What the two-year agentic AI window asks of a 20-person company.
On 4 May 2026 the Crown Prince of Dubai launched a two-year programme to move the emirate’s private sector to agentic AI. Most of the coverage has been about the city. This note is about one company in it: owner-run, twenty people, real revenue, running on spreadsheets and a dozen group chats. What does the window actually ask of that owner, and in what order?
What was announced
The Dubai Media Office announcement sets a two-year horizon for the private sector’s shift to agentic AI, described as self-executing and self-leading artificial intelligence, with the aim of making Dubai the world’s leading city in adopting these technologies economically and commercially. The stated objective is to empower companies to adopt technologies that boost productivity and expand business volumes. The instruments named are specialised training tracks for the business councils affiliated with the Dubai Chamber of Commerce, incubators for agentic AI companies and dedicated funds. Gulf Business summarised the same package, with the emphasis on productivity and lower operating costs. Khaleej Times reported that Dubai Chambers would train more than 14,000 private sector companies in using agentic AI for daily business tasks.
Nothing in the announcement obliges any particular company to do anything. It is a direction, with training and support attached, and a date by which the city hopes to be able to point at results. The agenda page sets out the wider context, including the D33 economic agenda and the Universal Blueprint for Artificial Intelligence, with the public sources linked. Here the question is narrower.
What it asks of a 20-person company
Read plainly, the window asks the owner to make the company run on systems that act, rather than on people who remember. Agentic AI, as distinct from the AI tools most people already use, is software that takes steps on behalf of the business: it reads the pipeline, raises the overdue item, drafts the follow-up, posts the cost against the job, flags the drift. For it to do any of that, the pipeline, the jobs, the costs and the people have to exist somewhere it can read them. In a 20-person company they usually exist in the owner’s head and in chats, and that is the real ask hidden inside the announcement. Before a company can run on agentic AI, it has to run on a system at all.
This is why the training is the right first instrument and also why it is not sufficient. A course can teach an owner what agentic AI is and show them tools that do it. It cannot move the operation out of their head. That part is work, and it is specific to each company. The agentic AI page sets out the distinction at more length.
A sequence to follow
For a company of this size, the order matters more than the speed. Done in the wrong order, agentic tools are bolted onto a business that still routes everything through one person, and they add to the noise rather than reducing it. Done in the right order, each step makes the next one smaller.
- Map the operation against one real job. Not a questionnaire. Walk a recent job from first contact to final follow-up and write every step and every hand-off on a wall, in order, with the point at which each thing is dropped. For most owners this is the first time the business has been described in full, and it is where the rest of the plan comes from. The method page describes how we run this session.
- Find the joins. The map will show where a person carries something between two tools or two people: the quote that becomes a job, the delivery that should become a debrief, the past client who should become a call. These joins are where the owner’s week goes and where agentic software earns its keep. Count them.
- Get the records into one place. Pipeline, work, money and people, as one record per job rather than fragments across tools. Whether this is bought or built is the subject of the build or buy note; for a 20-person company with a specific way of working it is usually built, and the operating system page describes what that looks like. Either way, nothing agentic can act on records it cannot read.
- Give the system one executive, not many bots. One senior AI that runs the system, changes hats as the question requires, and escalates to a named person, with no authority of its own. A pile of disconnected agents recreates the joins problem in software.
- Make the things that never happen structural. The alert before the deadline, the debrief after every job, the call at three and six months, the prediction set against the actual, the flag when a relationship goes quiet. This is where the productivity the announcement talks about actually comes from in a small company: not from doing the same work faster, but from doing the work that was never done.
- Train the team on the system, not on AI in general. Once the company runs on a system, the Chambers’ training and others like it land on something. Before that, they land on the owner.
Where to start
Start with the map, and start with the job that went wrong most recently. It will show you more than any general diagnostic. If you want to do it yourself, take a day, a wall and the people who actually run the work, and do not stop until the whole operation is on the wall in order. If you want it done with you, that session is the first thing we do with every company and the method page says what it involves.
Two years is long enough for a 20-person company to do all six steps properly and have a business that runs without its owner at the end of them. It is not long enough to do them twice. The companies that will be able to point at results when the window closes are the ones that started with the operation rather than the tools, and the ones that have already done so are described under work in production.
Written by the founders of Symbaiotic, Dubai. See who we are, and more writing.