JON

03.21.2023

Share this article

JON

03.21.2023

ChatGPT and GPT-4 use cases in Customer Support

In customer support, access to data and knowledge has become a differentiator in the market. Data is vital to providing high-quality, fast, delightful customer experiences that are expected today. With the rise of cloud-based solutions, It has become more common for support to have multiple tools they need to access to find the data they need to solve customer issues.

Integration of customer support tool stacks has become a typical use case for support agents to find information, troubleshoot tickets, and gather context about the customer. Requiring an agent to switch between multiple tools can be hard to teach, time-consuming to action, and have human copy and paste errors. High expectations from customers have increased pressure on companies to fix defects faster, reduce the friction in using products and getting support, and prevent issues from happening altogether. Combined this makes data access even more important to a support organization.

The advent of stronger artificial intelligence capabilities has made all of these expectations easier to meet. Over the past 10 years we have seen an increase in the capabilities of artificial intelligence and computational power. Machine Learning (ML) models can identify images, make decisions, and find patterns in data. Natural Language Processors (NLP) applies ML to text. Specifically attempting to derive knowledge and intent from spoken and written text.

Common applications of NLP include sentiment analysis, language translation, speech recognition, and text summarization. As these technologies have improved, OpenAI began developing GPT-3 in 2020. GPT-3 is a Large Language Model (LLM) which uses techniques from NLP to generate human-sounding text. This led to the launch of ChatGPT in late-2022.

These technologies will revolutionize how companies provide support services and will fundamentally change the traditional function of support within an organization.

What is ChatGPT?

It’s important to understand some terminology when discussing AI technologies. AI is often purported to be the end of jobs or take over society entirely. Its capabilities are exaggerated or are suggested to be able to solve any problem at all. Most experts say not likely anytime soon. What AI can do today is powerful, but in the context of this article we will be focusing on three key technologies: Large Language Models, GPT-3.5 and GPT-4.

Large Language Models

Large Language Models are a type of NLP model which is designed to understand and generate humanlike responses to inputs. They are trained using deep learning techniques known as neural networks to identify patterns within our language. LLMs are given vast amounts of data from relevant sources so they can learn about one or more topics. In doing so, they are able to generate text that appears as knowledge about those topics.

There are many LLMs in development. Google is working on LaMDA and PaLM and Meta has OPT-175B and GPT-4 was developed by OpenAI.

GPT-3.5

GPT-3.5 stands for Generative Pre-trained Transformer 3.5. It was trained using a neural network model with over 175 billion parameters coming from the internet. It represents a breakthrough in the development of LLMs and has gained a lot of attention from the public as well as the research community. GPT-3.5 has impressive performance with text completion, question-answering, and text generation. It is this technology that ChatGPT has been built on.

OpenAI released ChatGPT to the public and immediately the implications and use cases became clear. This technology could revolutionize numerous industries. One area in particular is in Customer Support.

GPT-4

In mid-March a new version of GPT was released by Open AI. Here are a few on the New Features that will enhance its capabilities and provide even more value to Support organizations:

Overall better accuracy, in testing GPT4 is 80% more accurate
Cuts reply time in half
Has the ability to analyze image inputs and answer questions about them. In support for example, this might mean automatically analyzing customers’ attachments
Add steerability, meaning better boundaries around specific topics, ability to specify the tone and style. An example support use case would be to only answer questions about a specific company and its products
Higher response word limit – size of GPT3.5 is currently constrained by around 3000 words. GPT4 doubles this limit, and you can even configure larger limits
Includes support for more languages.

Why is ChatGPT a Big Deal for Customer Support?

Support has the opportunity to fundamentally change the customer experience, the products, and services a company offers. LLM models have made it possible for small to medium-sized companies to create ML models of the same quality as big companies without the need for expensive data and engineering resources. Basically, LLMs are democratizing ML capabilities.

Customer Support organizations now have a low barrier to entry to understanding the vast amount of data they collect and auto-generate meaningful, non-robotic responses faster. The combination of these improvements means support can spend more time on the human side of support and less time on searching or interacting with tools. According to Zendesk’s 2023 CX Trends study, 59% of customers want data to be used to personalize support experiences. This expectation is easier than ever to meet.

This trend is starting to get management attention. Kustomer found that 84% of CX leaders think personalization will be most important in the next 3 years. ChatGPT can help bring more context to your replies faster. It’s not a matter of if you should look at adopting these technologies, it is how many use cases do you have where they can help.

Generative AI Use Cases for Support

Generative AI refers to a type of LLM that has been trained to generate new content, such as text, images, or music, based on patterns it has learned from a dataset. Whatever use cases you have for data, a generative AI model can likely assist in mimicking creation of it.

Chat Conversations

One of the most basic use cases is a technology that we are familiar with. Chat conversations, Chatbots or automated IVRs have been around since the early 2000’s and relied on keyword matching and, in some cases, required very specific phrases to be used. While these may not have been classes as AI’s, they were first iterations at automation, which can now be improved with ChatGPT.

ChatGPT adds value in two key ways:

It no longer relies on keyword matching. Because it can be trained on large datasets, it will be able to understand natural language intents and be able to provide service using a more natural way of communicating.
It has the ability to use the data to gather more context and relevant data points and adjust its responses on to be more relevant to the customer

Dynamic Macros

Macros or text expansions have long been a method for support agents to respond quickly to customers on repeatable tickets. They often look like this:

While a time saver, there is still manual effort to find relevant details, and “fill in the blanks.” As well, the written copy doesn’t change. Meaning as customers are with you longer and have similar issues, the text sounds robotic and lacks a personal touch.

