What actually is a good AI use-case?
Real talk, folks. Generative artificial intelligence (which I’ll endearingly call “gen AI”) has fantastic potential, but it is NOT a one-size-fits-all solution.
My goal for this two-part series is to save you from wasting resources on building out gen AI solutions for problems where it simply isn’t the best tool for the job or where it’s over-egging the pudding.
Newer is not always better. Sometimes a tried-and-true solution is the way to go. In the sciences we call this the principle of parsimony, shorthand for the simplest solution is usually the best one. When it comes to AI solutions, simpler equals traditional machine learning (ML).
“The simplest solution is usually the best one.” [paraphrased]— William of Ockham (c. 1287–1347)
Now picture some C-suite execs (and if you ARE a starry-eyed exec look away now…). Imagine those execs are so dazzled by stories of gen AI successes they believe it can magic up a solution to any problem. Swish and flick!
As data and AI strategists, our role then becomes to equip our colleagues and our teams with the knowledge to choose the right tool, not just the flashiest.
Choosing the Right AI Tool
Let’s take a closer look at these families of AI tasks. Broadly, they include our main contenders: generative AI (blue) and traditional ML (green).
For the illustration below, I also split traditional ML into three (green) sub-families (L-R): Prescriptive ML, Descriptive ML, and Predictive ML. These are all part of the traditional ML family.
Good AI use-cases fall into several families. (Image by author)
The power of traditional machine learning
Traditional machine learning algorithms offer a time-tested toolkit for a variety of automation tasks. From classification and regression to clustering and dimensionality reduction, they can be applied broadly across many domains, from manufacturing and scientific discovery to finance and healthcare.
A particular advantage of traditional machine learning algorithms is that they’re often interpretable. By which I mean we can look inside the black box to understand their inner workings, the logic behind their predictions. This has the happy benefit of reassuring us that we know what they’re doing and why, well enough to trust their outputs and simplify our debugging and error analysis. Interpretability is especially crucial in domains like healthcare and finance where understanding the “why” behind a decision can be critical.
Compared to compute- and data-hungry deep learning models, traditional machine learning algorithms can typically be trained with less data and require fewer computational resources. This makes trad ML easier to use on your local laptop and more efficient for real-world deployment, especially when data availability is limited.
Data quality
The success of machine learning hangs on the quality and quantity of data you’re able to feed it in training. Data work is a bit like being a detective. To draw clear conclusions about a case, you need to be confident in the quality of your evidence aka data.
Just like a detective relies on good evidence, ML algorithms depend on the availability of plentiful, accurate data that they can mine for patterns relevant to your question. High-quality data ensures the patterns they learn are genuine and robust, which in turn leads to reliable and trustworthy predictions. The alternative? Garbage in → Garbage out.
ML algorithms analyse the (preferably high-quality) training data they’re given, then use that training to predict answers. (Image by author)
Good use-cases for traditional ML
Traditional machine learning algorithms have quietly and effectively been powering a huge array of real-world applications.
For example:
Customer segmentation: For creating targeted marketing campaigns. Traditional ML algorithms can analyse customer data (think: purchase history, demographics) to identify distinct customer groups. This allows businesses to tailor marketing campaigns for each segment, maximising effectiveness.
Predictive maintenance in industrial settings: By analysing sensor data from equipment on the factory floor, ML models can predict potential failures before they occur. This enables preventative maintenance, which in turn reduces downtime and saves a company’s operations from unnecessary expenditures.
Fraud detection and risk management in finance: Trad ML algorithms can analyse financial transactions to identify patterns that indicate fraud. This helps financial institutions protect their customers and manage risk.
Spam filtering in email: ML algorithms can analyse email content and identify spam messages with high accuracy, keeping your inbox clean. What a relief!
Medical diagnosis support: Traditional ML can analyse medical scans and patient data to assist doctors in diagnosis. This can improve the accuracy of diagnoses and help diagnosticians manage their time more efficiently in healthcare.
Recommendation systems: Used extensively in e-commerce. Recommenders predict a user’s preferences, interests, or needs based on their purchase history, browsing behavior, and other relevant data. Personalised recommendations help users discover new content and keep them engaged with the platform.
The power of generative AI
Unlike traditional ML techniques that are confined to working with existing data, generative AI has the remarkable ability to synthesise entirely new information.
Generative AI goes beyond the pattern recognition and analysis of traditional ML. Pre-trained gen AI models (like GPT-4o, Gemini, Claude and others) are distinguished from trad ML in that they create entirely new content, be it text, images, computer code, or even music. It’s true they had to learn how to create content by ingesting vast catalogues of training data, but that training process is extensive and computationally expensive. For any use-cases building on the pre-trained models mentioned above, that process is abstracted away by the tech giants who own them (OpenAI, Google, Anthropic, etc.).
