SK Telecom enhances customer support for millions of Koreans with Claude in Amazon Bedrock

Title

SK Telecom enhances customer support for millions of Koreans with Claude in Amazon Bedrock

Content

SK Telecom, which has held the #1 spot in Korea's National Customer Satisfaction Index for 27 consecutive years, deployed Claude in Amazon Bedrock to power customer support for millions of Koreans — lifting in-call LLM response quality ratings by 34%, cutting low-quality responses in its Telco fine-tuned model by 68% versus the base model, and reaching ~89% of human agent quality in post-call processing. "Brand safety is crucial — anything that tarnishes the brand is just a no-go. Anthropic's huge focus on safety really stood out," said Eric Davis, Head of AI Tech Collaboration Group at SKT. The pressure: Korea is a culture of urgent service — "If you're one minute late, you're ten minutes late," Davis notes — and SKT's reputation rests on answering phone calls within minutes while bridging gaps between older customers who prefer voice and younger ones who gravitate to chat. Traditional call centers ran 9–6 Monday through Friday, leaving weekend and after-hours gaps. "If I call in, I can probably talk to a person within minutes versus having to wait on hold for 30 minutes," Davis explains. Why Claude: after evaluating multiple LLMs, SKT picked Claude on three grounds — brand safety, natural conversational ability ("Claude is more relatable and creative," with the cultural sensitivity needed for Korean dialects where other AI models generated formal, detached responses), and Amazon Bedrock's ability to serve custom models at scale. "We dedicated a lot of effort with Bedrock in terms of getting the performance there. Performance wasn't just about model accuracy — the throughput and latency had to meet our standards," Davis said. Combining SKT's in-house RAG model with Claude for end-to-end retrieval-augmented generation became what Davis calls "a killer combination" — boosting the Telco LLM performance score from 3.3 to over 4.3. Two solutions run in production: In-Call Assist delivers real-time document search and answers to agents, reducing cognitive load during live calls; Post-Call Processing automates summarization, topic classification, to-do list generation, and sentiment analysis with humans kept in the loop. The Telco fine-tuned model itself cut low-quality responses by 68% versus the base. Operational scale: agent evaluation scores rose from 3.3 to mid-4s across the agent base; Claude now covers both in-call and post-call workstreams with humans kept accountable. Davis frames the philosophy as "IA" — intelligence augmentation: "We recognized that our customer service agents are stressed, overwhelmed. We wanted human-in-the-loop technology to make their lives better. This is not meant to replace jobs — it's meant to augment the experience." Future direction: SKT plans a more multilingual, deeper model ("way more multilingual, way deeper model") while preserving the regional call-center network that handles dialect-specific care. The Anthropic partnership continues toward more sophisticated human-centered AI support.