Large Language Models (LLMs) and Generative AI (GAI) have monopolized public attention the past few months. And the launch and acquisition of artificial intelligence (AI) programs by a number of large technology companies has only further fueled interest in their capabilities.
Amid promises this technology will permanently impact the ways we work and do business, companies looking to leverage LLMs, especially in the contact center, must properly evaluate the potential effects they will have on their employees, customers, and reputations.
While the potential for this technology has already been proven, proper implementation in the enterprise is heavily dependent upon an organization’s ability to improve their accuracy and regulate their operation. Doing so will ensure that any program they’re incorporated into is able to give repeatable and reliable results.
While that work is underway, today’s LLMs are best suited as an added layer in an organization’s customer experience (CX) strategy. And not as the sole tool that will optimize their service channels.
What are LLMs?
At the most basic level, a LLM is a type of AI that is capable of mimicking human intelligence. To break it down even further:
- The model is large due to the amount of data it is trained on.
- Language because of the kind of information it is trained on, which are data pools of text.
- Model because of how the program is constructed to execute based on statistical analyses.
How Do LLMs Work?
If, for example, an agent representing an insurance company is asked by a customer if they are covered for flood damage, they can use the generative capabilities of a LLM to instantly adapt the tone and length of their response to better suit the situation.
To fully grasp how this technology is able to generate that response and all others, it is important to understand what level of training data they are given and how they are trained.
At the most basic level, a LLM is a type of AI that is capable of mimicking human intelligence.
LLMs are fed information like web pages, books, and articles and are then able to generate new content based on that information. Using a series of statistical models to analyze that vast amount of data, the LLM is able to “learn” and identify patterns and connections between words or phrases.
From there, an LLM takes those trends and patterns to generate new content. This is primarily in the form of text-based outputs, like an essay, an article, or a reply that is similar or approximate in its style to the original based on the user input.
In the example of the flood damage question, the model is likely to have countless articles and web pages within its training data on the subject. When the service agent lodges the request, the LLM quickly scans associated phrases within all of that content and pulls together a response.
What Does This Mean for the Contact Center?
In their current incarnation, LLMs are best suited to improve the performance of existing Conversational AI (CAI) solutions, rather than taking on the front-facing work altogether. Expectations that an LLM alone can take on the role of one or more support specialists aren’t reflective of their current limitations.
It is true that LLMs are able to analyze vast tracts of organizational data to supply approximate responses to queries via text generation. However, they’re unable to account for context, making them more optimal to layer in with tested technologies like natural language understanding (NLU) and CAI.
What this looks like in the contact center is leveraging LLMs and GAI to ease the workloads of human support agents and streamlining the training processes for other external-facing artificially intelligent solutions.
For example, a LLM is a useful tool in scanning and condensing an existing chat with a virtual agent. So that when passed to a human agent they’re able to quickly gather the salient information of the interaction. Or, prior to the deployment of a virtual agent.
Common Challenges and the Correct Response
One of the largest challenges in using LLMs is the ability to validate the content it creates is both accurate and reliable because it comes from various sources. Those sources could have potentially different motivations that may also open the technology - and its output - to bias.
To counteract bias, it is important to continue supervised training of applications, and institute specific standards for production and testing that are evaluated prior to technical deployment. The model’s training process involves a stage known as “fine-tuning” which is supervised and can help to reduce the propagation of harmful or biased content.
LLMs represent a dramatic opportunity to improve the user experience, however their use alone...is not enough.
While LLMs are able to satisfy user queries, they also present the risk of generating outputs that are inaccurate and misleading. These are commonly known as “hallucinations” and can be presented in such a confident manner by the LLM that, thanks to its ability to write well, they could lead to the generation of false information if not verified.
LLMs are unable to self-correct, which can be solved by incorporating tested uses of AI, like CAI, which are more reliable and can link a LLM to an accurate data pool. Other measures to take are incorporating a human-in-the-loop approach, especially when examining LLM use in an enterprise contact center.
Ensuring that someone verifies the content generated by a LLM prior to sharing with an end-user is of paramount importance.
Best Practices for Deployment
Current iterations of LLMs, though improving, are not accurate enough for large-scale business deployments. They show great promise, but if handling sensitive or customer-specific data, their use should be limited or verified by a human-in-the-loop approach prior to any LLM-generated output being shared externally.
Relegating LLMs for exploratory or development environment testing is the best starting point for organizations looking to employ their use. Avoiding placing them in environments that require high levels of sensitivity, like a task that could have legal implications for a business, is also wise.
LLMs Versus Other AI Models
There are countless AI models in research, in product development, and on the market. The most pertinent ones to the contact center include the following:
Narrow AI
- Narrow AI is a category of artificial intelligence encompassing AI that is designed to complete a specific task without human support. Unlike LLMs and GAI, Narrow AI produces replicable results. AI-powered chatbots are a form of Narrow AI.
Conversational AI
- CAI encompasses a wider breadth of technologies and technical approaches.
- Consider CAI to be the driving force behind something like a chatbot, which responds to user requests in a conversational manner, leveraging natural language processing, NLU, and machine learning (ML) models.
Chatbots
- Unlike LLMs, chatbots are trained on a limited subset of data and incorporate a range of intent data so they can provide repeatable, accurate results.
- When providing a response, a LLM is approximating based on a wealth of data, but it is not optimized for a specific function.
Deep Learning and Machine Learning
- Machine Learning (ML) is a subfield of AI, focused on equipping computer systems with the ability to learn patterns and make data-driven decisions through algorithms and statistical models.
- Deep Learning (DL) is an advanced form of ML that utilizes artificial neural networks to mimic the human brain’s learning process.
- When combined, ML and DL provide powerful tools for solving complex real-world problems, transforming the way business and organizations optimize their operations.
LLM Versus NLU
- An LLM is trained on a large pool of general knowledge and is incredibly capable at generating outputs that mirror accuracy. However, unlike NLU, LLMs cannot understand context and rely on hallucination to produce responses. NLU, for its part, is capable of parsing contextual information and helps something like a chatbot provide precise answers and does not make up answers.
What matters most to contact centers in the coming months and years is a measured approach to the implementation of GAI and LLMs. Research is still ongoing, and their disruption to more traditional enterprise operating systems has yet to be fully realized.
The key achievement of contact centers is providing timely customer service in a manner that feels both personal and encompassing of the users’ needs or issues. LLMs represent a dramatic opportunity to improve the user experience, however their use alone at this stage of their functionality is not enough.
Customer service departments should look to incorporate LLMs incrementally, primarily to speed up training of other external uses of AI like chatbots or website generation, to account for existing liabilities.