Ping Wu, CEO of Cresta, has pinpointed three common misconceptions contact centers have before implementing large language models (LLMs).
LLMs power generative AI (GenAI) use cases in the contact center, enterprise, and beyond.
Most will have utilized an LLM when experimenting with ChatGPT or Gemini. However, this first-hand experience has led to widespread misunderstandings of how contact centers – and the broader business – can leverage this technology.
As noted, Wu - who co-founded Google’s Contact Center AI Solution in 2017 - filtered these down into three central misconceptions.
First, people often think LLMs are just end-to-end text generation machines. You ask a question, and it gives an answer. However, it's more complex.
"There are two parts: comprehension and action (decoding)," explained Wu.
In many business contexts, LLMs are more useful for comprehension, understanding the user’s intent, while the output action is guided by business logic.
Secondly, many customer service leaders consider LLMs as being limited to question-answering.
Yet, as Wu suggests: "They are also very strong in synthesis, extracting key concepts from large amounts of text.” Case summarization is a mature example of this in the contact center.
Lastly, people often apply human intelligence intuition to LLMs, yet there are differences that business leaders must consider. Indeed, some tasks that are hard for humans are easy for LLMs, and vice versa.
Sharing an example, Wu stated: "LLMs can pass advanced placement biology exams but may get simple customer support questions wrong if not properly guided."
After running through these three misconceptions, Wu highlighted how contact centers can better guide LLM outputs, excellent examples of GenAI done well, and more in an interview with CX Today.
The interview is part of our 2024 CX Trends series and is available below.
Yet, for those wishing to skim through the interview, here are some more highlights.
The Two Methods for Implementing LLMs
According to Wu, there are two common approaches to implementing LLMs in the enterprise.
The first involves fine-tuning the LLM with use-case-specific data. For example, GitHub's coding copilot fine-tunes the model using a coding repository to generate better-quality code.
The second approach is retrieval-augmented generation (RAG). Explaining how this works, Wu said:
Relevant business data are first searched and retrieved, then fed into the LLM, which synthesizes the information to answer questions by analyzing multiple documents.
Businesses that embrace the RAG approach must ensure their knowledge centers – which store those documents – include accurate and up-to-date information.
High-profile customer experience AI fails from the New York City government and Air Canada offer tough lessons here.