Since the release of ChatGPT, there has been a lot of interest around the potential about Generative AI. But what are the Generative AI use cases as of today? Many people believe tools like Chat GPT will revolutionize their business or life, sometimes rightly so. Is this true? Where and how can these tools make the difference?
This article aims to be the definitive guide to Generative AI use cases. If a real generative AI use case exists, it should be here. However, there is a key difference from most other articles you will find online. This is not a list of “cool things” generative AI can do. No, this is a list of tangible use cases, things that are now possible thanks to Generative AI, and that were not possible before.
Let’s look at the table of content for what we are going to cover:
- Introduction to Generative AI Use cases
- Generative AI Use Cases: “Jobs to do”
- Generative AI Use Cases: “Productivity Multipliers”
- Myth Busters on Generative AI Use Cases
- Conclusion
We will start by introducing Generative AI. We will start to explain what it is, and how it differs from “normal” AI, and then define how we select use cases. Read this section, as it will explain how to come up with your own use cases.
Then, we move to the use cases, divided into two sections: Jobs to do, and productivity multipliers. Read more about those categories in the introduction. Then, we move on to a “Myth Busters” section, where we see some use cases that are not good for Generative AI. We conclude with resources and a way to connect with me and get help online.
Let’s start to dive deep into Generative AI use cases!
Introduction to Generative AI Use Cases
What is Generative AI?
Even if you think you know what generative AI is, read this section. There are a lot of misconceptions on this, and most of the things Generative AI can do, other types of AI can do as well.
To start with the basics, we need to start with the definition of AI (Artificial Intelligence). This is a collection of software programs, that instead of being explicitly told what to do, they figure it out by looking at data. For a traditional algorithm or program, a software developer defines a precise set of steps it must execute in a specific order. Instead, an AI program looks at tons of data and figures out what needs to be done (this is an oversimplification, of course).
Some types of AI algorithms are quite simple. For example, through a linear regression, we can write software that in a few lines of code identifies there is a relation between surface area and price of a house. After looking at many area-price combinations, the software can make predictions by finding a coefficient (say, $1k for 1sqft). Other types, such as deep neural networks, are much more complex and involve a complex network of “virtual neurons”. Generative AI uses the latter, obviously.
AI can be of two types: inference, or generative. The goal of inference AI is to make predictions. It looks at a lot of data and find patterns. Then, when you give it some inputs, it can predict some output. The case of predicting the price of a house based on its surface area falls in this category. Yet, we can have much more complex inference cases that rely on neural networks. For example, we can use inference to predict the future price of a stock. But predictions don’t have to be necessarily in the future, they are just about finding patterns. Thus, we can use this type of AI for things like finding a tumor in an X-Ray scan, or computer vision (e.g., do you see a traffic light in this picture?).
As you might expect by now, the goal of Generative AI is to create content. This content can be of any kind: text, video, images, audio, documents with special format, or whatever. But the thing is, the goal here is not to make a prediction or find some pattern. Instead, it is to create something new. The way Generative AI works is not that different from inference AI. It looks at a lot of data and figures out how to generate content.
For example, tools like Chat GPT are trained with tons of question-answer pairs so that they know what to generate when you ask them a question. They don’t have a question bank: the AI looks at many human-created question-answer pairs and “gets the hang of it”, and then can generate brand new answers independently. The same is true for tools that, for example, remove the background from images. They look at a lot of pairs of images before/after the background is removed and get the sense of it.
Under the hood, all complex AIs are the same. They have a neural network, which is a graph with nodes connecting to one another. Each connection has a numeric weight, and information that comes in (that is also converted into numbers, even if it is text or images), flows through various nodes based on the weights of the connections. Each node performs some mathematical function. Eventually, we reach the nodes at the end of the network and we get the numerical output, which is then re-assembled in an image, audio, text, or whatever media you need.
Those networks can have thousands or millions of nodes, and we cannot be sure why the AI decided to have some specific weight in some specific node-to-node connections. In other words, we can know that AI is generating good content in, say, 90% of the time, but we cannot be sure why or how it is doing so. Keep this in mind, it will come to bite us later.
How do we define “Generative AI use cases?”
Or, to say it differently, what differentiates a real generative AI use cases from a “cool thing to do with Generative AI”?
To test if a use case is real, I ask two questions:
- Is this driving value?
- Was this possible before, or with other tools?
