The 4 Types of AI Copilots and How to Choose Yours

So, you want to create an AI copilot?

You’ve made the right call! Since the launch of ChatGPT in November 2022, the appeal for generative AI has never been stronger. A 2023 study by ReportLinker shows that the overall Virtual Assistant market is expected to reach $46 billion by 2028, up from $11 billion in 2023. And according to this AWS study, 92% of US organizations plan on using AI by 2028.

This comes as no surprise. Virtually every industry can benefit from the AI revolution, and develop their own set of AI assistants. And with AI technologies improving at a staggering rate, we can expect to see higher-level industries, like healthcare, adopting AI tools very soon.

But for now, this technology is primarily used for all kinds of business purposes, with a keener focus on sectors like accounting, sales, and computer engineering.

From the moment OpenAI released ChatGPT, the race was on: LLM providers building better and safer models, GAFAs figuring out how to navigate in a brand-new environment, and businesses rightly viewing the technology as a powerful growth opportunity.

And this probably explains why you are here. You know your product needs AI and your users know it too!

This guide will help you shape and build your future AI copilot.

The 4 Types of Conversational AI Tools

At Blobr, we identified 4 types of AI copilots, each one adding a degree of sophistication but delivering an increasingly complex range of use cases.

The first step is to define the scope of your future AI feature. According to the kind of data you have on hand, you should be able to determine the level you need.

OpenAI’s GPTs

Primary Use Case: Specialized GPT
Technology Base: Vectorized databases from existing documentation, external APIs, function calling.

The first level of AI helper is the GPT, provided by OpenAI, which enables everyone to create specialized chatbots running on ChatGPT without coding. You can add documents, and make it use web browsing, Dall-E, and Code Interpreter — meaning that the GPT can run code and interpret files. If you want to embed your model externally, you can use the Assistant API.

Benefits

• Easy to set up with the ChatGPT no-code platform.
• Capable of performing basic use cases with your data and files.
• Available in the ChatGPT Plus plan without additional costs, unless you use the API.

Limits

• Only performs very limited use cases.
• Limited to GPTs on the ChatGPT interface if you don’t use the Open AI API.

AI Chatbot

Primary Use Case: Support Questions
Technology Base: Vectorized databases from existing documentation.

The second level of AI chat primarily focuses on providing answers based on existing documentation. If you've ever interacted with a customer service chatbot that draws on relevant information from a user manual or a FAQ section, you are already familiar with this level. By converting PDFs and other documentation into a vectorized database, these chats can quickly scan and present relevant data.

Benefits

• Efficiently handles high volumes of common queries.
• Reduces the need for human intervention for basic questions.
• Accesses vast amounts of documentation instantly.

Limits

• Human intervention still necessary for uncommon queries.
• No access to live data: if the data set changes, you need to modify it accordingly.

AI Assistant

Primary Use Case: Executing Actions
Technology Base: Real-time communication with APIs + SQL databases + Vectorized database

The third level is where AI chats truly shine, showcasing their power not just to inform but to act. Whether you want to schedule a meeting, adjust your smart home's thermostat, or initiate a complex business process, these AI Assistants communicate directly with APIs in real time to execute actions on your behalf. Think of them as your digital sidekicks, ready to assist in a myriad of tasks.

Benefits

• Seamless integration with various tools and platforms.
• Automation and simplification of complex tasks.
• Increased productivity through reduced manual input.

Limits

• Not all APIs are ready to be used by LLMs; some may require adjustments.
• 95%+ of use cases delivered but limited to the scope of the API.

AI Copilot

Primary Use Case: Advisory and Action Delivery
Technology Base: Real-time communication with internal and external APIs, SQL databases, and vectorized databases.

Venturing into the most advanced tier of AI chat, we encounter the AI Copilot. It's where the line between reactive response and proactive intuition blurs. This level is not just about answering queries or performing tasks, but about understanding context, predicting user needs, and providing advisory insights.

Benefits

• Deep integration across various data platforms and sources.
• Proactive suggestions and strategic advisory.
• Comprehensive action capabilities, from simple tasks to complex decision-making processes.

Limit

• Months of development required, possibly entailing the development of a specialized LLM.

Defining The Use Cases for Your AI Copilot

Now that you have a better idea of the different types of AI-powered tools you can create, you can start defining what yours should deliver. The nature of your use cases and their level of complexity will determine the kind of product you want to build.

The more value-loaded the use cases are, the more people will stick to using the AI chatbot. The pitfall in unfinished AI copilots is not providing sufficient responses or delivering enough tasks, making the copilot almost superfluous.

Here’s the method we recommend at Blobr to define the final list of use cases:

Let’s dive in further.

