How advances in ChatGPT are impacting CX and chatbot design

Let’s be straight from the start – ChatGPT is not going to solve your customer experience challenges on its own and overnight. It isn’t going to replace your customer service agents; you won’t be able to retire your website and make ChatGPT the primary way for customers to discover your products; and it won’t replace the chatbots that you have carefully designed to help your customers with everything from customer service queries to account registrations.

What ChatGPT will do is give you another useful tool that will help you provide the kind of customer experiences that people now expect, and some that may even surprise them.

In our previous blog we filtered out the hype and talked about what you could actually do with ChatGPT that would add value to your business and your customer’s lives. A lot of the ideas were theoretical as the well documented issues with the inaccuracy of the GPT-3 model and the limited availability of the service meant that it wasn’t a viable option for most production use cases.

As Ferris Bueller famously said, “Life moves pretty fast. If you don’t stop and look around once in a while you could miss it.”

He probably wasn’t talking about large language models, but it certainly applies.

In a few short weeks we have gone from talking about potential ChatGPT use cases theoretically, to using the latest version to create our own integration. Released today, we hope it will save people significant time and effort and give them a taste of how they could use ChatGPT themselves to enhance the lives of their customers.

More on that later, but first let’s catch up on some of the recent developments in the world of large language models, answer some pressing questions, and find out what has changed to enable us to create a real example of a ChatGPT integration.

The latest version of ChatGPT is 4, which was released in mid-March 2023.

Unlike ChatGPT-3, version 4 requires a subscription to ChatGPT+. You can sign up at https://chat.openai.com/ or if you have a Microsoft account, you can use the latest version of Bing which is integrated with GPT-4 and includes most but not all its features. The interface is arguably better designed for search-type queries, although it doesn’t accept images like the native version does.

If you already have a GPT-3 account, it won’t automatically be updated to version 4. Version 3 still works, but it now fesses up that it is still stuck in 2021, and you will be missing out on a host of improvements and new features.

Version 4.0 has some neat new features and improves on the previous version in just about every way. It really is amazing how far the model has matured in just four months since the public launch of version 3 back in November.

It’s almost as if they had both versions already lined up and staggered their releases as a clever PR tactic to extend the hype phase.

A few of the main improvements include:

So, on all these measures version 4 is allegedly better than 3 – quite significantly in some areas. For example, it can now accept up to 25,000 words as input where previously the limit was 3,000. People are testing it by submitting actual exam questions and it is acing them, where in most cases GPT 3.0 was simply passing.

But does it still have a tendency to talk nonsense?

Like its predecessor, GPT 4.0 is still just a model. It scans existing content and then uses it to provide answers, the accuracy of which will be dependent on the reliability of the information that it found on the subject. So, if you ask it to calculate Pi to a hundred decimal places then it will provide a very accurate answer – but ask it to provide proof of alien abductions then you enter the realm of conjecture and fantasy.

That said, OpenAI is claiming that version 4.0 is far better at identifying false claims and is less prone to factual hallucination.

Crucially, it’s also far more up-to-date and even knows last night’s match results.

We checked and the results are correct – even though the poetry is a bit suspect.

No – you now need a ChatGPT+ subscription which costs $20 a month. However, the latest version of Microsoft’s Bing search is based on ChatGPT 4.0 so you can try it out for free, although it is missing some features.

We don’t think so. Most businesses use chatbots in very specific ways and to perform functions that they have been designed and trained to excel at. Whether that is to help people sign up for a service, register a SIM card, or submit blood pressure readings to their medical practitioner. You don’t need an all-powerful large language model to achieve these specific goals, and in reality, it would be a distraction.

What ChatGPT will do is to help businesses to improve their chatbots and therefore improve the customer experience indirectly. Here are just a few examples.

Speed up the build process: ChatGPT can be used to supply training phrases and intents, which is a time and labor-intensive stage of the chatbot creation process. Organizations will therefore be able to create more and better quality chatbots that cover an increasing list of use cases.

Apply brand voice: When educated on a business’s brand guidelines, ChatGPT will be able to check that chatbot outputs are aligned with brand voice, tone, and typography. It can also be used to quickly update all of these for new markets.

Pattern identification: AI is very good at identifying patterns that humans may not notice but can also help explain the contributing factors and suggest how these can be manipulated for the benefit of the business.

For example, imagine a helpdesk for a global technology brand logging issues with the performance of its products. By cross-referencing the location of each caller and the type of problem they are reporting, AI could show that the service slows down during peak internet usage times in that time zone i.e. 7pm to 11pm. Chatbots can then be trained to include this information in their troubleshooting checklist.

Testing: The testing phase is crucial for the success of a chatbot. As well as providing training phrases, ChatGBT is very good at sanity testing large numbers of conversation flows to ensure that they all make sense, don’t end abruptly, and are sufficient to achieve the business goal of the chatbot.

You wouldn’t let a gorilla babysit your children, and in the same way you wouldn’t let your customers interact freely with ChatGPT in its wild and untamed state. As powerful as it is, it can’t be trusted to not give nonsense answers or provide negative information. Scraping the entire internet for content, it may pick up that single negative review you had from someone three years ago who was just having a bad day!

But in the right context and given the correct training, the King Kong of the AI world can be a fantastic tool for helping people discover your products and services in an intelligent and more conversational way.

At Infobip our mantra is that we learn by doing – so we went ahead and built an integration that we think will provide an exciting insight into how the power of ChatGPT can be harnessed safely and effectively.

Infobip has a rich library of product documentation that we publish on our website for all to access. Our global team of technical writers work hard to ensure that users of products have all the information and inspiration they need to get the most out of them.

From detailed product help for all our individual solutions, to step-by-step tutorials, and a detailed API reference library for developers, this adds up to a significant volume of information.

The documentation portal is designed to be intuitive and easy to navigate, but customers can now find the information they are looking for by interacting with our web chatbot or WhatsApp chatbot both powered by ChatGPT.

Why don’t you try it out for yourself? Just click the Omnia chat bubble in the bottom right of the screen, or if you prefer to use WhatsApp then scan the following QR code with your smart phone to start a chat.

Here are some questions that you could ask.

We have restricted the information that the AI has access to, so at this stage it will only answer questions about Infobip’s products and services. It will politely decline to answer questions that aren’t in its remit, so there is no risk of the gorilla escaping and smashing up your carefully curated brand image.

The final question is one that both content writers and nervous best man speech writers are keen to get answered.

No. Still a long way to go I’m afraid.