If the answer to the first question is yes, and the answer to the second question is no, then you have a valid use case.
The first thing is about driving value. In other words, you need to use Generative AI to do something useful. Use it to create something you need. For example, you can use it to write code in most programming languages, and then have your (human) software programmer refine it and check it does what it is supposed to do. Generating kitten pictures may be cool, but arguably it does not deliver any value.
Once we have established the use case drives value, then we need to ask if this is something enabled by Generative AI. If it is, then we are good, and this is valid use cases. The example of generating code in programming languages still holds here: previous tools, including inference AI, could not create something like that. However, we may have a ChatGPT-like tool that tells us about the weather in colloquial form. While this is technically enabled by Generative AI, it is not much different from looking at a traditional weather app. Therefore, even if that drives value, it would not pass the “enabled by GenAI test”.
With these two tests in mind, I prepared multiple use cases of Generative AI. But this was not enough. I wanted to select only the ones that can actually work in practice. How did I do that? Read the next section.
Criteria for identifying use cases – including your own
If you come up with a GenAI use case, the fact that it drives value and it was not possible before GenAI is not enough. You need to be sure this use case can be effective in the real world.
When talking about “What is Generative AI?”, I mentioned that the problem with deep neural network, which power GenAI, is that we cannot deterministically know how they decide to generate content. So, everything GenAI (or everything AI for the matter) is a game of chance, or statistics to be precise.
Explaining this on an inference (prediction) problem is easy: you can say your AI app makes a prediction right 99% of the time, 99.99% of the time, or whatever. But the same can apply to Generative AI, we can say it responds properly 99% of the time, 99.99% of the time, or whatever. Realistically, percentages will be more around 80% to 95%, but you get the idea.
No matter what is the actual percentage, there is always the possibility that AI may get it wrong. Thus, you need to consider two items when assessing the validity of a real use cases:
- Can I afford some uncertainty around this?
- What is the level of uncertainty I can tolerate?
Warren Buffett makes the example of a Russian roulette when talking about investments. You have a gun with all empty chambers but one which contains a bullet. According to Buffet’s view, you should not put that gun to your head and pull the trigger, no matter if you have 5 empty chambers and 1 bullet or 1 million chambers and 1 bullet. The impact, of losing your life in that case, is so big that you do not tolerate even the slightest chance. I completely agree.
Buffett then pivots to explain how the same thing applies to investment: if there is a company that has a chance of losing all your money, you should not bet all your money on it, no matter how slight the chance.
GenAI is like the gun with thousands of empty chambers, but still one bullet. You need to think if you can afford to get one bullet now and then in your use case or not. To assess that, you need to define the impact of getting one bullet, and also what are the alternatives (i.e., the alternatives of not playing this game at all).
For example, you could use inference AI to detect tumors in X-Ray scans. Sometimes, it will get it wrong, say in 0.01% of the cases. But what is the alternative? Doctors. Do doctors get it wrong in 0.01% of cases, more, or less? If AI is better than the previous system, then it is better to go ahead even if we still have the possibility to get it wrong, because we are still improving.
You may be able to use GenAI to write a blog post, for example. However, you don’t want to use it to write the annual letter to stakeholders that a company must file with the Securities and Exchange Commission. GenAI would be capable of writing such letter, of course, but what if it gets something wrong? In the blog post, no big deal. In the letter to stakeholders, you risk lawsuits from investors, or potentially even prosecution by the SEC. Probably not worth the risk.
So, to identify a real and valid use case, think where can I afford to have the level of uncertainty that GenAI is giving to me? If you find areas where this is acceptable to you, then you are in the perfect spot to have some Generative AI use cases. In general, everything that is generated in small chunks and reviewed by humans will be fine.
Categories of use cases
Based on the topics covered so far, we can identify two main areas of use cases for GenAI, or categories: “Jobs to do”, and “Productivity Multipliers”.
We consider jobs to do things that were previously done by humans, or not done at all, and that now GenAI can do independently, with a decent level of confidence. Here we don’t have humans to review the deliverable, or if we do it is in the same way as a supervisor would review the work of a subordinate.
The other category is about productivity multipliers. Here, GenAI is not on its own. It generates some content, which is then reviewed, altered, or enhanced by a human. In this case, GenAI is not independent, but works as the assistant of the person who is doing the job. Here we find things like writing code: the programmer starts to write code, GenAI proposes how to complete it, and then the programmer (human) decides if to accept the suggestion, and if so what to change in it.
