The Four Essential AI Use Cases

Six years ago, I stumbled upon Grant Sanderson’s YouTube video, What is a Neural Network?. At the time I was working as a consultant helping banks use statistics to catch financial crime. I knew a little bit about machine learning and I had heard that deep neural networks were the next big thing, so I turned to Sanderson’s YouTube channel, 3Blue1Brown, to educate myself on the subject.

What I learned from that YouTube video - and the hundreds of others I watched afterward - changed the course of my life and career. I became obsessed with deep learning and decided to devote my career to it and, in particular, to finding real, practical applications for it. I started applying for jobs in tech and eventually found my way to Yext, where I’ve been fortunate to do some incredible work on AI products like Yext Search and now Yext Chat.

I was floored by the power of the idea in Sanderson’s video: an algorithm that mimicked the human brain and could learn to perform arbitrary tasks. The video explained deep learning in the context of the famous MNIST example, in which a tiny neural network learns to identify hand-written digits. But it was clear to me that this same algorithm could fundamentally learn anything. There was no obvious limit to what it could do.

Alt Text

This was in 2017, and at the time deep learning was starting to pick up steam. The previous year, the New York Times published a piece on The Great AI Awakening describing how although neural networks had existed for decades, they had never quite found widespread commercial use, until now. Suddenly, they were beginning to solve challenges previously thought to be unsolvable. Later that year Google published the now famous Attention is All You Need paper, which introduced the revolutionary transformer architecture.

But if you had told me in 2017 all the progress we’d make by 2023, even I would have been astonished. It’s exceeded even my wildest expectations, and if the next six years are anything like the past six have been, we are going to be living in a very different world by 2030.

So what’s changed? What can we do in 2023 that we couldn’t in 2017? There have been many advancements, including improvements in GPU hardware, the introduction of powerful pre-trained transformer models like BERT, a cambrian explosion of the AI open source community on HuggingFace, and new breakthroughs in biology like AlphaFold. But the most important advancement of all - and the reason that you’re reading daily about AI in the press - has been the rise of generative AI, and especially large language models like GPT-3 and ChatGPT.

“Generative AI” refers simply to AI models that generate things, like text or images. Generative AI is not new. Generative adversarial networks - a predecessor to image generation models like Midjourney and Stable Diffusion - were invented by Ian Goodfellow in 2014. And autoregressive language models like GPT-1 and GPT-2 have been around for even longer than that.

What’s new is the scale of these models, as well as their training procedure. What makes GPT-3 so much more capable than any of its predecessors is that it is much larger - an order of magnitude larger than even BERT, which was at the time considered a very large model. And what made Chat GPT and GPT-3.5 even better than the original GPT-3 were improvements to its fine-tuning, specifically a process called reinforcement learning from human feedback (RLHF), in which the model learns to conform to human preferences.

Increasing the size of these models has made them far more intelligent and flexible, but perhaps most importantly it’s largely obviated the need for large training datasets. Acquiring training data has historically been the hardest part of machine learning, and the most prohibitive for many companies.

For example, if you needed a model to determine a user’s intent when speaking to a chat bot, in 2017 you would have had to train a model on tens of thousands of example messages, which means that you would have to label tens of thousands of messages, which is time-consuming and expensive.

But in 2023, you can simply “prompt” GPT-3 on the task at hand, describing it as you would to another human, and GPT-3 will simply perform the task based on your instructions. This is called zero-shot learning and it massively lowers the barrier to using AI.

The ability to simply prompt rather than train, along with huge upgrades in intelligence and coherence, makes LLMs a game changer for businesses and consumers. It makes possible things that would have been considered science fiction only a few years ago. It allows us to automate new types of work and create new types of digital experiences for our customers. I believe this has the potential to create enormous value for the economy and, more importantly, genuinely improve peoples’ lives.

