Most vendors will not put a number on a page. You get "it depends on your requirements", a discovery call, and a proposal two weeks later that you have nothing to compare it against. So here are ours, in dirhams, before you talk to anyone.
An AI agent build in Dubai typically costs between AED 10,000 and AED 50,000. Where you land inside that range depends on how many systems the agent touches and how much judgment it has to exercise. At the pegged rate of about 3.67 dirhams to the dollar, that is roughly USD 2,700 to 13,600. Running costs sit on top and are billed separately, by third parties, every month.
What does an AI agent cost in Dubai?
These are our published bands. They are what we quote against, which is the only sense in which any price list is true.
| Band | Scope | Range (AED) | Timeline |
|---|---|---|---|
| Single-workflow agent | One trigger, one or two systems, human escalation path | 10,000 to 20,000 | 2 to 3 weeks |
| Multi-system agent | Three or more systems, branching logic, error handling | 20,000 to 35,000 | 3 to 5 weeks |
| Complex agent | Custom logic, compliance constraints, self-hosted model, or heavy integration surface | 35,000 to 50,000 | 5 to 8 weeks |
Two things to note about the shape of that table. The timelines are calendar weeks, not billed effort, and they assume you can answer questions within a day or two. And the ceiling is a real ceiling: this is a value position, not an enterprise one. If a project genuinely needs six months and a dedicated team, it is not an agent build, it is a software programme, and it should be scoped and priced as one.
What actually moves the price?
Five things, in roughly the order they matter.
1. How many systems the agent touches. Each system is a set of credentials, a data model to learn, a rate limit to respect, and a failure mode to handle. Going from two systems to four does not double the work, but it is the single clearest predictor of which band you land in.
2. Whether those systems have documented APIs. HubSpot, Xero, QuickBooks, Stripe, Slack, monday.com and Asana all publish real API documentation, and integrating them is predictable work. A twelve-year-old internal system with no documentation and one person who remembers how it works is a different exercise. So is anything that only exports a CSV to a shared drive at 6pm. When a quote from us moves into the complex band, an undocumented internal system is the most common reason.
3. How much judgment the agent exercises. "Route this to the sales team if the subject line contains a price" is a rule, and rules are cheap. "Read this supplier email and work out whether it is a complaint, a reorder, or a payment query, and draft the right reply" is judgment, and judgment needs evaluation: a set of real examples, a measured pass rate, and a decision about what happens on the cases it gets wrong. That evaluation work is invisible in a demo and is a large part of the price.
4. Compliance and data-residency constraints. If your legal or banking counterparty requires that customer data never leaves a specific jurisdiction, a hosted model API may be off the table, and you are then paying for a self-hosted model and the infrastructure to run it. That single constraint is enough on its own to move a build into the top band. Ask the question early, because discovering it in week three is expensive.
5. How wrong an answer is allowed to be. This is the most underestimated driver on the list. An agent that drafts an internal summary can be wrong sometimes and nobody is harmed. An agent that posts entries into your accounting system cannot. Lowering the acceptable error rate means confidence thresholds, a review queue, an audit log, and an escalation path that a human actually monitors. You are not paying for a better model. You are paying for the machinery around it that catches the model when it is wrong.
What do you pay every month after the build?
Running costs are separate from the build fee, they are not included in the bands above, and they are billed to you by third parties rather than by us. There are three of them.
Model API usage, charged per token. Anthropic's published rates: Claude Sonnet 5 is USD 2 per million input tokens and USD 10 per million output tokens under introductory pricing through 31 August 2026, moving to USD 3 and USD 15 after that; Claude Haiku 4.5 is USD 1 and USD 5; Claude Opus 5 is USD 5 and USD 25. Anthropic's hosted web search tool is charged separately at USD 10 per 1,000 searches. (Verified July 2026 against platform.claude.com/docs/en/about-claude/pricing.) OpenAI's published rates for comparison: gpt-5.4 is USD 2.50 per million input tokens and USD 15 per million output tokens, gpt-5.4-mini is USD 0.75 and USD 4.50, and gpt-5.4-nano is USD 0.20 and USD 1.25. (Verified July 2026 against developers.openai.com/api/docs/pricing.)
