UAE sets a two-year target to run half of federal services on agentic models

The consequence for UAE and GCC organisations is blunt. When a licence renewal or a tenancy change can complete without a human, a chatbot that answers questions is no longer a digital service. It is a prototype. Budget and attention move from the conversation layer to the transaction layer, which means integration, identity and audit work. That is the work most teams have been deferring, and it is now measurable against a published public-sector standard. The UAE Government's own portal writes the mandate down as Agentic AI in 50 per cent of government sectors, services and operations by 2028, and points to the National Committee's unified guide, which is in Arabic only (https://u.ae/en/about-the-uae/digital-uae/digital-technology/Agentic-AI-for-Government-Services). The Cabinet announcement of 23 April 2026 is the source document, and it is unusually direct about accountability: the performance of ministers, directors general and entities is to be assessed on how fast they adopt the transformation (https://uaecabinet.ae/en/news/under-directives-of-uae-president-and-in-world-first-mohammed-bin-rashid-reveals-new-uae-government-framework-to-deploy-agentic-ai-across-50-of-government-sectors-operations-within-two-years).

Look at where the difficulty actually sits. TAMM AutoGov is not one service: it renews licences, handles utility payments and books routine healthcare appointments for 3.8 million users, and the platform resolves 95% of requests through AI, with more than 1.9 million service cases closed through conversational AI, 36 million customer visits avoided a year and 92.5% user satisfaction (https://www.dge.gov.ae/en/news/2025-dge-highlights). AutoGov shipped with TAMM 4.0 at GITEX Global 2025 (https://www.mediaoffice.abudhabi/en/technology/department-of-government-enablement-launches-tamm-4-during-gitex-global-2025/). Ajman's trade licence renewal is a chain across two bodies, because the commercial tenancy has to be renewed before the licence can be processed. In enterprise terms that is CRM, ERP, e-signature and a payment gateway inside one autonomous run, where every hop needs a defensible record. The reward is transactions completed with no human touch, which is a different metric from conversations started.

The governance scaffolding tells you what that record must contain. The National Committee's guide sets out the eight stages of an assisted journey and is explicit that no binding or sensitive action starts without clear approval from the customer, that the assistant applies an ask once policy instead of re-requesting data the government already holds, that it shows a plain plan and a log of the actions carried out in the customer's name, and that on a complex or sensitive case it hands the request, the context, the actions and the approvals to an employee so the customer does not retell the story. Its readiness checklist says every requirement needs evidence that can be reviewed and audited, not a descriptive confirmation, and government payments run through Sadaad UAE after explicit approval (https://uaemodel.egsep.ae/agentic_ai_guide_website.html). The Agentic AI Application Lab, opened on 22 October, exists so ministries can test with real users before rollout. Al Olama's control tower agent, audited continuously by agents and humans, is the same idea at portfolio level. The test for any agent is not whether it answers well. It is whether you can reconstruct, after the fact, what it did, on whose authority, and what it changed.

Technically we would treat this as a control problem. Put one OpenAI-compatible gateway in front of every model call, give each agent its own non-human identity registered at deploy time, and attach a hard budget and explicit tool permissions to that identity. Workload identity frameworks such as SPIFFE exist for exactly this: a cryptographic identity per workload, with short-lived credentials rotated automatically, so an agent is a principal you can authorise, revoke and attribute actions to rather than a shared API key (https://spiffe.io/docs/latest/spiffe-about/spiffe-concepts/). Write the action log at the gateway and at the tool layer, not inside the application, so it survives the app being rewritten, and use an open schema for model and tool call telemetry such as OpenTelemetry's GenAI semantic conventions rather than a bespoke JSON blob (https://opentelemetry.io/docs/specs/semconv/gen-ai/). Put a human approval gate on anything irreversible, and keep the run, the approval and the rollback in one workflow engine instead of scattered scripts. In our own stack that pattern is FastLLM Proxy for routing, budgets and access control in one place, with Dhole holding the approval points in the workflow and Kuvryn AI running each team's agents and models on your own GPUs in their own space. An illustrative policy for a licence-renewal agent:

agents:
  trade-licence-renewal:
    identity: "svc-agent/trade-licence-renewal@org.ae"
    models: ["gpt-4.1-mini", "command-a"]
    budget: {daily_aed: 250, hard_stop: true}
    consent: {mode: before_action, store: immutable/consent}
    tools:
      tenancy.renew: {approval: "human:licensing-supervisor"}
      licence.renew: {approval: "human:licensing-supervisor", log: immutable/audit}
    limits: {max_autonomous_steps: 6, on_anomaly: halt_and_page}

Where teams will get it wrong is predictable. They buy an agent framework before fixing identity and data access. They treat the audit log as a debugging artefact rather than evidence that must survive an investigation. They automate the happy path and leave no reconciliation step for the case where two agents touch the same record. They report conversations started instead of transactions completed without a person. The GCC advantage is real: shared rails such as UAE PASS and TAMM-style service catalogues make cross-organisation agents feasible here earlier than in most markets. The advantage goes to whoever can prove what their agents did.