As we will see, because of the uncertainty that is embedded in AI, most of use cases are productivity multipliers, and not jobs to do.
Generative AI Use Cases: “Jobs to do”
Customer support
The only viable case I could identify for Generative AI in the realm of “jobs to do” is customer support. This means having a chat where a customer asks questions, and GenAI produces answers about your company, product, and services.
GenAI can go a step further, and integrate with existing tools and APIs so that it can authenticate the customer and provide information about his own services, subscriptions, and products. He can solve issues like changing the address of an invoice, responding to inquiries, upgrade and troubleshoot products.
Some actions may be beyond GenAI because they are too risky, for example if you run an accounting software company you may want to avoid having GenAI touching the accounts of your customers, and the same if you have a banking app. But this is also true of human customer support agents: not all agents will have the same privileges and permissions.
As a result, you can use GenAI to do the job of all the low-risk agents, and keep human agents for higher risk activities, or for escalations when GenAI is not able to solve the problem.
Note that customers don’t like to chat with an AI, because they feel “you don’t even have the care of putting a real human to support me”. It feels like still self-service, and not customer support. So, you need to monitor clearly if GenAI is resolving the problems customers have and give them an easy valve to escalate to humans. It feels so frustrating for the customer if they don’t get the possibility to talk to a human because GenAI insists on handling the request itself.
Here we are not inventing anything new: customer support has been around for ages. However, with GenAI we can afford to have it open 24/7, guarantee a more consistent experience, and at a lower cost. As a side advantage, companies that before did not have customer support at all, can now afford to have it 24/7.
Generative AI Use Cases: “Productivity multipliers”
Code generation
Generative AI use cases find a lot of traction when it comes to multiply productivity. That is, help human do things they were already doing, but faster and better. Personally, as I often write code, I have been greatly helped by GenAI in that.
If you are not familiar with writing code, it is not that different than writing a Word Document. The main difference is that you have to know and follow a specific syntax, and that you work with many files that are interdependent from one another, not with a single file. As you write, the AI realizes what you are about to write and proposes you to auto-complete. This auto-complete can be a just finishing the word you are writing, up to a few dozen lines of code. You can hit tab and have the AI auto-complete for you, or simply skip it if it is not appropriate.
If it is appropriate enough, you would accept the proposed addition to your file, and then you are free to edit it any way you want. This is a good way to enhance your productivity, as it helps you to do all the scaffolding of your projects and simpler parts of your job faster. After speaking with some programmers, I understand this improves productivity by 5 to 15%. Taking an average of 10%, that means you save about 4h per week if you are working full time.
The best part is that this is available even for free now, thanks to tools like AWS CodeWhisperer or GitHub Copilot.
Summarization
Second in our list of Generative AI use cases we find summarization. Generative AI can parse a lot of information and give you a summary. This is helpful when you don’t need super-accurate information and can survive with the occasional grammatical error.
This can be helpful in many circumstances. For example, you can have AI writing summaries of documents, so that when people look at documents in the archive they don’t have to read through the whole document, but can skim a summary. Pairing this with an inference AI algorithm that categorizes and labels documents can boost your ability to find and retrieve information. This is particularly helpful for the legal and consulting world, where you have to deal with documents with many customers, and with many people working on those documents.
Another helpful use case is in customer support, where you can have AI write summary of customer cases or tickets, helping your agents understand what has been happing much faster, without having to read through all updates.
This may be a weaker use case compared to the code generation, but it still helpful, and in certain cases it can make the difference.
Assistant & Travel Planning
In big companies, executives and directors get to have an assistant. The assistant takes care of all the bureaucratic work for the executive, so that they can focus their time on making important decisions. Thus, the assistant will take care of scheduling meetings, circulating notes, book travel, submit expense reports, and so on.
Booking travel is something that can be easily automated with GenAI. Hooking some APIs from companies like Booking, Skyscanner, Google Maps and so on, Generative AI can build a travel itinerary and do all the necessary bookings. Of course, here we do not want to risk AI booking a wrong flight, or do it in the wrong name. But still, this is something easy to check.
We can have Generative AI build the itinerary and submit it to the executive for approval. Then, they just have to click a button or reject it, and if they approve the algorithm (now through reliable APIs, and not “guessing” AI) will submit the bookings.
