Improving the Online Ordering Experience with ChatGPT

As a software engineer, I love experimenting with new technology. It’s like being a kid with a new toy. So when ChatGPT, an artificial intelligence chatbot from OpenAI, was introduced I wanted to see what all the buzz was about. As Vice President of Online Ordering at Paytronix, I also wanted to find out if a chat bot could help our clients. 

The speed that ChatGPT went from techie tool to mainstream is simply amazing. OpenAI officially launched it in November 2022 and anyone can try it out and get familiar with how it works.  

I got my hands on a beta version a little earlier and experimented with it both as a feedback response vehicle and for its language translation abilities. I had tried other AI chat options, but ChatGPT was the only chat bot that offered what Paytronix needed.  

How ChatCPT Can Help 

The beta version of ChatGPT confirmed my suspicions that the chat bot would be a great customer communication facilitator for the Paytronix guest surveys and feedback tool, which is an integral part of Paytronix Online Ordering. When it comes to keeping customers engaged, responding quickly to customer concerns is essential.  

The Paytronix Online Ordering platform automatically engages guests within two hours of their order so that they can respond with the experience fresh in their mind. Guests are asked to rate their experience on a scale of one to five, and the results are immediately sent via email to a store manager who also can see the guest’s lifetime value. Designed to help store managers measure and mediate interactions with guests who place orders online, this process helps store managers identify guests with unhappy online ordering experiences. A quick response is often the difference between keeping a guest returning and losing them for life.  

Previously, store managers would respond to an unhappy guest by using either generic, pre-written emails or customizing one for their experience. Neither situation is ideal, since one isn’t personalized, and the other is time-consuming.  

Here’s where ChatGPT offered its best use-case. By creating queries that take into account a user’s experience, ChatGPT can return a customized email that can be read, edited, and sent out fast.  

As this technology continues to advance, the process of reviewing and responding to guest issues will move further along the automation curve. The responses will get better and the process of sending them will get even faster.  

Solving Integration Puzzle 

While much of the hype around ChatGPT has focused on the public version, Paytronix worked with OpenAI to license a premium version of ChatGPT, the best solution for embedding in commercial products like Online Ordering. Once we could access the ChatGPT application programming interface (API), Paytronix had access to ChatGPT 3.5 model and all its bells and whistles. 

The next step in making this a usable solution was “prompt engineering,” which requires engineers to write prose, not code, to test the AI chatbots. At Paytronix, our prompt engineers wanted to identify the most common responses in our survey tool, then help the chat bot come back with good answers to those questions.  

This is a case where content truly matters. When our engineers built prompts using direct quotes from guests the responses hit the right notes. But it’s not as simple as just sending the email over to ChatGPT, there needs to be some additional work happening to ensure we get the right responses. Even then it’s never going to be perfect, so it’s important to have a human to review and edit the response before hitting the send button.  

As we continue to develop this tool, we can take advantage of a general adversarial network (GAN), a type of machine learning capability, whereby one response can be played against another. The store manager can ask “is this appropriate to send? Yes/no.” Over time, the GAN will get better and better at understanding what is a plausible response and what is not.  

Beyond the email 

In addition to generating responses, ChatGPT can also summarize and categorize guest comments. Let’s assume a restaurant then received a barrage of negative comments in just one night. By adding TLDR (Too Long Didn’t Read) into the prompt, followed by the text ChatGPT can generate a synopsis in a fraction of a second. Now the manager can get a quick summary of what went wrong that night. ok 

On top of all that, ChatGPT offers a much better way to do sentiment analysis. To date, ChatGPT has demonstrated that it can ID data and statistical values when the proper input is provided. Likewise, it is getting better at spotlighting trends with the proper input. 

We’re now experimenting with other uses for ChatGPT. At its most basic it could be used to generate other copy like for menu item descriptions or personalized messages. We are also looking at it for translating information into various languages or even helping catalog things differently. This could help us rethink how we display or share menus.  

But no matter how we use it, there remains a human at the end making sure that what goes out is right and appropriate. AI is great, but we still can’t replace the human mind. 

Learn more about how to leverage artificial intelligence in your online ordering solution.