TLDR: Ryt AI is an LLM-native agentic framework powering Ryt Bank, the world’s first regulator-approved AI-native bank. It enables customers to execute core financial transactions (like fund transfers) through natural language conversations, replacing traditional multi-screen workflows. Built with an in-house LLM (ILMU) and specialized agents (Guardrails, Intent, Payment, FAQ), it prioritizes security, compliance, and human oversight, demonstrating that AI can reliably handle high-stakes banking operations.
A groundbreaking development in retail banking is reshaping how customers interact with their financial institutions. A new research paper introduces Ryt AI, an advanced AI-driven framework that allows users to perform core banking transactions simply by speaking or typing in natural language. This marks a significant shift from traditional, often cumbersome, multi-screen banking processes.
Ryt Bank, based in Malaysia, is at the forefront of this innovation, being the world’s first licensed bank to fully integrate an AI-native architecture. Its core, Ryt AI, is not just another chatbot; it’s the first globally regulator-approved system where conversational AI acts as the primary interface for executing critical financial operations. Previous AI assistants in banking were largely confined to advisory or support roles, never directly handling transactions like fund transfers or bill payments.
The Brain Behind Ryt AI: ILMU and Its Agents
At the heart of Ryt AI is ILMU, a powerful, closed-source Large Language Model (LLM) developed entirely in-house by Ryt Bank. This internal development was a strategic choice, driven by strict regulatory requirements in the financial sector that demand verifiable control over training data, inference processes, and updates. ILMU is specialized for banking, ensuring efficiency, compliance, and low latency.
Ryt AI operates through a modular, multi-agent system, where four specialized LLM-powered agents work together to orchestrate banking tasks:
- Guardrails Agent: This is the first line of defense, screening all user inputs for safety and compliance. It blocks malicious attempts, harmful language, or privacy breaches, ensuring secure interactions.
- Intent Classifier Agent: After passing the guardrails, this agent identifies the user’s intention (e.g., payment, inquiry, FAQ) from their natural language input, even if it’s informal or includes typos. It then routes the request to the appropriate action agent.
- Payment Agent: This agent handles all financial fund transfers. It extracts crucial details like recipient name, bank, account number, and amount from natural language or even images (using integrated OCR for bills and receipts). It then validates these details against banking rules and prepares the transaction for user confirmation.
- FAQ Agent: For general inquiries, this agent uses a Retrieval-Augmented Generation (RAG) pipeline to provide accurate and context-aware answers from a curated knowledge base, even reformulating vague questions for better results.
Banking as a Conversation, Not a Chore
Imagine transferring money without navigating through five to eight screens. Ryt AI transforms this experience into a simple dialogue. Users can express their needs in natural language, and the system intelligently processes, validates, and prepares the operation. For instance, a request like “Transfer RM100 to John for lunch” is understood, and the system prompts for any missing details before presenting a summary for approval. This conversational approach significantly reduces the time and friction associated with digital banking workflows.
The system also supports multimodal inputs, meaning you can even send a chat screenshot or an image of a bill, and the Payment Agent will extract the necessary transaction details, making banking more flexible and user-friendly.
Ensuring Trust and Security
Given the high-stakes nature of finance, Ryt AI is built with robust safeguards. A “human-in-the-loop” mechanism ensures that all critical transactions, especially fund transfers, require explicit user confirmation. Depending on the amount, this might involve two-factor authentication (2FA), adhering strictly to Malaysian financial regulations. This design gives users full control and provides a verifiable audit trail.
Furthermore, a stateless memory architecture means that sensitive user data and conversational context are never persistently stored after a session ends. This minimizes security risks and aligns with data minimization principles, crucial for privacy and compliance.
Also Read:
- RouteLLM: Intelligent Navigation Powered by Hierarchical AI Agents
- ReInAgent: A Collaborative AI for Smarter Mobile Task Automation
Real-World Impact and Future Outlook
Ryt AI is not just a theoretical concept; it’s a live system already serving over 50,000 users and processing approximately 80,000 transactions monthly. Its high language proficiency allows it to understand diverse inputs, including multilingual and informal expressions, which is vital for reliable banking operations. The system boasts a low hallucination rate, especially in high-stakes transactional flows, thanks to its strict validation and human confirmation steps.
While Ryt AI represents a significant leap, the paper acknowledges certain limitations, such as the substantial investment required for in-house LLM development, potential regional and linguistic biases outside its primary focus (Bahasa Melayu, English, Chinese, and Malaysian regulations), and the resource-intensive nature of model updates. However, these are seen as opportunities for continuous improvement.
This innovative framework demonstrates that regulator-approved natural language interfaces can reliably support core financial operations under strict governance, setting a new standard for user experience in banking. For more details on this pioneering work, you can refer to the research paper.
