Gauge Your Position on the GenAI Continuum

As a contact center professional, it’s likely your company is somewhere in the process of evaluating or implementing Generative AI capabilities. Large Language Models (LLMs) fall under the broader umbrella of Generative AI. This classification emphasizes an LLM’s capability to generate text, but it’s crucial to highlight that LLMs have two important functions.

An LLMs dual strengths are:

  • Natural Language Understanding (NLU): An LLM can interpret any arbitrary sequence of words, deriving meaning through the extensive world model it developed during training.
  • Natural Language Generation (NLG): With an initial string of words, an LLM skillfully predicts subsequent words, allowing it to craft coherent sentences and entire documents that remain relevant to the specified topic.

Although the emphasis on LLMs often centers on their text generation capabilities, their true superpower is rooted in “inference,” their ability to comprehend virtually any human language input without further training, even across multiple languages. This is particularly valuable in the context of contact centers.

To highlight just a few applications, LLMs can:

  • Discern the queries of customers during calls or texts.
  • Grasp the main points of recorded or transcribed calls, aiding in summarization.
  • Evaluate how closely a call center agent follows a prescribed script.

Assessing Your Current GenAI Positioning

Embarking on transforming your contact center with Generative AI may seem daunting, but starting is straightforward. Importantly, there’s no need for an abrupt overhaul of your existing NLP technologies. Instead, focus on comprehending the extensive benefits of LLMs and strategize on how Generative AI can enhance both customer and employee experiences.

To help you begin or stay inspired, we’ve developed a maturity scale. This scale is focused exclusively on your journey toward adopting Generative AI (LLMs) to power your contact center’s language understanding. (It does not address an content generation tasks or use cases).

  • Level 0: Unaware of the technology powering NLP systems currently deployed.
  • Level 1: Have a clear understanding of the existing NLP technology in use.
  • Level 2: Know the strengths and limitations of the current NLP system.
  • Level 3: Planning underway to integrate portions of the system with Generative AI.
  • Level 4: Pilot projects involving Generative AI for NLU are in progress.
  • Level 5: Substantial portions of NLU tasks are relegated to LLMs.

Hopefully this scale can serve to encourage your progress and help prioritize next steps in adopting Generative AI technologies for you language understanding needs.

Understanding Your Current (Legacy) Systems

Most companies are still using older NLP technologies that their teams have painstakingly handcrafted over years, sometimes decades. Dialog modules built with Nuance Mix for speech enabled IVRs are prime examples. More recently, Google Dialogflow and Amazon Lex were integrated into “voicebot” tooling as good examples of “pre-LLM” era Conversational AI systems.

These older systems, and the personnel that continue to care for them, continue to perform adequately. At the same time, LLMs have emerged as the new foundation for informing voicebots, chatbots and agent assistants. The path from present solutions to LLMs is not straightforward. The most experienced employees find the transition intimidating, possibly expensive, and risky, given its reputation of providing incorrect answers. As a result, companies are understandably hesitant about making changes.

Leveraging LLMs for language understanding offers numerous advantages over the older technologies. The journey towards GenAI adoption should therefore begin with understanding your current NLP system(s) and where they might be falling short.

Here’s a brief overview of the characteristics of the labor-intensive approach that characterizes pre-LLM era language understanding systems:

  • Intent-Based Design: Requires defining specific intents and training the system with sample utterances for each intent.
  • Slot Filling: Necessary for extracting specific pieces of information (like pizza toppings) from the user’s input.
  • Flow Design: Developers need to meticulously design conversation flows, mapping out each potential path a conversation could take.
  • Maintenance: Keeping the system up-to-date requires continuously adding new intents, sample utterances, and managing complex decision trees for conversation flows.

The key distinction between traditional NLP systems and LLMs is that traditional systems are not pre-trained to understand human language. Instead, they rely on manually programmed predictions of customer speech to recognize specific words and associate them with appropriate responses or workflows.

These systems, while powerful and effective in structured domains, often require significant upfront development and ongoing maintenance efforts. They are rigid, and extending their capabilities or updating their knowledge base has proven to be labor-intensive. They also still dominate the landscape. It may be years before they are replaced by LLMs and, when treated like many other information technologies and systems the prevailing rule is “if it ain’t broke, don’t fix it.” For the most important use cases, systems and processes are already in place to support continuous optimization. That’s why legacy systems are destined to stay in service for the foreseeable future.

