Khedmah - AI Chatbot

Short version: FNA Technology built a production AI customer support chatbot for Khedmah, Oman's digital services platform. It understands Arabic and English natively, looks up bill payments, services and orders in real time through secure backend integrations, runs on WhatsApp, web and mobile, and hands complex cases to a human agent with the full conversation history attached.

Project Facts at a Glance

Client	Khedmah, Oman's digital services platform (khedmah.om)
Built by	FNA Technology
Project type	AI customer support chatbot (LLM + RAG)
Year	2024
Languages	Arabic (including regional dialects) and English, with mixed-language support
Channels	WhatsApp Business API, web, mobile apps
AI & NLP	Large Language Models, LangChain, vector embeddings, intent-recognition models
Backend	Node.js, Python, PostgreSQL
Integrations	WhatsApp Business API, Khedmah core systems
Escalation	Confidence-based handoff to live agents with full conversation context

The Challenge: Scaling Support for a Digital Services Platform

Khedmah is Oman's trusted platform for digital services, bill payments and order tracking. As more customers came onto the platform, the volume of routine questions grew with it. Most of them were variations of "Where is my order?" and "Has my bill been paid?"

The support model was human-only, and that created bottlenecks during peak hours. Customers expected an instant answer, and a queue was the opposite of what a digital-first service should offer. Traditional rule-based chatbots were not a realistic fix. They depend on rigid decision trees, they handle regional Arabic poorly, and they cannot fetch live account data.

The brief for FNA Technology had three parts:

• Understand bilingual queries in Arabic and English, including dialects and mixed-language messages.
• Connect securely to Khedmah's backend to fetch real-time user data.
• Hand complex issues to human agents without dropping context.

How Does the Khedmah AI Support Agent Work?

The agent sits as a secure middleware layer between the customer (on WhatsApp, web or mobile) and Khedmah's core systems. Every message follows the same four-step pipeline, which keeps answers accurate and sensitive data protected.

1. Intent recognition: Specialized intent-recognition models classify what the customer wants, for example checking an order status, asking a general question, or requesting a human.
2. Entity extraction: The model pulls out the details needed to act, such as a transaction ID, phone number or account number, from natural conversational text.
3. API orchestration: The agent calls Khedmah's backend through secure APIs to fetch the live status, without keeping sensitive user data in conversation logs.
4. Bilingual response: The agent writes a clear, contextual reply in the customer's language, Arabic or English.

If a request falls below a set confidence threshold, or the customer asks for a person, the agent escalates instead of guessing.

Architecture and Technology Stack

We designed a decoupled architecture so each layer can be improved or scaled on its own, with speed, reliability and data security as the priorities.

Conversational engine: LLMs, LangChain and RAG

The conversational layer combines Large Language Models with LangChain for orchestration. For policy and FAQ questions, we use Retrieval-Augmented Generation (RAG) with vector embeddings. The bot retrieves the answer from Khedmah's verified knowledge base rather than improvising, which keeps responses grounded in approved content.

Secure backend: Node.js, Python and PostgreSQL

Account-specific questions never get a static answer. A Node.js and Python backend, with PostgreSQL, runs live queries against Khedmah's core systems. When a customer asks whether a payment went through, the agent authenticates the request, checks the transaction record and returns the actual status. Queries run over encrypted APIs, and data is held only long enough to compose the reply.

Multi-channel delivery: WhatsApp Business API

Customers prefer to ask for help where they already are. We integrated the agent with the WhatsApp Business API and deployed the same agent across web and mobile apps, so the experience and the logic stay consistent on every channel. If you are planning something similar, see our WhatsApp chatbot development services.

Solving the Bilingual Challenge: Arabic and English

Arabic support was the hardest technical requirement. Many off-the-shelf NLP tools struggle with Arabic syntax, regional dialects and right-to-left context, and customers often switch languages mid-conversation.

We handled this with preprocessing pipelines for Arabic input and by choosing models with strong cross-lingual performance. The agent understands mixed-language queries and replies in the language the customer is using. We apply the same approach to our Arabic chatbot development work for businesses across the GCC.

Privacy-First Data Handling

Support conversations about bills and orders involve sensitive information, so the agent was designed around privacy from the start. It does not store personally identifiable information (PII) or transaction data in its conversational logs. Every data query runs through an encrypted API, and the information is held in memory only long enough to compose the response.

This matters for two reasons. It limits what could be exposed if conversation logs were ever accessed, and it keeps Khedmah's core systems as the single source of truth. The agent reads live data, answers, and moves on, rather than building a second copy of customer records.

Human Handoff: Knowing When Not to Answer

A support bot is only as trustworthy as its willingness to step aside. The agent uses a confidence-scoring mechanism on every query. When confidence falls below a predetermined threshold, or when the customer explicitly asks for a person, it triggers a handoff to a live support representative.

The transfer carries the entire conversation history. The human agent sees what was asked, what the bot answered and which details were already provided, so the customer does not have to start again. This is the difference between a chatbot that deflects customers and one that actually resolves their issues.

What Is Next: Voice Notes

The agent currently processes text-based interactions. Because many WhatsApp users send voice notes, the architecture was designed so that audio transcription models, such as Whisper, can be added in a future iteration to handle voice queries without redesigning the pipeline.

Results: What Changed for Khedmah

The chatbot now handles a significant share of tier-1 support queries on its own, which has drastically reduced average response times and improved the overall customer experience.

• Instant answers on routine queries: Order, bill and service status questions are answered by the agent in real time instead of waiting in a queue.
• 24/7 availability: Customers can track orders and resolve issues at any hour, across WhatsApp, web and mobile.
• Fewer repetitive tickets for the team: Human agents are freed to focus on complex, higher-value conversations.
• No lost context on escalation: When a person does step in, they receive the full conversation history, so the customer never has to repeat themselves.

AI Agents vs. Traditional Rule-Based Bots

Feature	Traditional Rule-Based Bots	FNA AI Support Agent
Interaction model	Rigid decision trees and keyword matching.	Natural language understanding and context awareness.
Language support	Often limited; struggles with dialects.	Arabic (including regional dialects) and English, with mixed-language handling.
Data integration	Usually disconnected from core systems.	Real-time API integration for account-specific answers.
Knowledge answers	Fixed scripted replies.	RAG over a verified knowledge base.
Escalation	Drops context when transferring to humans.	Confidence-based handoff with full conversational history.

Build a Similar AI Support Agent

The Khedmah project is one example of how we deliver production AI agents: senior engineers, secure integrations with your existing systems, and a design that knows when to hand over to a human. To see the wider capability, explore our AI chatbot development services and AI agent development.