Short answer. On 5 October the Dubai Future Foundation launched an accelerator for agentic AI in government services. Its first phase brings together 33 government entities and assesses more than 300 public-facing use cases to pick the ones worth building. Phase one is sorting, not development. A twenty-person company has about twenty candidate processes, and an evening is enough: five tests, and two of them disqualify most of the list.
A city with a budget spent its first phase deciding what not to build. A company of twenty usually starts with a tool and finds the process afterwards.
What did Dubai actually launch?
A programme whose first phase is entirely selection.
On 5 October 2026 the Dubai Future Foundation announced the Agentic AI for Government Services Accelerator, run under Dubai Future Accelerators. The Dubai Centre for Artificial Intelligence organises it, with Digital Dubai and the Dubai Electronic Security Centre.
The figure that matters sits in the first phase. It brings together 33 Dubai Government entities and assesses more than 300 public-facing use cases to find the ones with the most potential for development on agentic AI. Nobody is building three hundred services. Thirty-three organisations are sifting three hundred candidates down to the few that will earn the work.
Two other details carry over. Participating entities work with technology companies certified under the Dubai AI Seal, which is a procurement filter applied before anybody writes code. And the security centre's chief executive, Yousuf Hamad Al Shaibani, put its position on sequencing plainly.
"Cybersecurity cannot be an afterthought in the adoption of AI. It must be embedded at every stage, from planning and design to operation and monitoring."
One caveat. All that exists so far is the announcement, and the Dubai Future Foundation's own case for the programme is that entities will be able to shorten traditional development cycles substantially. That is the organiser describing its own initiative before it has produced anything. The order of operations is the part to copy. It costs nothing and does not depend on the results.
Why does the order matter more than the tool?
Because a tool chosen first turns the selection into a search for something it can do.
The common sequence in a small company runs backwards. Somebody sees a demo, buys or installs it, and then looks for work to point it at. The process that gets picked is whatever fits the tool, and the usual winner is the most visible one: the inbox, the reporting pack, the thing the founder touches daily.
Visibility is a bad selection criterion. It tells you what annoys the person choosing. It says nothing about where the hours are or whether the work can be checked. We went through the narrower version of this decision, agent or plain workflow for a single job, in do you need an agent or a workflow. This is the step before that one: which job to look at in the first place.
The asymmetry is what makes the order worth fixing. Sorting twenty processes takes an evening. Choosing wrong costs a quarter, because the failure arrives slowly and in a form that looks like a tooling problem.
Which five tests do you run?
Five, and the first two disqualify most of the list.
| Test | How to check it tonight | What "no" means |
|---|---|---|
| Is there a recorded right answer? | Find 30 past cases where you know what correct looked like | You can only say whether you like the output, never measure it |
| Does the process leave a trace today? | Try to reconstruct last week's runs from what is written down | No baseline, so no improvement you can show |
| How often, and what does one run cost? | Count runs per week, time one honestly | Rare and cheap work does not repay a build |
| What does a wrong output cost, and to whom? | Name the person who absorbs it | If the client absorbs it, the bar is much higher |
| Does a person check before it reaches a client? | Look at the current path, not the intended one | No checkpoint, so no safe place to start |
Test one is the hard filter, and it is the same archive test we worked through in tested on history first. Without recorded outcomes there is no way to tell a good agent from a fluent one.
Test two catches more processes than people expect. You cannot measure a process that leaves no trace, before or after, so any claim of improvement is a feeling. That is also the cheapest thing on this list to fix.
Test four changes the answer more than any other row. The same process scores differently depending on who pays for a mistake. An internal draft that a colleague corrects is a reasonable place to start. A message that reaches a client unchecked is a bad one, whatever the demo looked like.
What if nothing passes?
Then the first build is a record, not an agent.
That is the normal outcome, and the team that brought the list should hear it from you. If nothing on the list has recorded outcomes and nothing leaves a trace, the first project is to make one process leave a trace: who did what, when, and what the result was. That work is cheaper than an agent. It produces the baseline any later measurement needs, and the agent would have needed it anyway.
The security centre's line about sequencing applies at this scale too. Deciding what an agent may reach belongs in the design, at the moment of connection, and a review afterwards is too late to do it. We described what that looks like in practice in what access an agent should have when nobody owns security.
Hold on to the list your team brought. The processes that fail test two this quarter can pass it next quarter, once they leave a trace, and the list is the only record of what you considered and why you said no.
What this means for your operation
- Sorting comes before tooling. A city with a budget spent phase one on it, which is the part that transfers.
- Pick the test that disqualifies, not the process that annoys you. Recorded outcomes and a visible trace are the two filters that do the work.
- The same process scores differently depending on who absorbs a wrong answer. Name that person before you decide.
- "None of them yet" is a real answer. The work it points at is a record, and it is cheaper than what you were about to buy.
- Keep the rejected list. It is the only written record of the decision.
We connect agents to mail, CRMs and document stores, and this sorting is where every one of those projects starts. What that looks like is on the AI integration page. You can run it without us: write down your processes, and mark the ones where you could name thirty past cases with a known right answer.
Danil Ivanov
Founder, KAIVIX
Builds AI systems for companies in the UAE and beyond.