AI Agents and AI Chatbots for Business in Dubai

An agent is worth building when it owns the task

An AI agent earns the name when it owns a task end to end. A chat window that drafts a reply for a colleague to check is a writing aid. An agent that reads the order, finds the shipment, answers the customer in the language they wrote in and closes the ticket is doing a job. Only the second one empties a queue.

An agent worth building does three things: answers customers, reads and writes to the systems you already run, and hands over to a person at a boundary you define — in Arabic and in English. iSky also runs a separate AI chatbot product at isky-ai.com; for a packaged website assistant rather than a custom build wired into your back office, start there.

Which tasks are worth giving an agent

Pick work that is high in volume, low in variation, and already written down somewhere — a policy, a price list, a database table. An agent applies a known rule to a new case at three in the morning. It is bad at deciding what the rule should be.

• First-line enquiries. Hours, branches, availability, returns, delivery times, “where is my order”.
• Order and booking status, read live from the system rather than promised by email and chased the next day.
• Appointments — booking, rescheduling and cancelling against a real calendar, written back where staff see it.
• Document intake and routing. A trade licence, an invoice, a CV: read, classified, filed, passed to the right desk.
• Internal lookups — the questions staff ask that need three systems open to answer one.

What we would not start with: pricing negotiation, credit and approval decisions, complaints needing commercial judgement — anything where being confidently wrong is expensive.

Arabic and English in one agent, and why that is harder than it sounds

Most vendors write “supports Arabic” on the box. Here that claim has to survive how people actually type. It breaks in four places.

Dialect. Customers do not write to you in Modern Standard Arabic. They write Emirati, Khaleeji, Egyptian, Levantine, spelled the way the word sounds rather than the way the dictionary has it. An agent tested only on formal Arabic answers the polished question beautifully and misses the same question as a real customer types it. Test with messages from your own inbox.

Code-switching. A Gulf customer opens in Arabic, names the product in English, quotes a number in Latin digits, closes in Arabic. Some write Arabic in Latin letters throughout. The agent reads one message in two scripts, then picks a reply language — almost always the one the customer opened with, held to the end.

Names. One person is Mohammed, Muhammad, Mohamad and محمد across four fields of the same CRM. Search fails, duplicates appear, and the agent tells a returning customer it has no record of them. That is a data problem before it is an AI problem, fixed with transliteration-aware matching, not a bigger model.

Direction. Arabic runs right to left; order numbers, SKUs and email addresses do not. Mixed strings flip in transcripts, logs and exported reports, so a reference reads back wrong to whoever reviews the conversation. Handled in storage and display from the start it is a detail; patched afterwards it ruins every report you produce.

The systems it connects to

An agent that cannot reach your data is a search box with better manners. The value is in the connections: your store, booking system, CRM, ticketing tool, WhatsApp. iSky builds e-commerce platforms, mobile apps and custom systems for retail, automotive, events, media and transport across the UAE and the Gulf — the same kinds of systems an agent reaches into. Where an integration does not exist we build it; where the data underneath is messy, we say so first.

Set the escalation rule first

Define the handover in writing, first. A workable rule has four triggers: the customer asks for a person; the agent is not confident enough; the request falls outside its named scope; or the conversation touches money, a complaint or anything with legal weight. When one fires the agent stops, passes the whole conversation to a named inbox, and says a person is coming — in the language of the conversation. A silent handover that loses the history is worse than no agent.

What data it touches

Decide three things up front: what the agent may read, what it may write, and what it must never see. Give it the narrowest access that finishes the task, mask payment and identity details in transcripts, and set a retention period rather than keeping conversations forever. If your data has to stay inside the Emirates, that is a hosting decision made on day one, not a retrofit.

How you would know it is working

Agree the measures before go-live and read them weekly: how many conversations the agent finished without a human; how many it escalated and why; how fast the first reply arrives; and how often the same customer returns with the same question — the honest test of whether the first answer was good. The escalation log tells you what to build next. Add a sampled human review in both languages — no metric tells you the agent’s Arabic sounds wrong.

How we work

Four steps: scope one task, connect it to the systems it needs, agree the escalation boundary, then put the agent in front of real customers in both languages. You get a written scope and a price before any work begins. Tell us what the queue looks like today, and we will say plainly whether an agent is the right answer.