This is something that can be implemented fairly quickly and cheaply, and yet can boost the productivity of large companies. In fact, while directors and executives have assistants, managers and senior managers rarely do. Yet, their time is still valuable, and a tool like this can help save a lot of their time.
Image editing
Another important use case of Generative AI is image editing, or video and audio editing for the matter. This use case works well because it is something that has to be reviewed and approved by a human, much like code generation or travel planning.
There are several tools that can edit images. Typically, you make an ask to those tools, such as “please remove the background”, or even “draw a picture of dancing SpongeBob”. They can do that, and then you can see if you like what they produce. You might, or you may need to tweak them somewhat, or discard it if the AI did not really get your request. Tools like Adobe Photoshop are starting to embed these capabilities.
Generative AI for image editing is great for a variety of professions. It can aid graphic designers and video makers to be faster, and in this case, it helps improve their productivity by about 20%. That is a big improvement! But those are not the only professionals benefitting from GenAI when it comes to image generation. In fact, the people who benefit the most are the ones who need images, but don’t create them themselves. Mostly, we are talking about content creators, website designers, or SEO content writers.
Those people need images in their projects and are probably flexible around the design. Instead of spending a lot of time creating images (since they are not qualified for it), or engaging a paid professional, they can rely on an AI tool. In this case, their productivity savings are much more impressive, hovering around 30%.
Image editing currently takes longer than text generation with GenAI, but already much less than if you were to do it manually. This is a mature use case: it may not help everyone with their productivity, but if you are in the niche that needs images edited, then this tool is great for you.
Content generation
Another valid Generative AI use case is content generation. After all, Generative AI is all about generating something. In this case, we use AI to generate content, typically text, like emails or messages to send to customers.
GenAI is useful to generate content if you want to do one of two things:
- Give structure and flow to some data you already have
- Come up with new ideas, brainstorming
However, as of 2024 I believe Generative AI is not ready to do both things together. This is why, when addressing the myths, we will see that GenAI is not well suited to writing blog posts. But in this section we are focusing on what GenAI can (and should) do, so let’s not worry about that for a moment.
The first use case is give structure to existing data. For example, in this case you may want to create an email to send to a customer. You know already what information you want to provide to the customer, but you don’t want to spend time thinking about how you can piece them all together. Here, you can give a summary of the information to GenAI and some instructions like “Write a formal email”, and it will generate something reasonably good.
At this point, you only need to proofread what AI has prepared for you and you are ready for copy-and-paste. This is a valid use case because you focus on the content, while you let GenAI figure out the how (the structure of the email).
The second use case for GenAI when it comes to generate content is brainstorming aid. Unlike giving structure, where you already know the content, here you don’t even know where to start, and you are looking for new ideas. GenAI can give you an hint, and let you refine it. Once you refined it and put some thought into it, you can then feed the refined idea back to GenAI to obtain a structure. To do this, take any GenAI-based chat bot and ask a question. The more specific to the kind of aid you want, the better.
For example, we can ask GenAI to help us start writing a book. It is often hard to find the first sentence, but everything else just flows after that, so why don’t we ask GenAI to prepare us the first sentence? Here is a prompt I used.
Write the incipit of a book set in New York City, writing should be dry and express a sense of helplessness. Write max 3 short sentences.
This is the result:
In the heart of New York City, where towering skyscrapers pierce the sky, lies a maze of bustling streets that swallow souls whole. Amidst the ceaseless chaos, individuals drift like lost ships in a sea of anonymity, their dreams swallowed by the city’s relentless grind. Here, in this concrete jungle, hope flickers dimly against the backdrop of an indifferent metropolis.
Not too bad, right? Even if this is not what we want to express, it gives us a lot of ideas. It gives us a bunch of words we can use to express the sense of helplessness, which then we can edit or rearrange the way we like: towering, maze, swallow, drift, anonymity, grind, jungle, dimly, indifferent. Would you have come up with all those words yourself? If you are like me, probably not. Of course, this result seems a little too much (at least for me), but gives us a great starting point to refine it.
Myth Busters on Generative AI Use Cases
Now that we have covered the real Generative AI Use Cases, let’s move to things that we believe are not genuine use cases for GenAI. Those are all things that were possible before GenAI, and where GenAI is currently not performing at the same level as traditional tools, or where it simply isn’t the best tool for the job.
Writing blog posts
This is the first item I wanted to address. You cannot use Generative AI to write blog posts. Sure, it can help you (like we saw in the “content generation section”), but it cannot really write an entire blog post by itself. Even if you proofread it and adjust it, the result will be of extremely poor quality.