But for that to happen, we all need to understand AI - what its strengths and limitations are, how to use it responsibly and effectively. As a practitioner and optimist about AI, my goal in this essay is to help other business leaders to wrap their heads around this technology, understand how it’s going to affect their organization, and make sound decisions about it.

Four Use Cases

Rather than talk about AI in the abstract, we’ll discuss what I consider the four main use cases for AI in 2023. I recognize that talking about use cases for AI in 2023 is a bit like talking about use cases for the computer in 1970. If, in 1970, you had tried to explain the potential use cases for computers, you might have come up with such examples as performing mathematical calculations on large data sets, or setting up a database for a company’s inventory.

But of course that would have sold computers far short. We use computers for much more than just doing math or keeping track of inventory. Computers suffuse every aspect of our lives, and today we use them for many things that in 1970 would have been completely unimaginable. I think that AI represents a platform shift as significant as personal computing, mobile computing, or the internet, and that it will come to suffuse every aspect of our businesses and lives just as those technologies did.

So identifying concrete use cases feels a little silly, and I also recognize that there is considerable overlap between the four use cases I discuss. But still, I think it’s the right place to start, because many organizations are struggling to understand where this technology fits in and what it can be used for. Furthermore, separating out different use cases allows us to make better decisions about risk, because depending on the use case the risks are very different.

For each use case, we’ll outline example applications, the specific technologies and algorithms it relies on, how human involvement works, and what’s different_ _now that we have generative AI.

AI as a Creative Assistant

The first use case is using AI as a creative assistant. In this case, AI is helping a high-level creative employee by providing suggestions, retrieving information, and supporting output. I’m intimately familiar with this use case because I use it almost every day in the form of GitHub Copilot and ChatGPT.

Copilot and ChatGPT help me write code, which is part of my job as a product manager. Copilot helps autocomplete my code or write code based on my prompted instructions. And I use ChatGPT as a sort of replacement for StackOverflow, asking it to help me debug things and think through novel problems.

It’s hard to overstate how much more productive these two products have made me over last year. Especially when I’m working in unfamiliar territory - e.g. in a programming language I don’t have much experience with - Copilot helps me quickly translate the ideas in my head into working computer code. ChatGPT helps me debug issues and understand new concepts much more quickly than reading documentation.

Copilot is a special creative assistant for software engineers, but in the future every industry and application will have some form of creative assistant embedded into it. There are already a number of other AI assistant products emerging for other industries such as Jasper and Copy AI for copywriting, or UIZard and Galileo for UI design. Additionally, many software products like Notion and Microsoft Office are beginning to embed AI assistants directly into their products.

I anticipate that this year we’ll see the emergence of other industry-specific creative assistants for fields like law, medicine, entertainment, advertising, academia, and finance. Here are some specific examples:

Legal creative assistants will help lawyers write briefs and contracts by quickly surfacing and summarizing relevant cases. Often the assistant will write the first draft and the lawyers will tweak as needed.
Medical creative assistants will help doctors to diagnose patients based on a written set of symptoms, or to write correspondence to insurance companies like this guy already is doing.
In entertainment, creative assistants will help screenwriters generate new ideas for TV episodes and even write the first drafts of the dialogue.
Advertising creative assistants will help write first drafts of commercials or other ad copy based on a prompt. They’ll also be able to create visuals or even entire videos.
Academic creative assistants will help professors write papers by surfacing and summarizing relevant academic literature.
Financial creative assistants will help bankers fill out complicated Excel formulas in their models and summarize relevant market data.

These applications are really only possible because of generative AI and LLMs. With the exception of Clippy, this category of product basically didn’t exist before 2022, but since then we’ve seen such an explosion of new startups. Before LLMs, AI models simply were not smart or flexible enough to provide much help in complex creative domains.

Importantly, when using AI as a creative assistant, the human is still very much in control. The AI will make suggestions, surface/summarize relevant information, or try writing a first draft. But the human - usually a highly capable knowledge worker - ultimately has the final say. The human gets to accept, reject, or tweak the AI’s suggestions.

