On 1 September, Dubai Chambers launched specialised agentic AI training tracks for more than 14,000 companies, delivered online through a new e-learning platform, Dubai Chambers Academy. The training is aimed at member companies of the business groups and business councils that sit under Dubai Chamber of Commerce. It is one piece of a plan Sheikh Hamdan bin Mohammed announced in May: moving Dubai's private sector to agentic AI within two years.
If your company is one of the 14,000, your people are about to learn what AI agents can do. That is the easy half. The harder half is what happens the week after, when someone connects an agent to the company inbox, the CRM or the customer WhatsApp line. Training does not decide whether that goes well. Five decisions made before it happens do.
What was actually announced
The detail matters, because the headlines ran ahead of it.
- Who: member companies of Dubai Chamber of Commerce's business groups and councils. Not every licensed company in Dubai.
- What: specialised agentic AI training tracks on Dubai Chambers Academy, an online platform. The programme has no separate brand name.
- Why now: a Dubai Chambers executive committee for agentic AI, formed in June, covers training tracks, incubators and dedicated funds.
- Not yet published: cost, course length, syllabus and how non-members can join. We found no official source for any of these, and anyone quoting them should show where they came from.
The government is moving faster still. In April the UAE set a target to move 50 percent of government sectors and services to agentic AI within two years, and in May the Cabinet approved training 80,000 government employees, its largest training programme ever. Businesses that deal with government will be dealing with agents on the other side of the counter.
The country is ready for it. Microsoft's AI diffusion report put the UAE first in the world in the first quarter of 2026, with 70.1 percent of the working-age population using AI. The question for a business is no longer adoption. It is control.
Chat is not agentic. That is the whole point
A chatbot answers. An agent acts. It reads the email, decides what it means, updates the record, drafts the reply and, if you let it, sends it. That is where the productivity is, and it is also where every risk lives. The training will cover the first half well. Here is the second half.
1. List the tasks, not the tools
Start with a one-page list of the repetitive work your team actually does: qualifying leads, answering the same five customer questions, chasing invoices, preparing weekly reports. For each one, write what a correct outcome looks like. An agent built against a clear task is measurable. An agent bought because the tool looked good in a demo is not.
2. Map the data before anything connects
For every task, write down which data the agent will read and where that data will go. This is not paperwork. The UAE's Personal Data Protection Law prohibits processing personal data without consent, subject to listed exceptions such as performing a contract. It also restricts transfers of personal data to countries without adequate protection, and most AI services process data outside the UAE. If your agent handles customer names, numbers or messages, you need to know which of those rules applies before switching it on.
Two caveats. Companies in DIFC and ADGM fall under their own data protection laws instead. And on the day we checked, the official legislation portal listed no executive regulations for the federal law yet. When they arrive, the law gives businesses no more than six months to comply. Building to the law now is cheaper than retrofitting later.
3. Split read from write
Give agents read access generously and write access grudgingly. Reading the inbox and drafting a reply is low risk. Sending the reply, issuing a refund, changing a customer record or messaging a lead is not. Every write action should either sit behind a human approval or have a hard limit, such as a maximum amount, a whitelist of recipients, or office hours only.
4. Keep a human in the decisions that affect people
This one is written into the law. Article 18 of the PDPL gives individuals the right to object to decisions made by automated processing, including profiling, and requires a human review of those decisions when the person asks. An agent that scores leads, rejects applications or sets a customer's price is making exactly that kind of decision. Design the review step in from the start, so a person can see what the agent decided and why.
It is also the principle the country has already published. The UAE Charter for the Development and Use of AI lists human oversight among its twelve principles, and the federal government's own agentic AI programme runs on the idea that the human leads and AI enables.
5. Log every action
Every action an agent takes should be recorded with what triggered it, what it did and what it changed. You will not read the log daily. You will need it the first time a customer says "your system sent me this" and you have ten minutes to find out why. A log is also the only honest way to decide, after a month, whether an agent has earned more autonomy.
If you sell to government, check the AI Seal
Dubai added a procurement rule many businesses have missed. The Dubai Centre for AI's Dubai AI Seal certifies AI companies, and in October 2025 Dubai Finance asked government entities to restrict AI-related service and consultancy contracts to Seal-certified companies. If government is a customer, or you want it to be, the Seal is now part of the tender, not a badge.
What we would do first
- This week: write the one-page task list and the data map. Two hours, one person, no tools.
- Before the training: decide who in the company owns agent permissions. One name, not a committee.
- Before the first agent goes live: read-only first, a human approval on every write action, and logging switched on.
- After a month: read the log, and relax limits only where it proves the agent does not need them.
We set out the full control framework in what to automate in 2026 and how to do it with control, and the legal picture in what actually applies to your business. For why agent permissions became the industry's biggest worry this month, see our piece on Anthropic's call to pace the frontier.
Training gives your team the vocabulary. Building agents that do useful work inside your systems, with permissions you can read on one page, is the brief of our AI automation work. Tell us which tasks you want to hand over and we will tell you which are safe to automate first.