Contrasting Approaches to a Popular Use Case

To whet your appetite for adventure, let’s look at the advantages of LLMs over traditional NLP for a well-understood use case, ordering a pizza.

Building a Pizza Ordering Chatbot with traditional NLP tools

  1. Define Intents: You first need to categorize user requests into specific intents (e.g., order pizza, change order, inquire about menu). For each intent, you manually define expected user phrases or questions that would trigger these intents.
  2. Sample Utterances: For every intent, you provide sample utterances that help the system understand the variety of ways users might express their intent.
  3. Specify Entities (Slots): Identify and define the necessary entities (slots) that need to be extracted from user inputs, such as pizza size, crust type, and toppings. These are predefined and must be anticipated by the developer.
  4. Design Conversational Flows: Map out how the chatbot should navigate the conversation based on user inputs, intents recognized, and entities extracted. This often involves creating decision trees or flowcharts.
  5. Training: Train the model with the provided intents and sample utterances so it can accurately recognize user inputs.
  6. Testing and Iteration: Test the chatbot extensively and iterate on the design by adding more sample utterances or refining intents and entities based on where users experience confusion or where the chatbot misinterprets inputs.

Now let’s examine the same process using Generative AI / LLMs.

Building a Pizza Ordering Chatbot with an LLM

  1. Crafting prompts: Instead of defining intents and entities, you craft prompts or instructions that guide the LLM in understanding the task (e.g., “You are a chatbot that helps users order pizza from our menu. Assist the user in choosing their toppings, size, and provide a total price.”).
  2. Contextual Understanding: LLMs can infer intent and extract relevant information (like pizza toppings or size) directly from the natural language input without predefined entities. This is due to their vast pre-training on diverse language data, which gives them a broad understanding of context and intent.
  3. Dynamic Conversations: The conversation flow with an LLM does not need to be explicitly designed. The model dynamically generates responses based on the input provided, maintaining context over the course of the interaction.
  4. Feedback Loop: Instead of traditional training, you refine the LLM’s responses through prompt tuning or by providing it feedback on its outputs, teaching it the specifics of your service or product as needed.
  5. Continuous Learning: Some LLMs can be fine-tuned with specific examples from your domain to improve their performance over time, although this is more about refining than the continuous learning of new content post-deployment.

Advantages of LLMs in NLU for Contact Centers

As you probably noticed, building a pizza-ordering chatbot with an LLM combines both language understanding as well as language generation tasks. Since the LLM is equipped to understand whatever the customer might say, the pizza ordering conversation will take natural twists and turns that require the chatbot to generate text on the fly. But let’s focus on just the language understanding part of the equation.

The advantages of using LLMs for NLU in contact centers are numerous. Firstly, LLMs provide a more flexible interaction model because they can understand a broad spectrum of natural language inputs without the need for explicitly mapping intents and entities. This flexibility reduces the setup time significantly, as there is no longer a requirement to manually define and train the system on specific intents and sample utterances. Additionally, LLMs exhibit a high degree of adaptability, allowing them to seamlessly handle unexpected user queries or shifts in conversational context, thus enhancing the naturalness of the conversational experience.

Start by Leveraging Generative AI Offerings Provided by Your Vendor

Nearly all CCaaS (Contact Center as a Service) providers now incorporate Generative AI solutions into their platforms, with call summarization being a common offering due to its straightforward nature and low risk. This feature showcases the dual capabilities of LLMs—Natural Language Understanding (NLU) and Natural Language Generation (NLG).

While the generated summaries are impressive in themselves, it’s crucial to recognize the underlying process: the LLM must thoroughly understand every detail of the conversation to produce an accurate summary. In contrast, the pre-LLM NLP systems we discussed previously would struggle with this level of detailed language comprehension, yet for an LLM, this task is relatively simple.

Many providers now offer self-service features that utilize LLMs to power customer-facing virtual agents. It’s wise to collaborate with your provider to identify opportunities where you might enhance or replace your existing NLP systems with these newer Generative AI capabilities. Chances are, there are many ways to begin harnessing the sophisticated language understanding abilities of LLMs, while managing the risks.

Categories: Intelligent Assistants