Hybrid LLMs: Turning Insight into Action

The launch of ChatGPT last year sparked a surge of interest in Generative AI, leading some enterprises to explore how this technology can improve their Sales, Marketing, and Customer Support. Large Language Models (LLMs) are expected to be particularly useful for these functions, especially when combined with the enterprise's proprietary data. In fact, such purpose-built LLMs may be the most valuable investments enterprises can make in the coming years: a recent McKinsey study concluded that Generative AI will add up to $4 trillion annually to the world economy, and that 75 percent of this value will be derived from customer operations, marketing and sales, software engineering, and R&D. While the business case for LLMs is strong, most enterprises can't expose their proprietary data to a Public LLM like OpenAI, nor can they afford to build their own fully custom LLM. So, what is the solution? Enter Hybrid LLMs. A Hybrid LLM combines the conversation and reasoning abilities of a Public LLM with the proprietary data and knowledge of the enterprise in an environment that the enterprise alone controls. The enterprise isn't required to share data back with the Public LLM and retains ownership of all accumulated learnings. A Hybrid LLM offers the best of both worlds, and cloud platforms such as Microsoft Azure make Hybrid LLMs possible today. There are a variety of use cases for Hybrid LLMs. A good example is Customer Support, where an LLM-powered service can interact directly with customers and/or aid human support agents. A Hybrid LLM tuned for Customer Support would likely include product & parts information, customer data, knowledge bases, warranty information, and returns policies. Because the service is specifically scoped and tuned for customer support, it can interact with customers in the same way a well-trained agent can. Ultimately the service can deliver support at lower cost and with fewer errors than most human agents, and learnings can be accumulated within the Hybrid LLM, making it smarter over time. There is, however, a catch: to make Hybrid LLMs work, enterprises need their data to be accessible, organized, and of high quality. As this is currently a challenge for most enterprises, data governance needs to be a high priority to ensure a strong foundation for Generative AI. As LLM pioneer Databricks likes to say, "the hardest part is the data part." For those enterprises that step up to the challenge, Hybrid LLMs will be a gamechanger. They will allow enterprises to interact with their customers more effectively than ever before, driving both revenue growth and cost savings. Generative AI technologies as a whole will continue to improve rapidly and come into reach for every enterprise, giving them the opportunity to out-innovate their competition.