Ryt Bank taps agentic AI for conversational banking

Forget the frustration of tapping through multiple screens to send some money to a friend. Since August 2025, Malaysia’s Ryt Bank has been enabling consumers to perform banking transactions through Ryt AI, its homegrown artificial intelligence (AI) assistant.

According to the digital bank, which is regulated by Malaysia’s central bank, Ryt AI can understand and process requests such as ‘send 50 ringgit to Siti’ or mixed-language expressions commonly used in the multilingual country, directly initiating fund transfers and other banking services.

While more banks have started to employ AI assistants, their capabilities have been limited to answering questions or providing account information. By contrast, Ryt AI has the authority to act on user instructions, marking the first time a regulator has approved conversational AI as the primary interface for core banking transactions.

“We were thinking, how do we build a bank from the ground up in the age of generative AI?” said Foong Chee Mun, chief product officer of Ryt Bank. “Instead of the rigid, menu-based user interface, we wanted users to use natural language to interact with our core banking system.”

According to an academic paper co-authored by Foong to be presented at the Empirical Methods in Natural Language Processing conference in China this week, the use of Ryt AI can reduce a typical fund transfer – which might take 30 to 45 seconds across five to eight screens – to a single conversational exchange.

Ryt AI’s capabilities are powered by ILMU, a proprietary large language model (LLM) developed entirely in-house by YTL AI Labs, the research arm of Malaysian conglomerate YTL Group, which partnered with Southeast Asian tech giant Sea Limited to establish Ryt Bank.

Foong said the company chose to build its own LLM, as off-the-shelf models from major tech companies such as Meta and others struggled with the linguistic diversity of Malaysia.

“When you fine-tune one of those models to be good in Malay, they might get dumber in something else, and so it becomes a case of whack-a-mole,” he explained. “We eventually gave up and decided to train a model from scratch,” he said, adding that ILMU is adept at handling conversations with a blend of English, Malay, and Chinese within a single sentence.

However, the decision to build an LLM was more than just a technical preference. In financial services, regulators typically demand financial institutions to have control over the entire AI model lifecycle to meet compliance and audit requirements, which can be difficult to achieve with third-party and open-source models.

Open-source LLMs, for one, have “unresolved challenges” such as opaque training data provenance, ambiguous legal responsibility, and vendor update cycles misaligned with jurisdictional compliance, limiting their use in core banking, Foong and his co-authors noted in their paper.

To ensure ILMU’s accuracy, Ryt Bank engaged Chemin – a Singapore-based firm that specialises in AI data sourcing, model refinement and training, as well as workflow design – to develop an AI testing framework. Chemin’s role was not to train the AI but to curate and verify a benchmark dataset of “exam questions” for the model.

This involved creating tens of thousands of prompt-and-action pairs, from simple commands related to paying someone to more complex, multi-step conversations, allowing Ryt Bank to rigorously evaluate the model’s performance and identify weaknesses, much like a teacher grading a student’s exam answers.

“If it’s scoring more than 80%, then you know it’s generally quite good. But if it’s less than 60%, then that’s bad,” Foong said, adding that new, targeted training data will be used to retrain the model in weak areas.

To address the risks of AI hallucinations and comply with regulatory requirements, Ryt AI is built on an LLM-native, multi-agent system that acts like an assembly line with built-in safety checks.

A guardrail agent first screens user inputs for malicious or inappropriate requests. An intent agent then determines what the user wants to do before passing the command to a payment agent to execute the transaction. An FAQ agent is also on hand to provide contextual responses to user queries.

More importantly, all transactions require a final, human-in-the-loop confirmation from the user before any money is transferred. This ensures users retain full control over financial decisions.

Through Ryt AI, Ryt Bank is already serving over 50,000 users and processing about 80,000 transactions a month. According to Foong, the conversational feature has helped with customer retention.

“We see in the statistics that repeated users of these AI features are quite sticky,” he said. “Hopefully that will remain true for quite some time, until our competitors catch up.”

Foong said the work with Chemin is ongoing as the bank expands Ryt AI’s capabilities, such as the ability to split and share a bill with friends. This will require the AI assistant to handle different conversational nuances and user intents that must be thoroughly understood and tested before deployment.

Furthermore, with plans to add support for Chinese language conversations in the near future, Chemin will be tasked with developing and validating an entirely new set of linguistic benchmarks. “All of this, as you can imagine, will require Chemin’s participation,” Foong said.