GPT-4 removes all of these negatives. Not only can it automatically grab the relevant information to be contextually accurate without human assistance, it can generate unique and natural sounding responses every time, so it never gets stale.

Identify Relevant SOPs

Onboarding support agents is hard. Depending on your business, you might have dozens of different workflows or standards of practice (SOP) that you use throughout the day. You might have tools that you need to teach, products that you need to train for, the process to do all the actions in those tools and products. This training can take days or weeks, or even months to become fully effective.

GPT-4 can remove a lot of this time and effort. By giving it data that compares your ticket text and language to the SOPs and workflows that your agent should do, you can train your agents in real time when those actions are needed. The AI not only can suggest the relevant workflow the agent should take, but with the right tooling can automate all the clicks necessary to make that action happen.

For example if you are delivering a product and in order to refund it you need to look up the order number in one system, and then go to the payment system to process a refund, ChatGPT will:

Read the ticket
Recognize a refund is necessary
Recommend the workflow to complete that refund to the agent
Provide the agent a method to run all the actions necessary in one step

This saves time and effort on the individual ticket as well as on the overall training of your support organization

Knowledge Articles

Knowledge Management concepts have been around in some form since support organizations started. In the early 1990s the Consortium for Service Innovation launched the knowledge centered support model (KCS) , which became a commonly used model for creating, revising, and updating knowledge content. But, like most models, it relied on humans who knew what they were talking about to write articles in a way they thought the customer could read, understand, and execute on. That is a lot harder than it sounds. Customers bring different levels of capabilities, perspectives, and language to your content. Writing articles that suit all of those differences is difficult.

With AI however, recognizing a new ticket or a solution that has not been given before becomes easier and given it has the context of your products and services it will have the capability to generate articles for internal use or one more dynamic content given it understands your customer context and your company’s voice and style . AI will fundamentally change the way knowledge is created and shared.

Summarization

The complexity of many support tasks increases when a voice channel is introduced. For example, Ticket Quality Assurance and Ticket escalations or handoffs historically have lost context or required either manually summarizing phone calls or other stakeholders listening to the entire call.

LLMs simplify this complex and time-intensive task. GPT-4’s NLP models can “listen” to your calls to transcribe and summarize them more effectively than previous technologies. These consistent and accurate summaries allow agents to lookup information more quickly, search previous tickets that were previously unfindable, and allows QA to understand the ticket better and score more tickets,

Maximizing the Value

Almost all Customer Service teams have one or more of the above use cases which can be improved with GPT-4-enabled technologies. However, it is critical to understand that maximizing the value from these tools requires some effort. When starting an AI journey at your company, ensure that the tools you review have a focus on deep training of its models and utilize, or make configurable, the settings that map to your needs.

Deeper Training

To maximize the value of your GPT-4 implementation, you should work closely with your tool vendor or partner or hire specific ML talent. Training a machine learning model like GPT-4 can be a complex process, and it requires knowledge of various techniques and methods to

achieve the desired results. Some of the available deeper training requires examples of the following:

Personalization involves customizing the model to a particular user or group of users to improve its accuracy and effectiveness.
Product lexicon is a specialized vocabulary or set of terms that are unique to your

business or industry

Multimodal input refers to the use of multiple forms of input, such as text, images, and audio, to train the model.
Languages refer to the different languages that the model can understand and generate.
Embeddings are a technique that involves representing words and concepts as high-dimensional vectors to improve the model’s performance.

To implement these deeper training techniques, you would provide existing content to GPT-4, including both positive and negative examples of each. This content would help the model learn and understand the patterns and nuances of your business or industry and improve its accuracy and effectiveness.

ChatGPT Settings

ChatGPT’s APIs allow your company or your vendor to tweak key communication settings that will help you deliver the experience your customers expect.Here are a few of the

Formality – You can train ChatGPT to use more formal or informal language based on your company’s brand identity and the user’s preferences. For example, a financial institution may want to use more formal language, while a fashion brand may want to use more informal language.
Empathy – While AI technologies are not known for their empathy, you can adjust the empathetic language and how ChatGPT will understand the user’s emotions. This can help create a more supportive and compassionate customer experience.
Politeness – In most Customer Services cases you want ChatGPT to use polite language and avoid using offensive or insensitive language.
Clarity – In general, you likely want ChatGPT to use clear and concise language, avoiding technical jargon or overly complicated language. Clarity can help to ensure that users understand the support provided and avoid any confusion.
Brand Voice – Having AI use a consistent brand voice is a major advancement in the field of LLMs. Maintaining your company tone and language will build a more cohesive service offering.

Evaluating the Right Tool

As Illustrated above, there are a lot of considerations when it comes to building an artificial intelligence strategy, or evaluating tooling solutions for your environment. Chat PT is the culmination of decades of research into AI, an advancement in ML,NLP, and LLM technologies, and a game changer for the customer service industry. For the first time, bots and automations won’t sound robotic, can recognize and respond to natural language, and they can be tuned and trained to meet your high service delivery standards.

This new wave of tools that ChatGPT will allow brings these technologies to small and medium-sized businesses in a way that hasn’t been possible or affordable until now. These tools will usher in a new way of serving customers that is more enjoyable for them and provides efficiencies to your business. Finding tools that solve one or more of the use cases above for your business needs to be a top priority for leaders in customer experience.