On 18 November 2026, Dubai approved its City AI Agent initiative and launched 93 specialised AI agents for human resources, finance and procurement. According to Gulf News, the agents are expected to run more than 34 million tasks a year across 85 government entities, save over 812,000 working hours and cut costs by more than Dh151 million.
Programmes of this size change the question leaders need to ask. A chat assistant answers. An agent acts: it raises a purchase order, updates a record or sends a message. Once software acts on your behalf, the question moves from "is the answer good?" to "who is responsible for what it did?"
Organisations already know how to answer that question for people. Every new joiner gets a role, a manager, a set of system permissions and a review. The same four things work for an agent, and they should be written down before day one.
Write the job description first
One page is enough. If the team cannot write it, the use case is not ready to build.
- Scope. The tasks the agent performs, the systems it works in and the volume expected each week.
- Limits. What it must never do. For example, approve a payment above a set value, or contact a customer without review.
- Hand-offs. The cases it passes to a person, and the name of the team that receives them.
Name one supervisor
One named person answers for the agent's work. Not a committee, and not the supplier. That person signs off the go-live, sees the exceptions and has the authority to pause the agent the same day. An agent without an owner tends to become nobody's problem within a few months.
Give it only the access the job needs
Treat the agent as you would a new joiner on their first morning.
- Its own identity. The agent never borrows a member of staff's login, so every action can be traced to it.
- Least privilege. Read and write access only to the systems and data named in the job description.
- A full record. Every action is logged, so the supervisor can replay what happened 30 days later.
Review performance against a baseline
Before the agent starts, measure how the work is done today: time per case, error rate and cost. Agree the thresholds the agent must meet. Test it on real cases at real volume, then review it on a schedule: weekly in the first month and monthly after that.
In the Gulf, test in Arabic as well as English. An agent that performs well in English can fall short on Arabic documents, dialect and requests that mix both languages. Build the evaluation set from your own Arabic cases and ask Arabic speakers to score the results.
Start with one agent, then repeat
The first agent takes the most effort, because the templates do not exist yet. Once the job description, the access pattern and the review cycle are written, the tenth agent can reuse them. Teams that skip these steps get a pilot that impresses in a demo and then stalls at security review.
A checklist for day one
- Job description. One page, signed by the business owner.
- Supervisor. One named person with the authority to pause the agent.
- Identity and access. A separate identity with access to the listed systems only.
- Baseline. Current time, error rate and cost per case, with thresholds agreed before the build.
- Evaluation set. Real cases in every language the agent will work in.