This means that the cost of mistakes is quite low. When I use GitHub Copilot and ChatGPT, they absolutely make mistakes. Occasionally they say things that are outright wrong. And even when they don’t make outright mistakes, I often chose to tweak or clean up their suggestions anyway. I almost never accept their suggestions blindly.

But that doesn’t mean they aren’t still extremely useful, or that they don’t save me an incredible amount of time. After all, it’s much easier to edit something than it is to write it from scratch. Sometimes all you need the virtual assistant to do is to point you in the right direction.

Overall, this use case has the potential to make knowledge workers much more efficient. It augments workers, rather than replacing them. And there is minimal risk.

AI as Business Process Outsourcing

The second use case involves using AI to automate entire business processes and completely replace humans, rather than augmenting them. This is similar to how many firms automate certain rote, highly repetitive processes to other firms and other countries (hence “business process outsourcing”).

One of the very first projects I worked on at Yext was in this area. I’ll spare you the boring details, but basically we needed an algorithm that could determine whether two listings on the internet were a “match”, i.e. they represented the same person or place in the real world.

We needed this because our Listings product would scan different publishers like Google Maps, Yelp, and TripAdvisor and we needed to reconcile the datasets from different publishers. (For example - does this McDonald’s on Google Maps match that McDonald’s on TripAdvisor?)

This was an enormously expensive business process for Yext. By some estimates it was costing us close to $2 million per year. And it was just complicated enough that we couldn’t quite write a program to automate it. There was just enough nuance that you needed a human to go through the listings line by line.

We quickly realized that, although a traditional program couldn’t perform this task, perhaps a machine learning algorithm could, and so we set out to train one. The resulting project taught us a lot about how to completely automate a human process using AI. We trained a random forest model model - a much older and simpler statistical model than the LLMs we’ve talked about so far - to perform this task, based on tens of millions of labeled examples we had compiled over the years.

As time went on, we gradually allowed the model to handle more and more of the human workload with less and less supervision. Soon enough the model became more accurate than humans, but we still required human review for some percentage of tasks to ensure we caught any signs of model drift.

A huge number of business processes like this one will be candidates for automation in the coming years. Many of them are esoteric little tasks like Yext’s matching problem that are specific to each individual company, so it’s hard for me to predict exactly what this will look like.

However, a few specific examples might include…

Classifying and routing support tickets - something that every call center does thousands of times per day
Medical coding, which is the process of translating diagnoses and procedures into standardized medical codes and costs the US economy more than $400 billion per year, according to a 2016 study
AML/KYC (Anti-Money Laundering / Know-Your-Customer), which is the process of financial institutions doing due diligence on new and existing customers

Of course, it’s been possible to automate tasks using AI and machine learning for many years now, and often older ML models are the right tool for the task, as in Yext’s case. However, LLMs significantly raise the bar for what can be automated while at the same time lowering the technical barrier.

Our Yext matching project required millions of rows of training data and a data science team to train and deploy a model. If we were starting that same project today, I’m not sure we would have needed either. We might have instead given GPT-3 instructions on how to perform the task.

It’s also important to remember that LLMs are not just capable of writing - they can also reason, predict, and classify. This means they can be used to automate a wide variety of tasks, not just ones that involve writing prose.

What’s challenging about outsourcing a process completely to AI is that there is, by definition, no longer a human directly in the loop. The whole point is to remove the costly human from the equation. This means that the requirements for accuracy become much higher, and additional safeguards are required to monitor and retrain the model over time.

One of the best ways to mitigate risk is to fine-tune or supervise the model, which means training the model on many thousands or millions of examples of the task, as opposed to using the zero-shot learning, where you merely give the AI written instructions. The more examples you have, the more accurate the model becomes, minimizing the risk of mispredictions.

