AI Agent Development for Dubai & UAE Businesses
We build AI agents that read your approved data and take controlled actions in your systems: looking up a booking, drafting a quote, updating a CRM record. Every action runs through defined tools, permissions and approval rules, and every decision is logged.
What we mean by an AI agent
An AI agent is a language model connected to a small set of tools it is allowed to use. Given a request, it decides which tool to call, reads the result and decides what to do next, until the task is done or it hands over to a person. The tools are ordinary functions we write: search the knowledge base, check stock, create a draft invoice, book a slot.
That is the difference from a chatbot, which only answers, and from a workflow, which follows a fixed path. An agent is worth building when the path depends on what the customer or employee says, and when finishing the task means touching more than one system. If the path is always the same, workflow automation will be cheaper and more predictable.
An agent is only as safe as the tools and permissions it is given. We design those first and choose the model second.
Who this fits
- Sales teams in real estate, car rental or tourism that need enquiries qualified, matched to inventory and booked at any hour.
- Operations teams answering status, availability and policy questions that require looking something up in two or three systems.
- Finance and admin teams preparing drafts (quotes, credit notes, supplier replies) that a person approves before they go out.
- Businesses with a clear process, documented rules and at least a few hundred similar requests a month.
It fits poorly when the data is scattered and out of date, nobody owns the process or every case needs a senior judgement call. Then AI consulting is the better first step.
Deliverables, scope and exclusions
- Agent specification
- The tasks the agent handles, the ones it refuses, its tone, escalation rules and the approval matrix, agreed in writing before build.
- Knowledge retrieval
- An index of your approved sources (policies, price lists, FAQs, product data) with ownership, update rules and citations back to the source.
- Tool layer
- Typed functions for each action, each with its own credentials, input validation and rate limits.
- Approval and handoff flows
- Queues where staff approve, edit or reject drafted actions, and a route to a named person when the agent cannot proceed.
- Evaluation suite
- Test cases from your real requests, scored automatically where possible and by your team where judgement is needed.
- Logs and dashboard
- A trace of every conversation, tool call and approval, with cost per run and error rates.
- Runbook and handover
- How to update content, change prompts safely, rotate keys and respond to incidents.
- Not included unless scoped: cleaning or migrating source data, building a new CRM or ERP, voice channels (see AI voice agents), and ongoing content writing for the knowledge base.
- We do not build agents that make final decisions on credit, hiring, medical or legal matters without a person approving them.
How a request moves through the agent
Request arrives
From WhatsApp, a web widget, email or an internal tool. The channel passes the message and any known identity, such as a CRM contact ID.
Context is loaded
The agent receives its instructions, the conversation so far and only the customer data its permissions allow.
Approved knowledge is retrieved
Relevant passages are pulled from your indexed sources. The agent is instructed to answer from them and to say so when they do not cover the question.
A tool is chosen
The agent proposes a tool call. Our code validates the inputs and checks the permission before anything runs.
Approval, if required
Actions above a threshold (a discount, a refund, a message to a group) are held for a person, who sees the proposed action and the reasoning.
Action and log
The tool runs, the result is returned to the agent and the whole step is written to the log.
Fallback
If a system is down, confidence is low or the customer asks for a person, the conversation goes to a human with a summary attached.
Stack choices
| Component | Options | How we choose |
|---|---|---|
| Model | Commercial models from providers such as Anthropic, OpenAI or Google; open-weight models on your own infrastructure | Accuracy on your evaluation set, cost per request, latency, Arabic quality and where data may be processed |
| Retrieval | Postgres with pgvector, a managed vector database, or keyword search with reranking | Size and structure of the content, how often it changes and whether you already run Postgres |
| Orchestration | Custom TypeScript or Python, or an agent framework where it saves real effort | We prefer plain code for production agents with few tools, because it is easier to test and audit |
| Hosting | Your cloud account, including UAE regions where required, or ours under a support agreement | Data residency needs, existing IT policy and who will own operations |
| Channels | WhatsApp, web chat, email, Slack or Teams, internal admin screens | Where the requests already arrive today |
Testing, evaluation and failure handling
Before launch, we collect a set of real requests (anonymised where needed) with the correct outcome for each, including awkward ones: mixed Arabic and English, missing details, requests the agent should refuse. The agent must meet the agreed pass rate on that set before it serves customers, and the set is re-run after every change to prompts, tools or model version.
In production, we watch for the failures that matter: wrong tool calls, answers without a source, repeated handoffs on the same topic and cost per conversation creeping up. Tool calls are idempotent where the target system allows it, so a retry does not create a duplicate booking. Rate limits and spending caps stop a loop from running up a bill. For deeper assurance work, see AI evaluation and testing.
Delivery process and proof
Discovery
Process walk-through, system access review, data audit and a draft specification.
Pilot build
A narrow agent with two or three tools, tested with your staff as the only users.
Evaluation and hardening
Pass-rate testing, permission review, approval flows and monitoring in place.
Staged launch
A share of real traffic first, then full rollout once error and handoff rates are stable.
Book a call and we will show a working agent and the traces behind it. The AI agent pilot checklist lists what to prepare before discovery.
What drives the cost
- Number of tools and systems the agent writes to, and how good their APIs are.
- State of the knowledge sources: clean and owned, or scattered across PDFs and chat history.
- Languages, especially Arabic and mixed Arabic-English input.
- Approval complexity: one approver, or rules by value, department and customer type.
- Evaluation depth required by your sector or risk appetite.
- Monthly volume, which sets model usage and hosting costs.
The AI agent cost guide separates setup, data work, integrations, evaluation, model usage and monthly operations, and sets out three scope tiers.
Questions buyers ask
Can the agent make mistakes?
Yes. Language models can misread requests or state things confidently that are wrong. That is why answers are grounded in your approved sources, risky actions need approval, and we measure error rates before and after launch rather than assuming them away.
Which model will you use?
The one that passes your evaluation set at an acceptable cost and meets your data requirements. The model can be changed later without rebuilding the agent.
Does it work in Arabic?
Current commercial models handle Arabic well for most business tasks, including Gulf dialect in informal messages, but quality varies. We include Arabic and mixed-language cases in the evaluation set so you see real results before launch.
Who can see the conversations?
Access to logs is role-based and set with you. We keep personal data in logs to what is needed for troubleshooting, with a retention period you approve.
What happens if the agent is down?
Messages fall back to your normal inbox or queue with a clear flag, and our monitoring alerts us. Customers are not left without a reply.
Ready to scope an agent? Send us the process, the systems involved and a monthly volume, and we will reply with a pilot scope. Get a project proposal.
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.