Those numbers are meaningless until you convert them into your own volume, so here is the arithmetic. Say one agent step reads a 2,000-token email and writes a 300-token reply. On Claude Sonnet 5 at the introductory rate that is USD 0.004 of input plus USD 0.003 of output, so USD 0.007 per run, about 2.6 fils. At 3,000 runs a month it is USD 21, or about AED 77. Real agents usually make more than one model call per task, so multiply by the number of steps rather than assuming one.
Hosting for the service that runs the agent. A DigitalOcean Basic droplet is USD 6 per month for 1 GB of RAM, 1 vCPU and 25 GB of SSD, USD 12 for 2 GB, and USD 24 for 4 GB with 2 vCPUs. (Verified July 2026 against digitalocean.com/pricing/droplets.) In dirhams that is roughly AED 22 to AED 88 a month. Most single-workflow agents fit comfortably in the smallest tier; a self-hosted model does not, and that is a different order of cost entirely.
Per-message platform fees, if the agent talks to customers. WhatsApp is the one that matters in the UAE. Since 1 July 2025 Meta charges on a per-message basis rather than per conversation. Marketing templates are charged on every delivery; utility and authentication templates are free inside an open customer service window and charged outside it; non-template service messages are free and can only be sent inside that window. Rates depend on the recipient's country, the template category and your volume tier, and Meta publishes them as downloadable rate cards including one denominated in AED. (Verified July 2026 against developers.facebook.com/docs/whatsapp/pricing.) We could not verify a specific UAE per-message figure from Meta's own page, because the page links the rate card rather than printing the numbers. Read the current AED rate card before you budget, and do not take a per-message number from a reseller's blog post. It is the single most frequently out-of-date figure in this whole category.
What is not included, and where projects go over
The band you are quoted covers the agent we scoped with you. It does not cover the following, and each of these has pushed a real project past its estimate.
- Undocumented internal systems. If the agent has to talk to something with no API and no documentation, the work of establishing how it behaves is not in the estimate until we have seen it. This is why we ask for system names in the first conversation rather than the third.
- Data that turns out to be dirty. Duplicate customer records, three spellings of the same supplier, invoice PDFs that are scans rather than text. An agent inherits every data problem you already had, and it surfaces them faster than a human did. Cleaning that is a project of its own, and it is usually yours rather than ours.
- Rules that change mid-build. Every stakeholder who joins in week three arrives with an exception nobody mentioned. Two or three of those are normal and absorbed. A steady stream of them means the process was not agreed before the build started, and the honest response is to stop and write the rules down rather than keep billing.
- Licences and third-party subscriptions. API tiers, connector licences, and the model and hosting costs above. Those are yours, in your name, on your card.
- Changes after handover. Ongoing changes and monitoring run on a monthly retainer, agreed separately. We would rather say that plainly than bury unlimited future changes in a fixed price that then has to be defended.
When is an AI agent not worth the money?
Often enough that a cost guide which never says "do not buy this" is just a price list. Five situations where we would tell you not to.
The task does not happen often enough. Do the arithmetic before anything else. A task that runs 20 times a month and takes 4 minutes is 80 minutes a month, about 16 hours a year. Even valuing that time at AED 200 an hour, you are automating AED 3,200 of annual effort with a build that starts at AED 10,000. It repays in year four at the earliest, assuming nothing about the process changes, which it will. Volume is what makes agents pay, and low-volume tasks almost never clear the bar.
Nobody can write the rules down. If the answer to "how do you decide which supplier gets priority" is "it depends, ask Fatima", you do not have a process to automate yet. An agent can encode judgment, but somebody has to articulate the judgment first. Trying to discover the rules during a build is the most expensive way to write them.
A wrong answer is expensive and nobody will check the output. This combination is the one to watch for. High stakes with a review step is a fine project. Low stakes with no review is a fine project. High stakes with no review is how an agent quietly posts wrong numbers into your books for a month. If you cannot name the person who will check the queue, do not build it.