1. List all the questions and prompts

The first step is like a wish list: it enumerates all the prompts you want your copilot to answer.

What would you and your users really want it to do?

If you’re making an AI copilot for software, take a look at your product to find ways AI would be able to ease the interface and user experience. Ask your power users and use your product’s data.

And if you’re making a chat connected to data, try to find some tricky requests covering the entire spectrum of the issue at hand.

In the example above, for a global HR product like Deel, we listed prompts and questions that are both based on common inquiries for the users, and actual tasks that we would like to see powered by AI.

Don’t set limits, step 3 will see to that. Try to cover everything from very basic prompts to more sophisticated ones.

2. Determine the use cases

Great! Between your inputs and those collected from your users, you should have a pretty decent list by now.

The next step is to look through the questions or prompts one by one in order to give the use case an actual name. Some may overlap with others, which is normal and indicates how common the use case is.

Depending on your list’s length, you should have between 5 and 25 use cases: you can now sort them into bigger “families” of use cases.

If we continue with our example, you can see that we assigned a use case to each prompt and a category to each use case: ones relating to HR, application tracking, or HR-related informative literature.

However, they will not all provide the same value for your end users and will not be that easy to implement.

3. Assess your resources and finalize the list

This is now the time to finalize your use case list and disqualify some of them.

For each use case, you can estimate a use case desirability index, meaning a way to prioritize the development, by assessing the following criteria:

• The number of prompts linked to each use case: a kind of popularity contest that will highlight the most valued ways to leverage AI.
• The scope of the use case: is it a niche one performing a limited set of tasks? Does it replace a time-consuming search?
• The difficulty to implement: a technical assessment that will lead to an estimated amount of development for it to work.

For each criterion, assign a score between 1 and 10, and divide the total by 3.

This will give you an idea of just how important the use case is:

If the use case score is lower than 3, it shouldn’t be a priority, if it scores between 3 and 6, it should be kept, and if it scores above 6, it’s a must-have.

Let’s see how this applies to our example.

Help sort applications

For my Application Tracking AI project, I would like to have a way to get the top applicants for a given job offer.

• Popularity: 8 — I have one prompt related to that use case.
• Scope: 8 — This use case has an limited scope (it can only achieve one task), but it is nevertheless really useful. Basically, my users can stop browsing through résumés and rely on the copilot to pick the top 5 applicants.
• Difficulty to implement: 3 — This will not be quick to develop. The API can retrieve the data from the applicants’ résumés but it will have to be stored on a vectorized database to limit hallucinations.

Total: 6.3 — This is an important use case for my product. As it requires several layers — the API and a vectorized database — I will need to develop an AI Copilot to make it work.

Examples of AI Copilots to Help You Start

Fin is Intercom’s AI feature. It takes the chatbot idea further, thanks to an OpenAI integration. Whereas the chatbot would have only used pre-defined response patterns, Fin crunches the documentation to deliver answers by itself.

It helps answer common inquiries and deals with the first layer of customer support, enabling customer support teams to focus on more important tickets.

What makes it a Chatbot?

Fin uses the product’s documentation and stores it in vectorized databases, from which it picks out the pieces of information it needs to form an answer. This AI chatbot doesn’t replace the need for human intervention but is used in workflows to resolve the most basic issues.

The scope of use cases is limited to the available documentation and customer support.

The chatbot is not connected to live data.

Mixpanel Spark uses generative AI to interpret natural language queries, crafting detailed data reports and visualizations to drive insights and guide decision-making.

For example, if a user asks for a breakdown per country of their sign-up trend over the past year, Spark will execute the query and produce a detailed report that illustrates the pattern of sign-ups throughout the year, sorted into different geographical areas.

What makes it an Assistant?

Mixpanel is a tool with a pretty steep learning curve. One has to spend some time on the product to understand the basics and start making advanced reports. Spark makes it possible for everyone to generate reports, but it won’t provide answers to analytical questions or produce insights into these reports.

Gives everyone the possibility to create reports, and helps level the learning curve.

Has access to live data, through APIs.

Doesn’t produce advice or insights into the reports it generates.

GitHub Copilot, which has been trained using billions of lines of code, can convert prompts written in everyday language into coding recommendations for a wide range of programming languages.

Users can ask questions about their code, debug, or seek help from the copilot to design all or parts of their code.

The copilot is available in GitHub, and tools like Visual Studio Code with an extension.

What makes it a copilot?

The ability to create code, provide advice, and help the developer from the first line of code right through to the final commit makes this tool a genuine copilot.

It can predict the user’s needs, according to what the user is coding.

Gives everyone the possibility to create reports, and helps level the learning curve.

Gives everyone the possibility to create reports, and helps level the learning curve.