The UAE Cabinet has 32 AI advisors. None of them votes

Short answer. The UAE Cabinet has launched a system of 32 specialised AI advisors. They read the documents, weigh the consequences, check them against global practice and keep decisions moving around the clock. The minister decides. A twenty-person company can copy the arrangement almost unchanged, and the arrangement is the part worth copying. The number is beside the point.

News like this normally arrives as a number and nothing else. This time the working arrangement came with it: who does what at each step.

What did the Cabinet launch?

The Cabinet AI Advisor. It helps the Cabinet prepare its decisions.

On 2 September the UAE Cabinet approved a system of 32 specialised AI advisors. Sheikh Mohammed bin Rashid Al Maktoum chaired the session at Qasr Al Watan, the first of the new government season.

The official description gives them six jobs:

1. Read policy and draft legislation.
2. Work out what it costs in money, for the economy, for people and for the environment.
3. Check it against global practice and against what matters nationally.
4. Find what the Cabinet has already decided on the subject.
5. Prepare a recommendation from an approved set of sources.
6. Keep decisions moving with ministers around the clock.

The same session approved an AI curriculum for every school, state and private, training for 22,000 teachers, and a framework for driverless vehicles.

What the description leaves out: who supplies the system, how accurate it is, and what happens when it gets something wrong. We are reading the design here. Nobody has shown a result yet, and judging one now would be premature.

Who makes the decision?

The minister. The system prepares advice and keeps the work moving, and the choice stays with a person.

The name says it all. It carries the word Advisor, and the word Decider appears nowhere. The phrase "presents recommendations to ministers" draws exactly the line that gets argued over in every other company. Where the assistant ends and the person who answers for it begins.

Break it into steps and see what fits a company of twenty.

Two rows out of six stayed with a person, and those two are what the salary is for. The other four come down to gathering, checking, writing up and reminding. They are what eats the week, and they go first.

Test it on yourself. If your founder spends five minutes choosing the option and three days on the other four steps, you already know where the agent goes. The same line runs inside a single enquiry, and we have a separate piece on that.

Why 32 advisors at once?

Because a narrow assistant gets things wrong less often than a broad one.

One assistant for the whole government would have to learn everything at once: finance, the environment, transport, schools. Thirty-two specialised advisors solve that problem differently. Each has its own area, its own sources and its own criteria.

A twenty-person company needs a different number. Two or three will do. The principle is the same: each has an area, a set of documents and a written note on what it does and what it stays away from.

Picking them is easy. Take the three places where people wait longest for an answer. In most companies that means incoming enquiries, document preparation, and internal questions about how work gets done here. Each gets its own agent with its own rules. No separate platform is involved: an agent here is a set of rules and documents on top of an ordinary model, living in the same messenger you already type in. Those agents are what an AI team is made of.

The pull towards one clever helper for everything is strong and almost always ends the same way. The helper answers everything passably and nothing well. A month later people stop opening it.

What does "an approved set of sources" do?

More than the rest of the description put together.

It means the advisor answers from the documents somebody selected and signed off. Whatever else turned up does not count. The difference shows up in the first contested case. When advice comes out of open search, you cannot check it, and where the number came from is anyone's guess. When it comes from an approved set, you open the source and look.

The minimum set in a small company looks dull. A price list, the discount rules, contract templates, a description of the services, answers to common questions. Five documents, each with an owner and a date. That is all.

The other half of the same question sits right next to it. The set of sources decides what the agent sees, and access rights decide what it can change. We have a separate piece on rights, and without it the first half runs on empty.

What changes when a system watches day and night?

The gap between deciding and doing disappears, and that gap is where things get lost.

The picture is familiar. The meeting agrees on something, someone writes it down, everyone leaves. Two weeks later two people understood it differently and a third was waiting to be contacted. Formally the decision stands, and in practice nothing moved.

Watching without a pause closes that gap. A robot is no stricter than a person, it simply does not forget and does not feel awkward asking a third time. "Let us look at it next week" stops working on its own.

The downside is visible straight away and worth naming. A system that reminds you about everything equally often turns into noise inside a week, and noise gets switched off. One rule fixes it: the agent tracks only the decisions that have a deadline and a name attached. It leaves the rest alone.

Your first step looks like this. Take the minutes of your last meeting and cross out everything with no deadline and no name. What survives is what you can hand over. About a third survives, in our experience, and that on its own is a useful discovery.

The government showed a design. Nobody has seen a result: accuracy figures are not public, the supplier has no name, and a year from now this system will be judged on other things. But the arrangement in it is right, and copying it costs nothing. An agent gathers, checks and reminds. A person chooses and answers for the result. You can start this week, with one process and five documents.