What the published rates actually say
Every hiring guide on this subject opens with a rate. It is worth starting there, if only because the published rates for the same role in the same city do not agree with each other.
| Model | Published figure | What it does not include |
|---|---|---|
| Freelance AI developer | AED 120–350 per hour, or roughly AED 15,000–40,000 a month | Your review time, and continuity when they take a salaried job |
| Freelance AI developer, second published range | USD 50–150 per hour, roughly AED 185–550 | The same role, quoted about 60% higher at the top of the range |
| In-house AI engineer | AED 25,000–60,000 a month | Roughly 6–12% in visa, insurance and gratuity on-costs |
| Agency or studio project | AED 50,000–250,000+ per project | Whatever the scope document left undeclared |
Put the two published ranges together and the price of the same job in the same city runs from AED 120 to about AED 550 an hour, more than four times end to end. That is not a market you can optimise by shopping the rate. It means the rate is not describing the same work, and comparing on it will mislead you.
The thing you are buying is not the model
The model is a line in a config file and it will be replaced within the year. What makes an AI feature work in production, and what survives the model being swapped, is everything around it.
- An evaluation set: a written list of real inputs and the answers you consider correct. Without it, nobody can tell whether a change made the system better or worse, and every release is an argument about vibes.
- The retrieval layer: what the model is allowed to see, kept current. Most disappointing AI projects are retrieval projects that were priced as model projects.
- Operations: cost per call, latency, what happens when the provider is down, and what the system does when it does not know.
- A human review path, because a system that scores, ranks, prices or screens people needs one under UAE law regardless of how good it is.
Ask which of those four a single hire covers in their first quarter. Usually one, sometimes two. That is the honest case for commissioning a first system, and equally the honest case for hiring once the system exists and needs tending.
When hiring is the right answer
- The AI work is continuous rather than a project: new use cases arrive every month from inside the business.
- Somebody on your side can already tell a good answer from a plausible one, and will make time to review.
- The data the model needs is already accessible in systems you control, so the first six months are not a data engineering project in disguise.
- You can keep the person busy. A specialist hired for one build becomes an expensive generalist afterwards.
When commissioning is the right answer
- You need one specific system working by a date, and the date matters more than owning the capability.
- Nobody internally can review AI output yet, so the estimate risk should sit with whoever can control it.
- The build spans retrieval, evaluation, integration and compliance, which is a team's worth of surface area for a first system.
- You want the capability afterwards. Commission it with the handover written into the scope, then hire into a system that already runs.
The line neither model covers by default
Both roads end at the same unanswered question. Federal Decree-Law No. 45 of 2021 gives a person the right to human review of a decision made about them by an automated system, which means an appeal path and a decision log are features somebody has to build. It appears in almost no hiring brief and almost no AI quote, and it is scope either way.