AI is rapidly becoming part of everyday life, and the telecommunications industry sits at the center of that shift. LG U+, one of Korea's leading carriers, generates more than KRW 15 trillion in annual revenue and serves over 30.71 million wireless subscribers as of 2025. Guided by its "Simply" customer philosophy, the company has made AI central to improving the customer experience, introducing AI-powered search and other services designed to deliver accurate, personalized answers. But as those services scaled, LG U+ ran into a structural problem. The company had become heavily dependent on external large language models (LLMs). With Databricks, LG U+ broke its reliance on any single vendor, bringing down operating costs and speeding up response times in the process. That same foundation is now carrying the company beyond AI search and into Voice AI services built on ixi-O, weaving AI more naturally into customers' everyday lives.
The reality of running LLM-powered services in production
Today, LG U+ prioritizes AI services that improve the customer experience above almost everything else. Its U+one AI Search platform moved past traditional keyword search to understand the questions customers actually ask and return answers that are accurate and personal to them. From there the company pushed personalization further into Voice AI, where ixi-O turns those capabilities into services customers can reach and use in their daily lives.
Keyword search could only go so far. It never quite grasped what customers were really asking, served up generic lists of related links and showed everyone the same results no matter the context. LLM-powered search cleared those hurdles, but it brought a new set of operational ones in its place. Now the team had to manage data quality, fold in business requirements, balance answer quality against response speed and keep the underlying infrastructure reliable.
The deeper, structural problem was how much the service leaned on external LLM providers. With nearly every request riding on an external LLM call, token costs, response speed, output quality and uptime all sat in the hands of outside vendors. LG U+ had built a multi-model architecture with proprietary traffic controls that rerouted requests automatically when something went wrong. But its primary model, Gemini, still ran on a single vendor's infrastructure, so any latency spike or outage on that side landed straight on service quality. The worst latency events still pulled engineers in to intervene by hand, and every switch to an alternative model started the validation process over from scratch.
"We realized we needed a more fundamental fix, something that would structurally free us from depending on any single vendor," said Cho Kyung-hyun, Principal Manager of the ixi Agent Media Business Enablement Team.
Reducing Vendor Dependence with Databricks Model Serving
LG U+ chose Databricks because it met all three of the things the team needed most: reduce vendor dependency, make for a fast migration with minimal code changes and hold performance and reliability at the levels the service already met.
Databricks Model Serving lets teams run multiple foundation models like Gemini, Claude and GPT right alongside one another. What mattered most for LG U+ was that Databricks hosts Gemini 2.5 Flash in a dedicated availability zone through a partnership with Google. The company could keep running the exact same model it already trusted while reaching it through a path that no longer depended on Google's own infrastructure.
The migration itself was fast. LG U+ integrated the platform while keeping its standard LangChain interfaces intact, establishing a reliable alternative serving path insulated from vendor outages. The move opened up far more than LLM serving too, bringing proprietary model serving, LLM tracing and Genie-powered analytics within reach and giving the team a platform it could keep building on.
"By adopting Databricks for LLM serving, we reduced our vendor dependence and actually improved performance at the same time," Cho said.
On top of Databricks, LG U+ rebuilt the U+one AI Search architecture from the ground up for speed and accuracy. Customer traffic now flows through an AI Gateway, and Query Orchestration picks the best search strategy for each request as it comes in. From there a Semantic Elastic Cache speeds up answers to similar questions while keeping costs down, Hybrid Reasoning Retrieval pulls together multiple data sources to surface the most relevant answer and Agentic Self-Correction Loops keep refining accuracy over time. The whole system runs inside an LLMOps framework built for continuous learning and optimization.
Faster Responses With Fewer Failures
In tests run under identical operating conditions, Gemini calls routed through Databricks significantly outperformed direct calls:
48% higher throughput, from 3.2 to 4.8 transactions per second
22% faster average response times, from 18 seconds to 14
90% lower failure rate, from 4% to 0.4%
The gains carried into production, with average response times dropping from roughly 5 seconds to 2, and 89% of responses now arriving in under 10 seconds. By moving to Databricks-hosted Gemini 2.5 Flash, LG U+ lifted performance and operational stability together in a single step.
Taming Complexity While Improving Cost and Performance
The most striking part is that performance and cost efficiency both improved even as the system grew more sophisticated. Measured against the original architecture, system complexity climbed 107%. Yet operating costs fell 36% and response speed improved 64% over the same stretch.
That combination is what makes this more than a simple infrastructure swap. With Databricks absorbing the complexity underneath, LG U+ laid a structural foundation for running sophisticated AI systems efficiently and at scale.
Looking Ahead
LG U+ plans to keep consolidating its distributed AI infrastructure on Databricks. The company also wants to put Databricks Genie in more hands across the business, opening up its data environment so that anyone can analyze key metrics by asking a question in plain language without writing a single line of SQL.
These capabilities will extend beyond U+one AI Search into Voice AI services powered by ixi-O, broadening the company's AI ecosystem.
"Building on the unified AI infrastructure we've put in place with Databricks, we plan to keep building services that are safer, more reliable and more practical for our customers," said Shin Jung-ho, VP and Head of the ixi Agent Dev Lab at LG U+.