But why is that? After all, Generative AI can give structure and can come up with new ideas. The problem is that Generative AI still isn’t too much context-aware. I will explain this in a bit, don’t worry.
When writing a blog post, you need to start by thinking about the purpose of that post. Why are you writing it for? Who are you writing it for? There are several answers to this question, none is right or wrong, but most commonly we see:
- For SEO, so that new people online can discover your website
- To inform your existing visitors about news, teach them something, share knowledge
- For yourself, just to log what happens in your life online
Now you can start to think how GenAI can (or cannot) help you achieve that purpose better, faster, and more efficiently. Unfortunately, at least as of today, it cannot.
Taking the SEO example, to write a proper blog post that has a decent chance of ranking with SEO, we need to do a few things and be aware of the context. First, we need to select a list of search keywords that have enough traffic. Then, we need to identify for which of those we can actually write something that would add value, that is not the same as what is already online. And for each of those, we need to look at what other sites are doing, and be honest if we are able to do something better. Once we have identified a keyword for which we can add value, do better than existing competition, and that is actually searched by people, then we can write a post.
By definition, GenAI is trained on material that already exists, and will generate something spun out of that. Your unique value in blog posts often come by your personal experience and narrating style, none of which GenAI can have. Thus, it will not help you to write a SEO blog post.
Stock trading
This one is funny, one of the generative AI use cases we definitely need to debunk. Simply put, you cannot use generative AI to trade stocks. Or, well, you could, but that would be a bad idea.
The first reason that makes stock trading a no-go for GenAI is that it is not really a generative task. Instead, it is an inference task: you have a set of data, and you want to predict something (future price, movement direction, trading volume, and so on). So, this is something GenAI does not add any value to: traditional AI can do the job just fine.
This alone would be the reason for which Generative AI use cases should not include trading. However, we can push ourselves even a little further, saying that AI is not a great tool for trading as well. At least, it is not for the individual trader, or to make trading choices.
The first argument is that you cannot really define why AI is making a prediction or another. Thus, there is a lot of opacity in the mechanism, and you can only rely on a statistical track record to see if your AI is effective at making trading (e.g., after 1000 trades you see that 65% of them result in a gain). This may be fine if you are trading by yourself, but imagine you are a hedge fund trading for your customers. How can you explain to them “Well, we lost your money, but we really don’t know why”?
The second argument is that you need to compare AI against what is already there. At the basic of investment, you can choose not to invest and do not lose nor gain money. If AI can make you even a single cent, then it is better than doing nothing. However, other than not investing, you have other investment options. A common benchmark is the S&P500, which is the collection of the 500 biggest companies in the US. You could also consider the MSCI All Countries World Index, which takes most of the quoted company in developed and developing countries. Either way, you can average about 7-9% returns per year. Is your AI able to beat the market?
Most likely, it is not. Most professional investors are able to only slightly beat the market, but such extra gains are eaten up by their commissions. It is unlikely that AI will give you such an advantage to the market. Most likely, it will be able to make some money, but to a lower return than the general market.
Medical sector
Just like trading stocks and financial markets, there are a lot of misconceptions about GenAI and medical sector. Specifically, there are a lot of Generative AI use cases that do not really stand the test of reality in this sector. The problem here is similar: GenAI hasn’t much to generate.
There have been studies where AI was effective at detecting tumors in X-Ray scans, and be even better than doctors that instead did not even notice a gorilla in a scan. However, this has nothing to do about Generative AI, it is just plain and old inference AI. GenAI can add some fancy words around it, but in the end, it is inference AI that detects pathologies.
Conclusion
Additional resources
This article is the ultimate guide about real world Generative AI use cases. So, if you find one that pass the tests we defined, we should definitely add it here if it is not. So please reach out to me (see below) and let’s add it to this guide. Additionally, here are the additional resources on this topic that me and other people have already provided.
- Amazon CodeWhisperer
- GitHub Copilot
- Hugging Face
- TensorFlow
- Zero to Mastery course on TensorFlow
- CNN Keras AI deep dive
Connect and get help
In case you want to discuss this article, share feedback, or simply thank me please connect with me on LinkedIn. I ask you just one kindness: add a note to your message and specify you found me through this article, so that I know why you are adding me. I accept everyone who connects with a note, but I reject people without a note because they are often spammer.
