# AI agent development for businesses in Iraq > An AI agent is software that understands a request in natural language, then decides what’s needed and actually does it: checks your stock, books an appointment or records an order in your system, then replies. At Al-Arjan we build Arabic agents connected to your data and systems, with defined permissions, human handoff when needed, and a full log of everything they did. Source: https://www.al-arjan.com/en/services/ai-agents · العربية: https://www.al-arjan.com/ar/services/ai-agents.md ## How is an AI agent different from a regular bot? **A regular bot replies from a menu or fixed script. An agent understands the question however it’s phrased and uses real tools — your database, booking system, accounting — to get the task done, not just answer about it.** Ask a traditional bot “do you have this medicine?” and it sends you a phone number. Ask an agent and it checks stock at your nearest branch, tells you availability and price, and offers to reserve it in your name. The difference isn’t nicer language — it’s the ability to act inside your systems. That’s why every agent project starts with one question: which task has your staff repeatedly answering questions whose answers already exist in your data? That’s where an agent makes the biggest difference. ## What can an agent do in your company? **Any repetitive task with clear rules and available data: answering customers, qualifying leads, booking, order tracking, answering staff questions, and extracting data from documents.** - WhatsApp customer service answering from your real data, 24/7. - A sales agent that asks the right questions, sends the quote and logs the lead in your CRM. - A booking agent that offers genuinely free slots, confirms and reminds. - Order and shipment tracking by order number without staff involvement. - An internal knowledge assistant answering staff from company policies and procedures. - Extracting data from invoices, contracts and forms into your system for review. ## How do we build an agent that doesn’t make things up? **With three layers: the agent answers only from your data and documents (RAG), it can only call specific tools with specific permissions, and it hands off to a human when it doesn’t know or the matter is sensitive.** Large language models are fluent, but left unconstrained they can “invent” a confident answer. So we never rely on the model’s general memory for anything about your business: prices, availability and policies are pulled live from your systems and documents, and the agent is instructed to say “I don’t know — let me connect you with a colleague” rather than guess. Before launch we test the agent on hundreds of real questions from your past customer conversations, measure answer accuracy, and iterate until results reach a level you sign off. - Answers grounded in your data and documents, with sources tracked internally. - Specific tools only: read stock, create booking, log order — no open permissions. - Sensitive actions (refunds, discounts, data changes) require human approval. - Smooth handoff to staff with a conversation summary. - A full log of every conversation and every action the agent took. ## Where does your company’s and customers’ data go? **Together with your team we choose where processing happens based on data sensitivity: cloud models from major providers configured so your data isn’t used for training under their terms, or open-source models running on your own servers for sensitive data.** We send the model no more than it needs for the task, mask unnecessary personal data before processing, and set retention periods for conversation logs. For hospitals and government bodies we often prefer running an open-source model inside the organisation’s own infrastructure. ## How do we measure an agent’s success? **With numbers you can review monthly: share of conversations resolved without staff, human handoff rate, answer accuracy in random audits, team time saved, and customer satisfaction.** - Self-resolution rate - Handoff rate and reasons - Answer accuracy in an audit sample - Average response time - Sales or bookings completed by the agent - Post-conversation customer rating ## How long does an agent take and what does it cost? **A pilot agent for one task on one channel usually takes 3 to 6 weeks. Cost has three parts: build and integration, monthly model usage by conversation volume, and channel fees such as WhatsApp.** We always recommend starting with one clear task and measuring it, then expanding to more tasks and channels. Details are in our article on bot and agent costs. ## Features - **Arabic, including Iraqi dialect**: Understands MSA and dialect, replies in your brand’s voice, tested on real conversations. - **Works inside your systems**: Connected to stock, bookings, CRM and accounting via defined tools. - **Answers from your data**: RAG over your documents and databases instead of the model’s general memory. - **Human handoff**: Instant handoff to staff with a summary when needed or requested. - **Permissions & logs**: Every action scoped, logged and auditable. - **The right model**: Leading cloud models or open-source on your servers, case by case. ### Traditional bot vs AI agent vs staff member | Criterion | Menu bot | AI agent | Staff member | |---|---|---|---| | Understands free-form questions | Weak | Strong | Strong | | Performs tasks in systems | Very limited | Yes, within permissions | Yes | | Availability | 24/7 | 24/7 | Working hours | | Complex & sensitive cases | No | Hands off to a human | Yes | | Cost per extra conversation | Very low | Low | High | ## Technology Claude, Gemini, GPT (OpenAI), Llama, Mistral, LangChain, Python, FastAPI, Node.js, PostgreSQL, Qdrant, n8n ## Frequently asked questions ### Does the agent understand Iraqi dialect? Modern models handle Iraqi dialect well in most cases, and we test on your customers’ real conversations before launch to fix weak spots. ### Can the agent make mistakes? Yes, like any system — so we constrain it to your data and specific tools, require human approval for sensitive actions, and audit conversation samples regularly. ### Does the agent replace customer service staff? It works best taking the repetitive questions so staff can focus on cases needing human judgement and relationships. ### Which channels does it work on? WhatsApp, Telegram, Messenger, Instagram, your website, your app and internal staff tools. ### Is our data used to train models? We choose providers and settings that don’t train on your data under their published terms, or run an open-source model on your servers. ### Can we start with a small pilot? Yes, and we recommend it: one task, one channel and clear metrics, then expand. --- Contact Al-Arjan: +964 773 545 2792 (phone & WhatsApp, 24/7) — [Request a quote](https://www.al-arjan.com/en/quote)