Advancing Government with Agentic AI

The UAE has set a deadline: half of government sectors, services and operations on agentic AI within two years. Saudi Arabia built the layer underneath first, and now leads the ten-country public sector AI adoption index. 

Both routes meet the same constraint, and it is rarely the model. It is authority: who is allowed to decide what, once an agent can act. 

On 23 April 2026 the UAE announced that 50 percent of government sectors, services and operations will move to agentic AI for autonomous execution and decision making within two years. 

It is not a target to remove half of the federal workforce. It is a target for how work is performed. On 18 May the Cabinet approved a capability program for 80,000 federal employees across five occupational categories, and the framework states the principle it rests on: people come first. 

Five months from target to second phase: 

• 23 April  The 50 percent target is announced. 
• 4 May  A parallel initiative moves Dubai’s private sector onto the same two-year path. 
• 18 May  The Cabinet approves roles for every ministry, the first services for conversion and the 80,000 employee program. 
• 20 May  The first four federal agents are launched, for procurement, tax auditing, customer happiness and technical support. 
• 14 June  The AI and Data Authority is established. 
• 9 September  The second phase begins, with the FedAI ecosystem, a customer experience lab and internal challenges. 

From the bottom up, and since before any conversion target: 

• 2019  SDAIA is established as the national data and AI authority. 
• Sep 2024  The AI Adoption Framework is launched, with AI offices in 23 government entities. 
• Jul 2025  A National AI Index measures entity readiness across three pillars and seven dimensions. 
• Mar 2026  2026 is designated the Year of Artificial Intelligence. 
• Jul 2026  A national AI risk management framework is issued. 
• SAMAI  More than one million citizens certified in AI, and a second phase with eleven ministries. 

The result: Saudi Arabia ranks first of ten countries in the Public Sector AI Adoption Index 2026, and roughly two thirds of its public servants use AI at work every day, the highest rate in the index.

Draw the line with two questions. Does the decision create, remove or refuse a right or an entitlement belonging to a person? And is it contestable? If a citizen can appeal it, a person has to have decided it. 

The proposed rule: automate the approval where the criteria are met, and route every refusal to a person. What stays with a person:

1. Refusals and revocations 

2. Anything decided on hardship or exception grounds 

3. Anything that sets a precedent 

4. Enforcement, penalties and sanctions 

5. Decisions about the entity’s own employees

It is a pattern from earlier AI and automation programs in the region. A well-run pilot beats the manual baseline and everyone agrees it works. Then, in month five, legal is asked whether an action taken by a system carries the same authority as one taken by a named employee, and the honest answer is no. 

So the agent recommends and a person retypes its output into the official record. Cycle time does not improve, and the entity now pays for both. The model is not your constraint. Your delegation of authority manual is. 

1. Plan in activities, not positions. 

2. Redesign job architecture in parallel with the service, not behind it. 

3. Create the four roles most job catalogues do not yet define, grade them and give them a career path. 

4. Confront the inversion of the competency model. 

5. Decide the redeployment destination before go-live, and publish it. 

6. Move the national entry point up. 

7. Rewrite performance measures before go-live. 

8. Put HR on the body that approves autonomy levels, and convert HR’s own processes first. 

The HR strategy question is not how many people we will need. It is what those people will be doing in the week after the agent goes live, and whether anyone in the building owns that answer.

Technology is only part of the answer. In the 7th Annual Horváth CxO Priorities Study, based on interviews with more than 1,000 CxOs across 32 countries and 16 industries, the data foundation ranks as the biggest barrier to a successful AI transition, followed by workforce capability and skills. An agent cannot run a reliable process on fragmented data, and a pilot cannot become an operating model while process ownership and decision rights stay where they were. 

Capability is the other half. In Saudi Arabia, 84 percent of public servants have received employer AI training, but half of those say it focused on compliance rather than practical skills. Adoption is no longer the binding problem in this region. Depth is.

A percentage target creates counting behavior. If the denominator is the number of services, entities convert the services that are easiest to convert, and the few that carry most of the workload stay untouched. Ninety percent of services and ten percent of the volume is a report, not a transformation. 

The deadline is right. What it needs beside it is a rule that no service enters the program without a recorded before state. If a minister only had one question: show me the before number. 

Seven things, in sequence. None of them requires a platform decision first. 

1. Write the decision inventory for one high-volume service, before the strategy and before the platform. 

2. Record the before state for every service entering the program: time, cost, quality, employee effort and service outcome. 

3. Set decision rights in the design, not at sign off. 

4. Change the workforce plan from positions to activities, and start the job architecture work in parallel with conversion. 

5. Make capability role-specific, and test supervision rather than attendance. 

6. Treat data ownership as a governance decision. 

7. Redesign the target operating model around the new allocation of work.

Two years is long enough to convert services, and short enough to convert only the easy ones. The governments that have something to show at the end of it will be the ones that settled three questions early: which decisions an agent may take, what the people who did that work do next, and what counts as progress. 

Horváth Middle East works with government entities on exactly these decisions. 

By Saad Hamam, Partner and Managing Director, Horváth Middle East.