Governance
A model can be talked out of an instruction. A tool’s check cannot.
Summary: On the 10ⁿ Agentic Service Platform the controls sit in code the model cannot change. The tools that write to a record check identity, ownership and a confirmation before they act; a different person approves what an officer recommends; and every model call and tool call goes to an audit trail. This page lists each control and where it runs.
How AI governance works on the platform
| Control | Where it is enforced | In the government services demo |
|---|---|---|
| Approval before any write | In the tool: state-changing tools carry an “always require approval” mode and check the customer’s confirmation and UAE PASS step-up | Yes |
| Confirmation bound to the data | The confirmation is tied to a hash of the exact payload, so a changed field needs a new confirmation | No; built in another of our demos |
| Separation of duties | An officer recommends, a different approver decides; the tool and the platform both refuse self-approval | No; built in another of our demos |
| Ownership | Tools return only the records the system holds for the signed-in Emirates ID | Yes |
| One filing per draft | One idempotency key per draft and a lookup after any uncertain outcome | Yes |
| Facts from the source | Fees, rules and reference numbers come from the system; an answer quoting a number the system did not return is withheld and logged | Yes |
| Input screening | Prompt Shields, an Arabic and English blocklist and a phrase list; regulated requests answered without a model call | Yes |
| Planted instructions | Instructions hidden in uploaded documents are flagged to the reviewer and never followed | No; built in the case-management demo |
| Audit trail | Every model call and tool call, with model version, argument hash, outcome and latency | Yes |
| Hash chain | Audit rows chained by hash, so a changed row shows | No; built in another of our demos |
| Release gate | The test set runs three times per case; a second person approves; a failure blocks the release | No; built in another of our demos |
| Stop switch | Turns the customer assistant off without a deployment | No; built in another of our demos |
| Agent registry | Purpose, owner, tools, permitted data, approval points and prohibited actions for every agent | Partly: 17 roles listed |
| Production changes | Through a pipeline with human approval; coding tools never touch production data | Production design |
Approval sits in the tool
A prompt that says “ask first” is advice. On our platform the filing tool refuses to act until the customer has confirmed the exact details on screen and signed in with UAE PASS on the same session. The agent’s only way to write is through tools, and each tool carries its own checks.
Separation of duties
Where a service needs an officer’s decision, the officer recommends and a second person approves. In the demo where we built this, the approval tool refuses when the same person tries both, and the platform refuses again at its own interface, so a bug in one layer does not open the other.
An audit record of every call
Every model call is recorded with the deployment and the model version the service returned. Every tool call is recorded with a SHA-256 hash of its arguments, the outcome and the latency. In another of our demos the rows are also chained by hash, with an on-screen check that shows whether any row changed; production designs carry the chain.
Refused before the model sees it
Some requests never reach a model. An instruction to ignore the rules, a request to set a fee to zero or to approve one’s own application, in Arabic or English, is caught by the input screen and answered with a fixed line: fees come from the system of record, and the reviewing team decides. Each block is counted on the operations dashboard.
Designed against the UAE agentic AI framework
The UAE Government Agentic AI Project Operating Framework, issued by the National Committee for Agentic AI in August 2026, and its government experience design guide set out how agents may act on government services. For each service we prepare the evidence a reviewer will ask for: an agent action catalogue, a human and agent responsibility matrix, a risk register, the review and override procedure, accessibility tests in Arabic and English, and the identity design for UAE PASS.
Want to see it on your own service? We run the demo live with you, in Arabic or English.
Related: Agentic AI Platform, evaluation, human in the loop AI, AI outputs are advisory until someone signs.
Frequently asked questions
Can an agent change a record without a person?
Only after the customer confirms the exact details on screen, and only through a tool that checks identity and ownership first. Decisions the entity reserves for staff are never given to a tool the agent holds.
How do auditors check what an agent did?
They open the session in the audit browser. Each row shows the model version, the tool, a hash of the arguments, the outcome and the latency.
What happens when a model update changes behaviour?
Our release gate reruns the test set, three times per case, before the change goes live. A failure blocks the release, and a second person signs off the change.
Who is accountable for an agent’s output?
A named service owner at the entity for each service, and a named owner at 10ⁿ Tech for each agent in the registry.