But the risk is never zero. In evaluating this risk, however, it’s important to always ask “What’s the alternative?”, which is usually to have humans continue doing the task. Humans have the disadvantage of being both costly and slow, but if they are a more accurate/safer option then that cost is justified. However, humans are not always a more accurate or safer option.

As we found in the case of Yext’s matching problem, our machine learning algorithm quickly surpassed human performance. In other domains like self-driving cars and medical diagnosis, AI is quickly catching up to and exceeding human performance. And those are relatively complex tasks - for many routine business operations, AI today will surely be able to match or exceed human performance. Therefore, the question should not be “Can AI do this perfectly?” but instead “Does AI improve the status quo?” and in many cases the answer will be “yes”.

The lowest hanging fruit here will be highly repetitive back-office work in the information economy. Importantly, AI still struggles with applications in the physical world, as evidenced by the surprising lack of adoption of self-driving cars in the last ten years. It turns out that getting AI to interact with the physical world in a safe, predictable way is extremely challenging.

It will be a very long time before we have AI plumbers, chefs, or farmers. But AI medical coders? Or KYC analysts? Or paralegals? These are right around the corner, and it will behoove every business to automate this type of rote, manual work. Not only will it save enormous amounts of money, it will free your employees from some of the most thankless, unfulfilling work. This form of AI is likely to replace many jobs (we don’t do anyone any favors by pretending it won’t), but they are jobs that we will not miss.

AI as Digital Experience

A third use case is using AI to create or augment digital experiences. This is the use case I focus on in my work at Yext - helping businesses create AI-powered digital experiences. I use “digital experience” as a broad term to refer to any touchpoint between a business and its end users. In this use case, AI is a medium through which the user interacts with a business or uses its services.

This use case is already well-established. Most people have already been interacting with AI-powered digital experiences for years now. We use an AI-powered search engine every time we interact with Google. We use AI-powered recommendation engines to browse movies, songs, and products on Netflix, Spotify, and Amazon. We use AI-powered conversational agents like Siri and Alexa to look up information and perform tasks. AI is increasingly the medium through which we interact with digital businesses.

When it comes to AI-powered digital experiences, we usually aren’t trying to replace a human but rather provide an experience that fundamentally can only be provided by software and AI. The human’s role - to the extent that there is one - is more so to monitor the AI’s behavior and curate the content that it has access to.

AI-powered digital experiences have existed for nearly a decade now, but generative AI has the potential to enhance them significantly. It’s already enhancing search engines like Bing and Google, who have begun using LLMs within their search engines to allow users to have detailed conversations about search results.

I would be shocked if Siri and Alexa didn’t introduce similar updates this year. These improvements will only make these products more integral to our lives, so it’s important to use products like Yext Listings to manage your company’s presence on these platforms and search engines.

It’s also possible that software platforms will start generating media content on the fly. What if Spotify could not only recommend music for you, but also generate new music based on your preferences? I think this is at least several years off, but it’s not impossible to imagine.

While building AI-powered digital experiences has historically only been possible for the largest, most advanced tech companies, advances in AI are making it possible - and in fact necessary - for average companies to deliver these experiences to their customers as well.

Increasingly, consumers will expect every business they interact with to have personalized recommendations, an intelligent search engine, and a helpful virtual assistant. This is what we’re focused on at Yext: making it easy for any business to harness this technology to improve the digital experience for their customers.

This takes many forms, such as…

A high-tech company providing a virtual assistant that can help users navigate their documentation and answer complex questions about it. (A lot like the bot you’ll find on Yext Hitchhikers.)
A retailer providing personalized recommendations for complementary products, as well as fast semantic search to help users find the precise product they’re looking for.
A health system offering a virtual assistant that can help customers identify the right doctor to treat their symptoms and get an appointment scheduled.

As AI gets more powerful and widespread, consumer’s expectations will rise, and they will no longer tolerate clunky, outdated websites and chat bots. They will expect to be served relevant, dynamic content and to get instant, AI-generated answers to their questions. If they don’t, they’ll take their business elsewhere.

