The United Arab Emirates' federal government has started building a formal classification framework for agentic AI — a system that sorts government decisions into tiers based on how much autonomy an AI agent is allowed to exercise before a human has to sign off.
That's a notable shift from how agentic AI has been rolled out in the public sector so far. Most early deployments — chatbots triaging citizen requests, agents drafting permit approvals, systems flagging tax discrepancies — were sold as productivity tools: faster queues, lower headcount costs, fewer backlogs. The classification effort described here is different. It's not asking whether an agent can do a task competently; it's asking who is allowed to say yes on behalf of the state, and under what conditions that authority can be handed to software.
According to AI News, regulators are treating this as a governance problem rather than a technical one — the open question isn't model capability, it's where the line of accountability sits when an autonomous system, not a person, initiates or completes a government action.
Why delegation is the harder question than capability
In private-sector agentic deployments, the failure mode of a bad output is usually a redo, a support ticket, or at worst a refund. In government, decisions carry legal weight and are frequently hard to reverse — benefits eligibility, licensing, enforcement, immigration status. That asymmetry is why delegation boundaries, not accuracy metrics, are reportedly the starting point for the UAE effort. A model that's 99% accurate is still wrong for one in a hundred citizens, and in a state context there's no small-claims equivalent to quietly absorb that error.
What a workable framework has to specify
Based on how the UAE and other regulators are approaching this, a workable delegation framework needs to answer at least four questions:
- Which decision types are reversible, like a re-issued permit, versus irreversible, like a benefits denial or an enforcement flag
- Who carries legal accountability when an agent acts on its own — the agency, the software vendor, or the civil servant who approved its use
- What triggers a mandatory human checkpoint before an action executes, not after it's already been taken
- How the system preserves a record of the agent's reasoning and tool calls, not just its final output, so a decision can be reconstructed later
The exact tiers the UAE settles on weren't detailed in the source reporting, but in our estimation they will likely follow the pattern used in other high-stakes autonomy domains, such as aviation or medical devices: narrow delegation first, wider autonomy only after a track record accumulates.
What this means for builders outside government
None of this stays confined to public administration. Any team wiring an agent into money movement, healthcare records, hiring decisions, or legal filings will hit the same delegation question — the UAE case is just the version playing out where the stakes are public and the deadlines are regulatory.
- Score every agent action by reversibility and impact before wiring it up to execute — not just to draft
- Decide in advance which actions require human confirmation, and enforce that in code rather than in a policy document
- Log the agent's reasoning and tool calls, not only its final output, so any decision can be traced back later
- Treat a model's confidence score as insufficient on its own to justify an irreversible action
AiiN's takeaway
Agentic AI's hardest problem was never getting a model to call the right tool — it's deciding, before deployment, which calls it's allowed to make without asking first. Governments are being forced to answer that question in public, because their mistakes are visible and litigated. Most companies get to answer it quietly, in a Slack thread, after something has already gone wrong. The UAE's classification effort is worth watching less for the specific tiers it lands on and more for the fact that it forces the delegation question to be answered explicitly, rather than left to whatever an agent framework happens to default to.