An existing tool already does it. If the workflow is a straight line from trigger to action, Zapier or your CRM's own workflow builder will do it faster and cheaper, and we will tell you so. We have written the full version of that argument in Zapier vs custom integration, including the part where a custom build does not pay for itself out of subscription savings. The reason to build custom is logic you need to test and version, latency a polling interval cannot hit, or failure handling you have to own. It is rarely the invoice.
The underlying system is about to be replaced. If you are migrating your ERP next quarter, an agent built against the old one is disposable work. Wait, or build the integration layer against the new system.
A worked example, from scope to number
This scenario is illustrative. It is representative of the kind of work we do, not a description of a specific client engagement.
A distributor in Dubai receives supplier invoices as PDF attachments into a shared mailbox, roughly 800 a month. Someone opens each one, keys the header fields into Xero, matches it against the purchase order, and either files it or emails the buyer about a discrepancy.
Scope. One trigger, the mailbox. Three systems: the mailbox, Xero, and Slack for notifications. Branching logic, because a variance over 5 per cent goes to a human instead of being posted. Error handling, because a supplier will eventually send a scanned image instead of a text PDF, and the agent has to route that somewhere rather than guess. Documented APIs on both Xero and Slack.
Band. Three systems with branching and error handling puts it in the multi-system band: AED 20,000 to 35,000 over 3 to 5 weeks. Because the APIs are documented and the variance rule is a written policy rather than a judgment call, it sits in the lower half of that band. An indicative quote would be AED 26,000 to 28,000.
Running cost. Assume two model calls per invoice: one to extract the fields from the PDF, one to compose the discrepancy email when needed. At roughly 3,000 input tokens and 400 output tokens per call on Claude Sonnet 5 introductory pricing, that is about USD 0.01 per call, USD 0.02 per invoice, USD 16 a month at 800 invoices, about AED 59. Add a 2 GB droplet at USD 12 a month, about AED 44. Call it AED 100 to 150 a month on published rates, before anyone sends a WhatsApp message.
What would push it up. If a third of those invoices arrive as scans, add OCR and a review queue for low-confidence extractions. If the finance team requires an audit trail showing which invoices the agent posted without human sign-off, add the logging and the interface to review it. Either one moves the quote toward the top of the band; both together move it into the complex band. This is the same intake-versus-decision split we described in our write-up of an email triage agent: the intake is cheap and the decision-making is where the money goes.
How to brief a vendor so the quotes are comparable
If you get three quotes for the same agent and they are AED 12,000, AED 30,000 and AED 60,000, the usual explanation is not that one vendor is greedy and another is cheap. It is that all three answered different questions, because all three were given a paragraph rather than a brief. The cheap quote assumed documented APIs, clean data and no review queue. The expensive one budgeted for all three.
Write one page and send the same page to everyone. It should contain:
- The named systems the agent must touch, and for each one whether it has a documented API, an export, or neither. Not "our CRM". The product name.
- The trigger and its volume. What starts the agent, and how many times a day or month that happens today.
- What the agent decides versus what it escalates. One sentence for each. If you cannot write that sentence, that is the finding, and it is worth more than any quote.
- The acceptable error rate, and who checks. Name the role that owns the review queue. If the answer is nobody, say so, and expect the honest vendors to push back.
- Any data-residency or compliance constraint, stated up front rather than discovered in week three.
- A request for the build price and the monthly running cost as separate lines, with the model and hosting assumptions written out. A vendor who will not separate them is either not sure what the running cost is or does not want you comparing it.
- A request for what is excluded. The exclusions list tells you more about a proposal than the price does.
Then read the three quotes against each other rather than against your budget. The one that asks the most uncomfortable questions before quoting is usually the one that has built this before. If you want the systems layer underneath the agent scoped at the same time, that is what our API integration service covers, and the broader operational picture sits under AI automation. If you want a number for your specific case, send us the one-page brief above and we will price it against the bands on this page.