Should Your Company Launch a ChatGPT Plugin?

The springtime launch of generative AI plugin features created a major opportunity for brands and technology companies. Initial plugins for OpenAI’s ChatGPT platform—an advanced language model that can engage in dynamic and realistic conversations—include names like Shopify, Expedia, Instacart and more. These plugins enable chat platforms to access company data systems and take actions, from booking a trip to building a grocery order.

Should your company launch a plugin? What do brands need to know before development? Are there risks? To learn more, The Bounteous Innovation Lab team developed a Proof of Concept (PoC) plugin to learn about platform functionality and provide client recommendations. Much of what we learned applies across platforms, but our team focused on ChatGPT for testing. Tips and ideas are featured below, but first, why should brands consider plugin development?

Why Building a ChatGPT Plugin is a Worthwhile Investment

According to OpenAI’s website, “Language models today, while useful for a variety of tasks, are still limited. The only information they can learn from is their training data. This information can be out-of-date and is one-size fits all across applications. Furthermore, the only thing language models can do out-of-the-box is emit text. This text can contain useful instructions, but to actually follow these instructions you need another process.”

In other words, generative AI text tools output text, but can’t natively take actions on information like scheduling a meeting, adding oat milk to your shopping cart, or booking a vacation. Enter plugins! 

Plugins are designed to overcome these function and data limitations. Brands are now able to build plugins that allow ChatGPT to access real time information and perform actions on the behalf of users via the brands API. If your company helps people solve problems, sells something, or has a mobile app, you may already have reasons to build a plugin for ChatGPT. 

Testing A Custom ChatGPT Plugin - Developer Learnings

In order to understand any functional strengths or limitations, our Innovation Lab built a PoC brand plugin to learn how businesses would launch a plugin. 

Use Case Development

Bounteous partners with a number of leading retail, dining and c-store brands. We offer an excellent ordering recommendation system. We also knew from past experiences that chatbots struggle with order intake when customers request multiple items or change their order at the last minute. For testing purposes, we built a plugin PoC that acts as a dining ordering assistant.

Easy to Get Started

When companies start to build a ChatGPT Plugin, we recommend creating a manifest file that provides a clear description of what the plugin can do. It’s important to document things in a way that the Large Language Model (LLM) will understand. Our team found setup processes to be easy and painless.

Plugin Strengths

During testing, we found the Plugins feature was extremely smart and knew when to trigger. Users can be having a conversation with ChatGPT about something completely unrelated, but the moment there is a reason to, ChatGPT will trigger the Plugin and call any relevant API endpoints such as getting a menu, retrieving information from the test restaurant like hours of operation, starting an order, or adding/editing things on the order.

ChatGPT is smart enough to do several things at once. When tackling an input like:

“I would like to order two cheeseburgers, one with no onions. Also I would like a pepperoni pizza and a shake.” 

Our custom plugin and ChatGPT knew to ask for the menu to make sure these things are available, create an order, then add the items, along with including customizations like no onions. 

Developers don’t need to focus on math. In our PoC we included an order summary and ChatGPT used this information and presented it back to the sample customer accurately.

Challenges with ChatGPT Plugins

As with any software project, there were a few challenges and limitations. APIs today are written in a way that makes sense for building custom applications, but often not in a way that a LLM can fully understand. We recommend using a middle layer to make your brand’s API more friendly for the LLM and including additional information that you typically wouldn’t need for a custom application.

Users have to seek out your company’s plugin and install it. Currently there is little as far as discovery options and users will likely have to explicitly search for your plugin. We expect discovery to improve over time.

Expect little control over what and how ChatGPT presents information to users. In some tests we found that the chatbot would provide the images we gave it; other times it wouldn’t. Developers can guide it to a degree, but the chatbot ultimately decides how it wants to communicate back to the user. Overall, ChatGPT does a great job of communicating to the user, but it may be challenging to accurately display your company’s branding.

ChatGPT currently limits end users to three active plugins at once. This means that your brand will be competing for attention. This also means that unless your brand’s plugin provides daily value, users are likely making a decision to use your plugin, activating it, then performing their task versus activating as needed based on normal use.

Testing can be difficult, because developers are limited to using web interfaces. In addition to challenges with automation, developers can provide the same input and get different results due to the nature of the degree of randomness involved with ChatGPT. To further complicate things, ChatGPT state management is the entire conversation. This means that it may require several prompts to get to a state that you want to test. 

Tips for Building Your Company's ChatGPT Plugin

Start Small and Focus on User Value.

You may have several services or features to offer your users, so focusing on a particular service or feature can lead to a faster launch. Your team will learn a lot from a simple test. Especially early on, focus on things that users may want to seek out to make things easier for them if they can do it in a conversational interface.

Build a Separate API/Middlelayer for ChatGPT to Use

Plan on building a specific interface or way for ChatGPT to leverage your services independently of current APIs. This provides the flexibility to combine information and structure it in a way that is more user-friendly for LLMs.

Make a Testing Plan

Plan on manual testing, have several steps to your testing, and test the same scenario more than once. Because there is a degree of randomness in the responses, it is best not to assume that just because a test passed once it will pass every time. Also planning an end-to-end scenario for testing is necessary to test features that may appear further into a conversation.

Ask Users to Perform the Final Action Manually

Allow ChatGPT to do the prepwork and actions, but always ask the user explicitly to perform the final action. 

Some examples are:

• ChatGPT: draft an email and User: manually sending it
• ChatGPT: create an order and User: manually reviewing the order and submitting it
• ChatGPT: Finding a vacation destination and starting a booking and User: Manually reviewing and booking the vacation 

Focus on User Value vs Branding

Remember, ChatGPT is acting as an assistant to accomplish things for the user and not specifically as an assistant of your brand. It is easy to want to have a ChatGPT personality to match your brand, but it does a fantastic job of being friendly and helpful to the user. With that in mind, we recommend that your team focus more on creating value in what the users can do, and a bit less on brand safety, as you test and experiment.

Our Take

Start building! Our Innovation Lab expects LLM plugins for platforms like ChatGPT and other tools to quickly gain traction in the market. This is a landmark opportunity for brands to experiment with new types of value delivery.