Blog

Designing ChatGPT-Enabled Customer Service Chatbots

Chances are, you’ve heard of ChatGPT from Open AI. Trained on the largest text dataset to date, ChatGPT represents a leap in artificial intelligence technology that offers natural, conversational ways of interaction.

Users can give ChatGPT natural language prompts, to which it will respond with human-like responses. Unlike traditional customer chatbots designed with conversation flows, keyword triggers, pre-defined options, and planned responses, you could train ChatGPT with your brand’s customer support knowledgebase and it will know how to answer any inquiry you throw at it, provided the information is available from the provided data.

This means that ChatGPT could help your brand improve your customer satisfaction rate while keeping your labor costs down. By integrating ChatGPT with your brand’s customer service chatbot, you can automate your most common inquiries in a conversational way and free your support team to focus on higher value tasks.

In this article, we’ll discuss some considerations for designing delightful customer service experiences for your brand with the aid of ChatGPT. We’ll focus on UX and design aspects and leave technical implementation specifics to other authors and future articles.

Interaction Model: Pure ChatGPT vs. Hybrid

Based on the brands I’ve worked with in the past, chatbots have been typically designed using conversation trees in chatbot platforms like Chatfuel. In this model, users are presented with pre-defined options in the form of buttons and carousels. 

Example of Traditional Chatbot Conversation Flow

Upon choosing one of these options, the user is presented with more options or taken to a human operator based on the flowchart logic. 

Example of predefined “quick response” choices in traditional, rule-based chatbots

Now that we have the technology to parse any potential user input and provide human-like responses, should we chuck away this entire model out the window?

Not necessarily. A hybrid approach provides a safety net in the form of predefined choices and responses that have been vetted in advance for accuracy while allowing ChatGPT to handle all other unknown inputs.

Using a pure, freeform ChatGPT approach removes the need for designing conversation flows and keyword triggers, though it doesn’t eliminate the need to do fine tuning and testing to improve the accuracy of responses.

I expect most projects will fall somewhere in between, combining some predefined options with the conversational power of ChatGPT.

Tone of Voice

Word choice, sentence structure, and formatting style all form part of your brand voice and play an important role in how customers perceive and relate to your brand. 

In traditional chatbots, conveying the brand voice is a matter of getting your UX writers to adopt your brand personality as they write the chatbot’s predefined responses based on your conversation flow.

With ChatGPT, the tone of voice will largely be determined by its base dataset, i.e. the entire publicly available texts from the internet, any custom texts (FAQs, KB Articles, Transcripts, Support Documents) that you teach your GPT-3 model through OpenAI’s fine-tuning process, plus any tweaks in the prompt.

What that means is that you need to prepare a set of texts, prompts, and expected responses that reflect your brand voice. Your UX writers will be involved in this process, but it’s a very different way of approaching chatbot design. In some sense, it’s more like training a dog compared to how you would normally write.

Persona and Identity

As powerful as ChatGPT is, it’s only as good as the text data it’s trained on. Garbage in, garbage out. It could struggle with complex questions since it only gives the semblance of cognition, not actual reasoning.

And for that reason, it could be dangerous for your brand. Because ChatGPT’s responses are so human-like, it can be very frustrating for your users when it falls short.

That’s why it’s important to communicate to your users that they are talking to a bot and to offer a way to talk to a human operator. You can introduce the bot as your brand or give it its own name. Think Google Assistant vs Siri. 

Whatever choice you make, the important thing is to set clear expectations that the user is talking to a bot and explain what the bot is able to do.

Localization

Since GPT-3 is trained on a large dataset of text from the internet, it’s able to understand and generate text in a wide variety of languages, including my native language Tagalog (Filipino).

However, you still need to fine tune the model so that it answers in the language you expect. In the example below, I ask it initially in Tagalog but it answers in English by default. While that’s fine for me, it might not be desirable depending on your target user persona.

Example of Localization

Only when I specifically instruct it to answer in Taglish—Tagalog and English code switching that’s popular in the Philippine capital—does it respond in the expected target language. Even then it’s still answering in a stilted, overly formal Tagalog. 

Escalation and Handoff

When ChatGPT fails to answer customer inquiries—and it will—you’ll need a human to take over. Your main considerations are when and how to perform such a transfer. Let’s list out a few of those options.

Upon detection of frustration or anger

ChatGPT supports sentiment analysis, which gives you an opportunity to handoff to a human operator when you detect anger or frustration from the user.

On unknown questions

At the time of writing, ChatGPT doesn’t offer a confidence level for its answers so you it’s not possible to set a threshold below which a human should step in. However, through prompt engineering, you can force ChatGPT to state whenever it doesn’t know the answer and handoff to an agent everytime that scenario occurs.

Post-interaction survey

Another way is asking the user about their experience at a certain duration—say, 10 minutes of idle time—after the interaction has stopped. This gives your chatbot the opportunity to handoff to an agent if the user responds that they’re unsatisfied by the answers they’ve received.