Generative AI models have learned the underlying patterns and relationships from their training so well, that they’ve gained the capacity to create new and original outputs resembling the data they learned from.
Gen AI’s “creativity” opens a whole new range of applications that are impossible for traditional ML.
Creating data
We’ve started to glimpse a world where generative AI can conjure up photorealistic portraits of people who never existed, complete with intricate details and lifelike expressions. This isn’t some distant sci-fi future, or if it is, the future is now. Generative models — available to anyone with an internet connection — are achieving this feat in everyday ways, creating stunning visuals for everything from architectural concepts to video game characters.
Generative AI models have been trained to know a lot about the world we live in. Their training (+ probabilities) allow them to create new content in answer to your prompt. (Image by author)
When it comes to text-based creation, generative AI can craft compelling narratives, compose music lyrics for different genres, or even generate realistic news articles that mimic various writing styles.
This ability to create entirely new data extends far beyond human-consumable media. For programmers, gen AI is catalysing futuristic development in every corner of the internet.
Generative AI can draft new lines of code, automatically complete complex functions, or even write entire scripts from scratch. Think about the boost to developer productivity. CAUTION: be sure a skilled human developer is verifying the code quality!
For the data scientists out there, generative AI can be harnessed to manufacture synthetic datasets specifically designed for machine learning models.
Need a vast amount of data to train an image recognition model? Gen AI can create realistic images with specific attributes, allowing you to build a robust and diverse training dataset without ever needing a physical camera.
You want to run a sentiment analysis? Gen AI can craft textual data for you, rich with a range of emotions and opinions to augment your model’s understanding of language nuances.
Even the foundational building blocks of machine learning, dataframes — tabular structures you might know as spreadsheets — can be generated by AI. Imagine filling in missing information within a dataset or creating entirely new data points based on existing trends. This ability to generate dataframes populated or supplemented with realistic but fictitious information unlocks a new level of flexibility for machine learning endeavours.
CAUTION: Be sure a skilled human is verifying the code / data / output quality
Generative AI’s impact extends far beyond these few examples (but I had to stop somewhere!). It’s fundamentally reshaping the way we create and interact with data, opening doors to entirely new applications and experiences for public consumption.
Good use-cases for generative AI
Let’s break down some key use-cases where the unique ability of generative AI to create entirely new data prove to be most advantageous.
Customer segmentation: I listed segmentation above, but Generative AI can go a step further than trad ML and create entirely new customer profiles based on real data, allowing businesses to explore hypothetical customer segments and test marketing strategies before full-scale implementation.
Image generation: Product designers can use gen AI to rapidly create and iterate on different product concepts. By inputting specific requirements or constraints, designers can rapidly generate a range of prototypes, helping them to identify the most effective and appealing designs.
Drug discovery: Generative AI can create entirely new molecules that haven’t been previously documented. This allows researchers to virtually explore vast chemical spaces, accelerating the discovery of new drugs with desired properties.
Automated question-answering: Think of this as a system for answering specific questions, not just by pointing you to relevant documents but actually summarising an answer directly from the source. Generative AI can be fine-tuned on a collection of documents to understand the context and the factual content it contains. This allows users to ask detailed questions and receive concise, informative answers, saving them the time and effort of wading through mountains of text themselves.
Generating realistic synthetic data: As mentioned, generative AI can create entirely new data points that mimic real-world information. This is incredibly valuable in situations where acquiring real data is difficult, expensive, or ethically sensitive. For instance, in the medical field, generative AI can create synthetic patient data for training healthcare models without compromising patient privacy.
Data Augmentation for Imbalanced Datasets: In machine learning, imbalanced datasets with a scarcity of certain data points can hinder model performance. Generative AI can create synthetic data points to artificially balance the dataset, leading to more robust and generalisable models.
Remember, generative AI is a powerful tool, but it’s not a one-size-fits-all solution. As data and AI leaders, it’s our responsibility to guide our colleagues towards effective AI implementations.
By understanding the strengths and limitations of both traditional machine learning and generative AI, and also when/where to use them, you can you can equip your teams with the knowledge to make informed decisions and choose the right approach for your specific circumstances.
The next time a colleague approaches you with an AI idea, don’t hesitate to ask probing questions and offer guidance.
Don’t be swayed by the generative AI hype. Take a step back, assess the problem.
The principle of parsimony and your knowledge of AI families will enable you to identify an approach well-suited to your challenge. That’s where you’ll deliver real value.
Looking ahead
Now you’re bursting with ideas for good AI use-cases, and you’re trying to decide whether generative AI is right for you? Am I right?
Or perhaps your AI-enthusiast colleague has convinced to consider generative AI for their use-case?
Either way, you’ll want to put that idea to the test before adding a gen AI solution into your product roadmap.
Watch here for Part II in this Good AI Use Case series, where I’ll outline a simple two-step test for selecting a good AI use-case.
Comments