AI in government entities: what capability programs must build

Why is understanding AI no longer enough in government?

The ambition is not new. In 2018, PwC Middle East estimated that AI could add USD 320 billion to the Middle East economy by 2030, about 11% of GDP (PwC Middle East, 2018; forecast). What has changed is what is being asked for. The goal is no longer that officers know what AI is, but that a real service runs with it, safely, in front of beneficiaries.

Government services sit close to the centre of that change. The International Labour Organization finds that one in four workers worldwide is in an occupation with some exposure to generative AI, that most jobs are expected to change rather than be replaced, and that clerical work is the most exposed (ILO Working Paper 140, 2025). Enquiries, applications, licences and correspondence are the daily work of many government departments.

Access to tools does not produce the value on its own. Three in four AI high performers have fundamentally redesigned their workflows, against one in four of the rest (McKinsey, The state of AI in 2026; Indicative). Redesigning a government workflow needs someone to decide, someone to apply and someone to build. That is what capability programs must produce.

Which three audiences should public sector AI training serve?

Each audience makes a different decision and leaves with a different output. Mixing them in one room produces general awareness and no change to the service.

The sequence starts with leaders for a practical reason. Where employees see a clear AI strategy, 80% report measurable impact from AI, against 55–60% where it is limited (BCG, AI at Work 2026; Indicative). In a government entity, support does not mean enthusiasm. It means written decisions: what is allowed, who approves, and how impact is measured.

What does one request look like, end to end?

The request passes through five steps:

1. Intake. The enquiry arrives through the department's channel and the agent classifies it. Anything outside the permitted types goes straight to an officer.
2. Retrieval. The agent reads from the version of the regulation the department owns and keeps current, not from public sources.
3. Draft. It writes a reply in Arabic and cites the article used, so the approver sees the source, not only the answer.
4. Approval. A licensing officer approves, edits or rejects the reply. Nothing reaches the beneficiary before this step.
5. Record. The enquiry, reply, cited article and approver's name are logged, so any reply can be audited later.

On the Delegation Ladder, this agent sits at Draft, with binding consequences because the reply reaches a beneficiary. It needs a named approving officer, and it moves up a level only on evidence from the department's own records.

Where does the data stay, and which models are used?

Officers should not be trained on a tool they will not be allowed to use at work. The principle Cader Group works to: "Sovereign by design: your models, your data, in-country."

In practice, that means three things:

• Models the entity has approved. Training runs on the entity's approved models, whether hosted models such as Claude, Gemini, OpenAI GPT and Codex, or open-weight models such as Qwen and Falcon that can run in-country.
• Data that mirrors records without exposing them. Exercises use sample data that matches the entity's records in structure and language, without putting real beneficiary data into unapproved tools.
• Residency decided at design. Where data is processed and stored is set when the agent is designed, not after it is built.

How do you sequence programs without a new procurement round for each?

The sequence we propose for government entities:

1. The Readiness Snapshot measures the starting point; its results become the baseline against which impact is measured.
2. Leaders decide what is allowed and who approves.
3. Teams test workflows on their actual tasks.
4. Builders build agents on the services leaders have approved.
5. A live pilot is measured against the baseline and ends in a scale-or-stop decision.

Government programs usually fail between stages, not within them. A leaders' session ends with a good decision, the next stage waits for a full procurement cycle, and the decision loses momentum. We therefore propose contracting the sequence once, in stages with defined outputs, so each starts without a new procurement round.

Measure where your own team stands: take the free Readiness Snapshot →