agent authority over preparation and routine steps, and keeps the signature on decisions with a named officer, with an audit trail of what the agent did.
With a service inventory ranked by volume, rule clarity and risk. High-volume, rule-based services with a clear evidence trail usually move first; discretionary decisions move last.