Build cost is only the first number
Published Dubai ranges for custom AI work, from NomadX's 2026 breakdown:
| What you are building | Build | Per month to run |
|---|---|---|
| Customer-support agent (chat, WhatsApp, email) | AED 55,000 – 200,000 | AED 3,000 – 18,000 |
| Internal operations agent | AED 120,000 – 350,000 | AED 5,000 – 25,000 |
| Finance or reconciliation agent | AED 200,000 – 550,000 | AED 8,000 – 40,000 |
| Multi-agent orchestration | AED 450,000 – 1,200,000 | AED 20,000 – 70,000 |
Note the right-hand column. The rule of thumb is that running an AI system costs 15 to 30 per cent of its build cost every year, indefinitely.
As with every published price guide, these come from agencies selling the service and sit at the upper end of what an SME would actually spend. Read them as the shape of the problem rather than your quote.
Why the monthly number surprises people
Conventional software costs almost nothing to run once it is built. AI is different, and the difference catches finance teams out.
Language models charge per use. Every question asked, every document read, every reply generated consumes tokens you pay for. An assistant handling fifty conversations a month costs very little. The same assistant handling five thousand costs real money — and the better it works, the more it is used.
Then there is the part nobody budgets for: models change. Providers retire versions roughly every six to twelve months, and what was tuned for one model behaves differently on the next. Migration is a recurring engineering cost, not a one-off.
So the only number worth asking for is three years, not one: build, plus thirty-six months of running, plus the migration you know is coming.
Build or buy
A great deal of AI work should not be built at all. Off-the-shelf tools now cover customer support, meeting notes, document search and sales assistance at AED 90 to 550 per user per month.
Buy when your problem is one most businesses have. Someone has already built it, tested it on far more data than you have, and maintains it for you.
Build when the task is specific to how your business works, when it must connect to systems nobody else has, or when the data genuinely cannot leave your environment. That is where custom work earns its price.
Watch the per-seat maths, though: at AED 350 per seat across 120 people, a SaaS tool costs about AED 504,000 a year. Above a certain headcount, building becomes the cheaper answer — so do the arithmetic for your own numbers before assuming either way.
How to spend less and still get it
- Start with one task and one number. Not an AI strategy. One task that costs you money, and a figure to beat. One cycle is enough to find out whether it works.
- Measure before you build. How many hours a week does this take now? What is the error rate? Without a baseline you cannot tell whether it worked, and most failed AI projects failed by being unmeasurable rather than by being bad.
- Use a smaller model where it is enough. Model costs vary by an order of magnitude. Sorting and classifying rarely needs the most expensive option; the architecture should route the cheap work to the cheap model.
- Check whether AI is the answer at all. A deterministic rule, a better form, or fixing the process upstream often beats a model — cheaper to run, easier to trust, and it does not need migrating next year.
- Get the running cost in writing before you sign. In dirhams per month, at your expected volume. Any supplier who cannot estimate it has not thought about your project properly.
How we price it
The build is quoted per twenty-eight day cycle, with the number agreed before work starts. We estimate the monthly running cost up front, in dirhams at your expected volume, because discovering it after launch is how AI projects get switched off.
A first cycle is deliberately a small commitment: one task, one measurement, one working result. If it beats the number, expand it in the next cycle. If it does not, you spent one cycle finding out rather than a year. Bring us a task that costs you money and roughly what it costs you now.