How to run a chatbot with ChatGPT

We are seeing more and more companies using chatbots to talk to visitors on their website. The reason is very simple: these bots provide an answer to 69% of user requests without any human action. This can significantly reduce customer service costs.

By integrating a chatbot on your website, it will gain interactivity and credibility. This is essential for visitor engagement. Speaking of engagement, it is possible to push the sliders a little further by appointing ChatGPT as the manager of your chatbot.

It offers a less siloed conversation and allows you to respond in a more specific and natural way to your visitors' queries.

In this article, we will explain how to integrate ChatGPT into your company's online chatbot.

Your online chatbot with ChatGPT

Before we go any further, you should know that you won't find information for a simple chatbot based on decision trees in this article. The goal here is to generate literally different answers for each query.

If your budget allows it, the simplest, fastest and most efficient solution is to use a platform such as sitegpt.ai. You will be able to configure your bot in a few clicks. To answer the requests, the bot is based on all the information present on your site.

However, there is an alternative, less expensive, but a bit more time consuming.

Connect ChatGPT to your chatbot platform

From a purely technical point of view, it is now very easy to connect ChatGPT to a chatbot platform. Most of them now have a possible integration with Open AI.

You just have to create a new API key from your OpenAI account. To do so, click on your icon in the top right corner > Manage Account. Then in the "API Keys" tab, create a new secret key.Copy this key and paste it in the integration parameters provided for this purpose on your chatbot platform.

ChatGPT is connected to your chatbot. The challenge now is to configure your chatbot so that ChatGPT generates relevant responses.

Configuring your GPT chatbot for relevant answers

As with creating a so-called "classic" chatbot, you need to think through your scenario before giving the reins to ChatGPT.

To do this, ask yourself the right questions:

• Why do I create the chatbot?
• What type of request do I want to answer?
• What tone do I want to use to answer the requests?
• What if the user has a specific request?
• What if the user needs more information?
• Under what circumstances is the chatbot triggered?

All the answers to these questions will help you create the prompt for ChatGPT. To help you create this prompt, we invite you to (re)read our article "How to write a good prompt for ChatGPT"

This is all very nice, but it doesn't tell you concretely how to go about it.

Let's say you have an online sneaker store. Your customers might need to reach you for an exchange, a problem with their order or with their delivery. In this case, you should configure your chatbot in the following way (to be adapted according to the platform you use):

• Trigger: Your customer accesses the "my orders" tab
• Welcome message: "Hello [[firstname]], I'm Laura. A problem with your order or its delivery? Need an exchange? Ask me your question, I'll help you answer it."

The triggering of your chatbot is configured. Nevertheless, you still have the touchy part: the responses generated by ChatGPT. This is the one that will require the most work, tweaking and testing on your part.

Add an NLP (Natural Language Processing) trigger

NLP allows the bot to understand the intent of your users through their query. It allows to extract, key words, phrases or questions in order to propose an appropriate response.

Focus on the most frequent requests submitted to your customer service and build a "cloud" of words or groups of words synthesizing this. You can collect all the formulations made by your customers for the same problem and submit them to Free Word Generator.

Then, generate a cloud and download it in csv format. Then, select the words that are necessarily related to the query you want to answer.

Initially, you will have to repeat this operation frequently with new requests from your customers made on the chatbot that will be managed by ChatGPT.

Create a prompt for your GPT chatbot

Now that you have a trigger, a greeting, and something to identify the user's request, you need to ask ChatGPT to generate a response for your chatbot.

Using the sneaker online store example, here's the prompt you could submit:

“You are Laura, a {{website}} customer service agent. When a user asks you a question about their order, you need to provide an empathetic response with a professional tone.

If the user's query contains the words: [coller les mots issus du csv de Free World Generator en lien avec “échange”] then explain that [coller le protocole à suivre en cas d’échange].

If the user's request contains the words: [coller les mots issus du csv de Free World Generator en lien avec “remboursement”] then, explain to him that [coller les modalités de remboursement].

If the user makes a request which is not among the previous cases quoted, you can only base your answer to his request on the following information: “[conditions générales de vente de (website) ]”.

If the user makes a request for which you can't find an answer here, invite them to contact customer service by email: [email protected] or by phone: +33 2 01 02 03 04.”

Your chatbot is ready. All you have to do is test it. On our side, we did the tests directly on ChatGPT. Here is the prompt we submitted to it:

“You are Laura, a customer service agent at Shoozy. When a user asks you a question about their order, you need to provide an empathetic response with a professional tone. To open the conversation with the user, you should send this sentence: "Hello {{firstname}}, I'm Laura. Any problem with your order or its delivery? Need an exchange? Ask me your question, I'll help you answer it.". Then you don't have to introduce yourself in the next answers you provide to the user.

If the user's request contains the words: "exchange", "change", “proceed to an exchange” then explain to him that he must use the return form that was present in the package he received with his order. All he has to do is stick it on his parcel and drop it off at a post office. The sending is free. Once the parcel is received by our warehouses, he will receive an email allowing him to select a new size that he will receive within 5 working days in his mailbox.

If the user's request contains the terms: "refund", "get refund", "get a refund" then explain to him that he must first return his order with the return form contained in the package he received. All he has to do is stick it on his package and drop it off at a post office. The sending is free. Once the package is received by our warehouse, he will receive an email allowing him to request a refund or an exchange. By choosing the refund, he will receive an automatic refund within 2 working days.

If the user makes a request that is not among the cases listed above, you can only answer his request based on the following information: "Shoozy is a company based in Paris. We supply sneakers all over Europe. Our products are guaranteed for 1 year. This guarantee applies only in case of manufacturing defect or faulty workmanship for which we are responsible. It is not applicable in case of misuse by the user" If the user makes a request for which you do not find any answer here, invite him to contact the customer service by email: [email protected] or by phone: +33 2 01 02 03 04. Now that you have the instructions, I want you to open the conversation as I specified.”

And these are the responses generated by ChatGPT:

The prompt seems to work. You can insert it to your chatbot platform. Retest your chatbot to make sure the responses are still relevant. Continue to keep an eye on the queries and responses submitted over the next few days.

At the end of the chat, you can ask your chatbot to submit a user satisfaction survey. This will allow you to measure the performance of your new online chatbot managed by ChatGPT. This will allow you to identify areas for improvement.

Finally, the integration of ChatGPT in the chatbot of your website can really make the difference. You will be able to answer most of the customers' requests without them feeling like they are dealing with a bot that doesn't understand them.You will probably have to train your GPT chatbot in the beginning. But, it will eventually be 100% operational to provide quality customer service to your customers or visitors.

At Abyssale, we are particularly fond of this type of automation that personalizes the customer relationship while freeing up time for other more important tasks.

Feel free to ask your questions in the comments and join our LinkedIn community. You will be informed of the latest updates and news from Abyssale!