Example post-interaction survey

Conclusion

Creating delightful customer service experiences is more than just slapping the ChatGPT API onto your chatbot and calling it a day. Sure, it’s a powerful tool that enables natural conversational interaction.

However, you still need to consider the main interaction model, tone of voice, persona, localization, and handoff to keep customer retention and acquisition high as you automate your support operations with the power of ChatGPT.

Need help designing an AI-powered customer support chatbot for your brand?

I’m a UX practitioner and technology consultant who specializes in B2B and fintech products. Let’s prosper together. Send me an email at niaw@cognible.com.

The Beauty of Sales Copywriting

Sales copywriting is beautiful. Just as a sculptor whittles away a block of wood to reveal a figure, a copywriter carves and shapes the information until only the words that best represent the product or service remain—in a form that helps the customer understand how it will improve their life. In the hands of a noble writer, this can help a client find something that perfectly fits their needs and desires. It only becomes nefarious when it is used to sell things that don’t add value to the purchaser or the world. In this manner, copywriting is a neutral tool that can build as much as it can hurt.

You’ve Successfully Logged In

“You’ve successfully logged in” is pretty much like a waiter saying you’ve successfully entered a restaurant. Why not just say “Hello” or “Welcome back”?

I see this quite often in the user interfaces I audit. Somehow, when it comes to information systems, people discard any semblance of common sense and humanity the way they would would throw a banana peel into the compost.

Swim Lanes and More

There’s a nice chapter in the book “How to Make Sense of Any Mess” by Abby Covert that lists down different ways of representing information. It’s great if you want to expand your toolset.

Though I can readily make block diagrams, flow diagrams, mind maps, and hierarchy charts, I realized I’m not fully maximizing the use of things like swim lanes and quadrant diagrams to communicate data better.

Marie Kondo Ultra

It’s 2030. The world is in chaos. Rising from the ashes of a tumultuous decade, an extremist group spreads around the world forcibly organizing homes and discarding things that don’t spark joy–with or without the owner’s consent. They’re called the MKU or Marie Kondo Ultra. Their tenets evolved from a fundamentalist interpretation of Marie Kondo’s bestseller, which they consider sacred.

Power Concentration

40 families control 76% of the wealth in the Philippines. These families dominate politics because of their resources and influence and the way that our electoral process favors candidates who can spend the most funds. So unless we change how we elect our politicians, elections are always going to be an exercise in futility—a choice between the silent, yellow oppression of agricultural oligarchs or the overt, red oppression of violent populists.

I don’t know the answer though. Is it a parliamentary form of government where only ministers are directly elected to avoid the popularity trap, or campaign fund caps to avoid the graft and corruption trap?

Fixed Schedule Productivity

http://calnewport.com/blog/2008/02/15/fixed-schedule-productivity-how-i-accomplish-a-large-amount-of-work-in-a-small-number-of-work-hours

Cal Newport’s idea of “fixed-schedule productivity” is similar in principle to the personal finance idea of “pay yourself first.” Simply put, allocate for the most important things first and then work with what’s left.

For personal finance, take out a percentage of your income for savings and investment, then work with the rest. For time management, block out time for the most important work, then work with the rest.

The Glitchy Glory of “Into the Spiderverse”

Just when I’d about given up hope of experiencing something truly amazing in a cinema, here comes Spider-Man: Into the Spiderverse in all its glitchy glory.

It’s damn near perfect and feels exactly like what a comic book movie is supposed to be—in both form and spirit.

While epic visuals are a staple of big budget Marvel movies, they can often feel manufactured and boastful—visuals for visuals sake. Into the Spiderverse is anything but. The look exudes heart; all the decisions feel deliberate—from the animation style reminiscent of stop motion and collage of art styles representing the different spideys and their worlds, to the little touches of comic book elements that can make you feel as if you’re living inside the page.

That’s not to say it isn’t epic. At times, I can’t believe the spectacle taking form in front of my eyes and how such a thing could have been created. I can’t get over how the movie brings to life the spirit of comic books without literally copying its form.

As far as plots go, it doesn’t stray very far from the Marvel formula. But the storytelling is superb; it does a good job of making its characters fun and relatable and unraveling the truth of its themes of hope, courage, and responsibility without being preachy. I also love how well it can rope you in and make you feel the confusing emotions and struggles of Morales as he comes to terms with his newfound powers. I won’t go too much into the characters to avoid spoilers so I advise you to just see it for yourself.

Beyond the story, I found the movie inspiring in three ways. One is how the individuals that made up the production team came together and combined their talents to produce something worthwhile. Second is how the envelope can always be pushed farther. Third is how the core theme of taking a leap of faith is so crucial to making great things happen, especially when creating something new.

If anything’s wrong with it, it was that the two and a half hours went by too quickly. Comic book movies will never be the same as this masterpiece sets the bar very high indeed. And I’m not even a Spiderman fan!