With more than 11,000 stores across Thailand serving millions of customers, CP All, the country’s sole licensed operator of 7-Eleven convenience stores, recently turned to AI to dial up its call centers’ service capabilities.
Built on the NVIDIA conversational AI platform, the Bangkok-based company’s customer service bots help call-center agents answer frequently asked questions and track customer orders. The bots understand and speak Thai with 97 percent accuracy, according to Areoll Wu, deputy general manager of CP All.
This kind of innovation is a key value for CP All, which partners with several industry groups and national agencies on an annual awards program to encourage research and entrepreneurship in Thailand.
CP All’s system uses NVIDIA DGX systems and the NVIDIA NeMo framework for natural language processing training, and the NVIDIA Triton Inference Server for AI inference and model deployment.
Keeping Up With the Calls
CP All’s 7-Eleven call centers manage customer inquiries in many business domains — including e-commerce, finance and retail — which each have area-specific expert representatives. The centers typically get nearly 250,000 calls a day, according to Kritima Klomnoi, project manager at Gosoft, a subsidiary of CP All.
“Reducing hold time for customers is a key measure of our service performance,” Klomnoi said. “NVIDIA technologies offer us a 60 percent reduction in the call load that human agents must handle, allowing employees to efficiently tackle more unique and complex problems raised by customers.”
Using AI-driven automatic speech recognition services, CP All’s customer phone calls are transcribed in real time. When a customer service bot recognizes a question based on the NVIDIA-powered intelligent FAQ system, it immediately provides an answer using text-to-speech technologies.
Otherwise, the AI quickly analyzes and routes calls to the appropriate employee who can assist in resolving the query in its specific business domain. CP All has also automated all e-commerce order-tracking inquiries using AI.
Adapting to the Thai Language
When first exploring conversational AI, the CP All team faced the challenge of getting the model to recognize the nuances of the Thai, Wu said.
Standard Thai uses 21 consonants, 18 pure vowel sounds, three diphthongs and five tones — making it a complex language. NVIDIA NeMo — a framework for building, training and fine-tuning GPU-accelerated speech and natural language understanding models — helped CP All work through the intricacies.
“The toolkit’s pretrained models and tools made the process of deploying our service much less daunting,” said Wu. “With the help of NeMo, we were able to quickly build and improve our AI language models, which are now optimized to understand and speak the unique Thai language.”
According to Wu, the NeMo framework enabled a 97 percent accuracy in CP All’s Thai language models, more than tenfold the accuracy achieved previously.
Looking forward, CP All plans to expand its AI services to more business domains and scale to millions of concurrent sessions on NVIDIA GPU inference architecture.
Learn more at CP All’s panel at GTC, running March 21-24.