AI Agent Cost in the UAE
An AI agent has a build cost and a running cost, and the running cost is where most budgets go wrong. This guide separates setup, data work, integrations, evaluation, model usage and monthly operations, and sets out three scope tiers.
The short answer
The build cost of an agent is driven by what it is allowed to do: how many systems it reads from and writes to, how clean the knowledge it relies on is, and how strictly it has to be tested. The running cost is driven by volume: how many conversations it handles, how long they are and which model it uses.
An agent that answers questions from a tidy set of documents is a modest project. One that qualifies leads, checks inventory, books appointments and updates a CRM with approval rules is several times the work, mostly in integrations and testing, not in the AI itself.
This guide describes scope and how costs are calculated, not prices. A quotation follows discovery and states the agent's scope, tools, evaluation criteria, exclusions and running-cost assumptions.
Assumptions
- Quotations are in AED, with VAT added on top.
- The agent uses a commercial model through an API, billed by the provider on usage.
- Your systems have APIs we can use; building new ones is priced separately.
- Model prices change often, so we recalculate usage at quotation using the provider's published prices on that date.
Six separate cost lines
- Setup and design
- Process mapping, the agent specification, prompts, tone, refusal rules, the approval matrix and handoff design.
- Retrieval and data work
- Collecting, cleaning and structuring your knowledge sources, indexing them and setting ownership and update rules. Often underestimated when content lives in PDFs, chats and old spreadsheets.
- Integrations and tools
- Each action the agent can take is a function with validation, permissions and error handling. Every extra system adds build and test time.
- Evaluation
- Building the test set from real cases, scoring it, fixing failures and re-testing. Regulated or customer-facing agents need more of this.
- Model usage
- Paid per token to the model provider, every month, in proportion to volume.
- Monthly operations
- Hosting, monitoring, log review, content updates, re-evaluation after changes and support.
The first four lines are paid once, during the build. The last two continue every month the agent runs. Ask for them separately in any quotation you compare, because a low build price with no evaluation and no operations plan usually moves the cost into the months after launch.
How model usage adds up
Providers charge per token of input and output, and input is usually the larger share for agents. Each turn sends the instructions, the conversation so far, retrieved passages and tool results back to the model. A ten-turn conversation therefore costs much more than ten times a single question, because the context grows each turn.
A rough monthly estimate is: conversations per month, times average turns, times tokens per turn, times the price per token. We measure tokens per turn on your evaluation set rather than guessing. The levers that reduce it are shorter instructions, retrieving fewer but better passages, prompt caching where the provider supports it, trimming old turns and routing simple requests to a smaller, cheaper model.
| Input | Example value | Where it comes from |
|---|---|---|
| Conversations per month | Your volume | Current enquiry or ticket counts |
| Average turns per conversation | Measured in pilot | Evaluation set and early traffic |
| Tokens per turn | Measured in pilot | Instructions, context, retrieval and tool results |
| Price per million tokens | Set per model, separately for input and output | Provider's published pricing on the date of the quote |
Scope tiers
| Tier | What it includes | What pushes it up | Running-cost profile |
|---|---|---|---|
| Lean | A knowledge assistant for internal staff: read-only, one knowledge base of policies and product data, answers with sources, web chat, English only | Messy or scattered source documents, a second language, more than one knowledge base with different access rules | Mostly model usage and hosting, and the most predictable, because staff volume is steady |
| Standard | A customer-facing WhatsApp lead agent: English and Arabic, qualifies leads, answers from approved content, books slots, writes to the CRM, hands over to a person | Each extra system it writes to, complex approval or handoff rules, dialect-heavy Arabic, peaks in volume | Model usage, hosting and Meta's WhatsApp message charges, which can be the largest line |
| Advanced | A staff-facing operations agent: reads from 3 or more systems, drafts quotes or credit notes for approval, full audit log, detailed evaluation suite | Number of tools, approval rules by value and department, sector requirements for evaluation and data residency | Model usage with longer contexts, hosting, log storage and regular re-evaluation |
Running costs in every tier depend on volume, so we size them from your current enquiry or ticket counts and refine them with usage measured in the pilot. The Standard tier also carries Meta's message charges, explained in the WhatsApp automation cost guide. We give a fixed quote for the build after a scoping call, with the running-cost assumptions written down next to it.
Monthly operations in detail
| Item | What it covers | How it is calculated |
|---|---|---|
| Model usage | Provider token charges at your volume | Conversations times turns times tokens per turn, priced at the provider's published per-token rates for input and output. Billed by the provider monthly. |
| Hosting and data | Application hosting, vector or search index, logs storage | Cloud usage billed monthly: compute, database and index size, and log storage, which grows with volume and how long logs are kept. |
| Channel fees | WhatsApp message charges or telephony, where used | Meta charges for WhatsApp messages at rates it publishes by message category and the recipient's country; a BSP may add a markup. Telephony is billed per minute by the carrier. |
| Monitoring and evaluation | Log review, re-running tests after changes, quality reports | Engineering time each month, which depends on how often prompts, content or models change. |
| Maintenance and support | Content updates, API changes in connected tools, fixes | A monthly support agreement sized by the number of connected systems and the response times you need. |
Ongoing cover is available through AI monitoring and support.
Where agent budgets go wrong
- Scoping too many tools in the first release. Two or three well-tested actions beat ten fragile ones.
- Skipping data work. An agent grounded in outdated price lists gives confident, wrong answers.
- Treating evaluation as optional. Without a test set, every prompt change is a gamble.
- Choosing the largest model by default. A smaller model often passes the same tests at a fraction of the usage cost.
- Forgetting the channel. WhatsApp message fees, BSP fees or telephony can exceed model usage.
- No spending caps. A looping agent or a spike in traffic should hit a limit, not your card.
If your process has a fixed path, a rules-based workflow costs less to build and much less to run. The AI agent pilot checklist helps you decide what belongs in a first release.
Questions buyers ask
Why not just use an off-the-shelf chatbot subscription?
For answering from a website or FAQ, a subscription tool may be enough. Custom agents make sense when the agent must act in your own systems, follow your approval rules or keep data in your infrastructure.
Will model prices go down?
Prices per token have generally fallen over time, but newer, more capable models are often priced higher. We design so the model can be swapped, and we recalculate usage when you switch.
Can we cap monthly spending?
Yes. We set usage limits at the provider and in our own code, with alerts before the cap is reached and a defined behaviour, such as handing over to staff, when it is.
Is a pilot cheaper than a full build?
Yes, because it limits tools, channels and volume. It also produces real usage data, which makes the full-build and running estimates far more reliable.
Does Arabic cost more?
Arabic text can use more tokens than the same content in English with some models, and Arabic evaluation cases take extra time to build and review. Both are included in our estimates when Arabic is in scope.
Tell us the process, the systems the agent would touch and your monthly volume, and we will return a scoped estimate with build and running costs separated. See our AI agent development service and pricing.
Ready to build your growth system?
Send a short brief or message us on WhatsApp. We reply with questions, a suggested scope and the sensible next step.