Dubai opens its AI Sandbox: real-world trials with regulators, and a route to exemptions

For companies selling AI in the UAE and the wider GCC, this is a commercial signal more than a policy one. Dubai is competing on time to deployment. If your blocker in education, healthcare, transport, logistics, tourism, media or government services is a rule written before your product existed, the AI Sandbox programme page sets out a documented route: real-world trials alongside government entities, including the Dubai Centre for Artificial Intelligence, with technical, policy and commercial support and tailored regulatory guidance. Where the use case warrants it and the relevant authorities approve, the programme can explore exemptions. For enterprise buyers in the region, a completed trial is procurement evidence: someone has already argued your case to the people who write the rules. The mistake most teams will make is treating the sandbox as funding or a badge. It is neither. The entry bar is explicit: MVP stage or beyond, commercialisation readiness, and a use case that generates evidence policy makers can use.

The scaling problem is well measured. The 74% figure behind the launch, the share of companies that have yet to move AI beyond experimentation, comes from the World Economic Forum's AI in Action work. The obstacles it names are not technical readiness alone: regulatory requirements, data access and policies that do not fit emerging technology. Adoption is not the constraint in the Gulf. Governance-ready deployment is.

What regulators will actually need is evidence instrumentation. Risk assessment wants eval sets, pinned model and dataset versions, failure rates and an audit trail of what the system did and who approved it. And a specific barrier beats a vague one every time. A grading product blocked from Dubai schools because the requirements around it are undefined will be read; a claim that AI regulation is unclear will not.

Two concrete examples of the shape we would build. First, put a governed gateway in front of every model call so the trial evidence exists by construction: FastLLM Proxy is an OpenAI-compatible gateway in front of your own LLM servers and 80 hosted providers, with routing, budgets and access control in one place, logging every request and response. Second, run the agents and models on your own GPU servers with Kuvryn AI, each workload in its own space, so student, patient or passenger data stays on infrastructure you control. Put a human at the decision point, the pattern we build into NovaGrade, where exam answers are marked criterion by criterion and teachers review instead of mark. Gate the engineering with Procoder so untested or unfinished work cannot pass as done. A minimal sketch of the pattern:

# Trial architecture: auditable by default
gateway:          # FastLLM Proxy, OpenAI-compatible entry point
  upstreams: [self-hosted-llm, hosted-fallback]
  access_control: sandbox-evaluation-team
  budgets: { monthly_usd: 5000 }
  audit_log: request + response, retained for the trial period
platform:         # Kuvryn AI: each agent in its own space, on your GPUs
  human_review: required before any scored decision

Where teams get it wrong, twice: picking a use case that is technically interesting but carries no genuine regulatory barrier, so they are not selected, and building the demo before the evidence architecture, so they spend the trial window fixing logging and versioning. Azrty is based in Dubai, and our team here does the consulting, the deployment and the support. We would scope one barrier, instrument the trial from day one, and write the risk and evidence pack before writing a line of application copy.