AI as Prediction and Forecasting

The final use case is a relatively unsexy one that you won’t read as much about in the press, but is extremely valuable: AI can be used to predict the future. Astute readers will point out that all AI is predictive; LLMs are predicting the next word in a sequence, diffusion models are predicting the ideal set of pixels for an image, recommendation engines are predicting whether a user will like a product.

In this case I’m referring specifically to predicting business outcomes in the real world. For example:

Predicting consumer demand for a specific product based on historical trends
Predicting what the stock market will do tomorrow based on a variety of factors
Predicting how likely a person is to repay a loan based on their payment history
Predicting the weather tomorrow based on IOT data

You might think of these things more so as “statistics” than “machine learning” or “artificial intelligence”, but, as it turns out, those are all the same thing! AI and machine learning are statistics, applied at an extraordinary scale and to problems that we don’t traditionally think of as involving numbers.

This use case is the oldest of the four. We have been using math and statistics to predict the future for many decades now. What’s different today is that the algorithms can handle much more data than they previously could.

For example, WeChat - the Chinese super app that spans social, commerce and payments - is rumored to predict users’ creditworthiness based on hundreds of thousands of seemingly unimportant data points that WeChat collects about its users, like how much battery their phone has left and how many different people they text on average.

Historically, we would feed predictive models only a carefully curated set of signals or “features” that were known to have strong statistical relevance. (For example, if you wanted to predict the price of a house, you’d want to use square footage as a feature.) We might use ten or twenty of these data points to train a simple predictive model.

But, over the past ten years, deep learning and large neural networks, combined with other innovations in data processing, allow us to train much larger models with many more features on massive datasets. These models are able to learn complex relationships and representations that would have been impossible to learn with more primitive statistical methods.

Here once again the whole point of the model is to outperform a human, and usually these models do so handily. In fact, these models excel the most in the exact situations where human intuition fails. If you asked the average person to predict movements in commodity markets based on historical time series data, most surely couldn’t. This is a job for an algorithm. An algorithm can process far more data and observe statistical patterns that no human could.

This use case is probably the most unaffected by generative AI. In general, generative AI does not move the needle very much on our ability to produce statistical forecasts. For the most part, older ML methods like recurrent neural networks and even transformers work better here. But there are some exceptions - for example Deepmind recently released Graphcast, an algorithm that predicts the weather with state-of-the-art accuracy using generative AI. Perhaps generative AI will revolutionize this field as well.

The use cases we’ve discussed are by no means exhaustive, and many applications don’t fit neatly into a single category. (For example, I’m not sure where the facial recognition software on your iPhone fits in.) But still, this framework hopefully gives you a start for thinking about AI applications for your business.

AI can be used to assist creative knowledge workers such as programmers, lawyers, doctors, teachers, or screenwriters. You should adopt this sooner rather than later because there’s very little risk and, speaking from experience, AI can be extremely helpful.
AI can be used to automate entire business processes, wholesale, but this is riskier and requires high accuracy and additional monitoring/safeguards. But often the risk is worth taking, and you should compare AI to the status quo, not to perfection.
AI can be used to create immersive, helpful digital experiences. The human’s job here is to curate content and tweak the algorithm’s behavior. This is relatively low risk, and consumers will increasingly demand AI-powered experiences of all companies, not just Google and Netflix.
AI can be used to forecast the future based on historical trends and other data. This is a valuable application but relatively little has changed in this arena.

In reality, both the risk and the reward of all these applications has been overstated lately. It will take a very long time to retrofit existing business processes and digital experiences with new AI technologies, and I expect that when 2023 comes to a close we may look back and feel disappointed at how little changed, compared to the AI revolution that was promised.

But as Bill Gates famously said, “Most people overestimate what they can do in one year and underestimate what they can do in ten years,” and while we will assuredly be disappointed by what AI accomplishes this year, I think we might be utterly floored by what it accomplishes this decade.