Khedmah - AI Chatbot

Short version: FNA Technology built and deployed a production-grade AI support chatbot for Khedmah, Oman's leading digital services platform. The agent deflects over 65% of incoming support tickets, handles bilingual queries (Arabic and English) natively, and executes real-time bill and order lookups via secure backend integrations, delivering a response time under two seconds.

The Bottleneck: Scaling Support for a Digital Nation

Khedmah is Oman's trusted platform for utility payments, mobile recharges, and government services. As user adoption surged, the volume of tier-1 support queries—predominantly users asking "Where is my order?" or "Has my bill been paid?"—began to overwhelm their human support team.

During peak hours, response times stretched, creating friction in an otherwise seamless digital experience. Traditional rule-based chatbots failed because they lacked conversational nuance, struggled with regional Arabic dialects, and required users to navigate rigid decision trees.

Khedmah needed an intelligent system that could interpret natural language in both Arabic and English, securely query user data, and provide instant resolutions—all while knowing exactly when to escalate to a human agent.

How Does the Khedmah AI Support Agent Work?

The AI support agent operates as a secure middleware layer between the customer (via WhatsApp or Web) and Khedmah's core backend systems. When a user submits a query, the system follows a deterministic pipeline to ensure accuracy and data security.

1. Intent Recognition: The incoming message is processed by a fine-tuned Large Language Model (LLM) to determine the user's core intent (e.g., check_order_status, general_faq, human_escalation).
2. Entity Extraction: The model extracts critical parameters like transaction IDs, phone numbers, or account numbers from the conversational text.
3. API Orchestration: The agent triggers secure API calls to Khedmah's backend systems to fetch real-time statuses without storing sensitive user data.
4. Bilingual Synthesis: The agent synthesizes a natural, context-aware response in the user's preferred language (Arabic or English) and delivers it in under two seconds.

Core Architecture and Technology Stack

To achieve high reliability and low latency, we designed a decoupled architecture prioritizing speed and security.

1. Conversational Engine (LLM & NLP)

We utilized a combination of LangChain and advanced Large Language Models for the conversational brain. By implementing Retrieval-Augmented Generation (RAG) with vector embeddings, the bot can accurately answer policy and FAQ questions by retrieving information directly from Khedmah's verified knowledge base.

2. Secure Backend Integrations

The agent does not rely on static answers for account-specific queries. It uses a secure Node.js and Python backend to execute live API queries against Khedmah's core systems. For example, if a user asks, "Did my electricity bill payment go through?", the agent authenticates the request, checks the transaction ledger, and replies with the exact status.

3. Multi-Channel Deployment via WhatsApp Business API

Recognizing that users prefer interacting on platforms they already use, we integrated the AI agent directly into WhatsApp using the WhatsApp Business API, in addition to the web and mobile app interfaces.

Overcoming Bilingual Challenges in AI Chatbots

A significant technical hurdle was ensuring the agent performed equally well in Arabic and English, particularly with Omani dialects and "Arabizi" (Arabic written in Latin characters).

Many out-of-the-box NLP solutions struggle with Arabic syntax and right-to-left context. We addressed this by implementing robust preprocessing pipelines and selecting LLMs that demonstrated high proficiency in cross-lingual transfer. The resulting agent natively understands context switching and can seamlessly handle a conversation that begins in English and shifts to Arabic.

Verifiable Metrics and Business Impact

The implementation of the AI support agent delivered immediate and measurable business outcomes for Khedmah.

• 65%+ Ticket Deflection Rate: Over 65% of tier-1 support queries are now resolved autonomously by the AI agent without human intervention.
• Sub-2-Second Response Time: Users receive instant, accurate answers to their queries, eliminating wait times during peak usage hours.
• 24/7 Availability: Customers can track orders and resolve issues at any time of day or night.
• Enhanced Human Agent Efficiency: By automating repetitive queries, Khedmah's human support agents are now freed to focus on complex, high-value customer interactions.

Comparing AI Agents to 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.	Fluent in English and regional Arabic dialects.
Data Integration	Usually disconnected from core systems.	Real-time API integration for account-specific answers.
Escalation	Drops context when transferring to humans.	Seamless handoff with full conversational history.