Artificial Intelligence 101: Core functionalities and foundational use cases for digital publishers

Large Language Models (LLMs) like ChatGPT, Bard, Alpaca and Flan-T5 have shown the digital content and marketing industries that creative processes now fall within the category of business tasks which can be supported by Artificial Intelligence (AI) systems. But there is already a range of use cases which do not require techniques from the forefront of AI research to be impactful for digital subscriptions businesses. These businesses should consider building AI solutions to support decision-making, automate repetitive processes and improve customer interactions, and embedding these within an AI capability which will grow to require a unique set of people, processes and tools as it scales.

AI describes algorithms that can perform tasks which would typically require human intelligence, such as recognising speech or images, writing text, or playing games like chess and Go.

In practice, “AI” commonly refers to “Machine Learning” (ML), which are algorithms that learn from data without being explicitly programmed. This is fundamentally a different approach from codifying and applying domain expertise to a problem. Instead, algorithms are trained on large datasets, iteratively attempting the problem and observing the result. Based on these results, the parameters which configure the algorithm's output are tuned to perform well on the training data. In production, the algorithm is then applied to new, unseen input data.

Supervised learning is a type of ML where the model is trained on labelled data, which means the input data is paired with a known, correct output. The goal is to learn a mapping function from input to output that can be used to make predictions on new data. Some examples of supervised learning use cases include fraud detection and customer churn prediction. In contrast, unsupervised learning is a type of ML where the model is trained on unlabeled data, which means the input data is not paired with any output. The goal is to learn patterns and structures, such as clusters of data points or ways to compress the data.

As models get bigger and more complex, they require more computational power to train and run. This is increasingly accessible due to both cloud computing and consumer hardware. Similarly, big data sources and processing techniques are increasingly prevalent, providing more raw material for training ML algorithms. As a result, we are seeing more advanced behaviours emerge from AI models than ever before. For example, while it is true that at its core GPT learns correlations between input words and output words based on sequences present in real world data, because the model has such a high amount of parameters with which to optimise its performance it is possible to interpret its behaviour as first generating internal algorithms which themselves then generate the training data. In other words, GPT-4 may, in effect, be taking an input sequence, attenuating to the parameters which are activated and effectively implementing a relevant generator algorithm (for example a set of internal neurons which work better on code instead of natural language), and then producing its output.

The essential abilities of AI models are:

• Processing unstructured data - transforming unstructured data (e.g. images, text, audio) into a numerical format allows it to be further analysed and visualised using more traditional techniques
• Regression - this involves predicting a continuous outcome variable based on one or more predictor variables. For example, propensity models can predict the numerical likelihood of a customer taking a specific action, such as making a purchase or subscribing to a service.
• Classification - involves predicting a categorical outcome variable based on one or more predictor variables. For example, sentiment analysis can classify customer feedback into different categories like positive, negative and neutral.
• Clustering - involves grouping similar data points into clusters based on their similarities and differences. For example, this could be used to create a novel customer segmentation, or to detect anomalies.
• Sequence generation - an emerging area of functionality, so-called Generative AI models are able to generate sequences of text in various forms (natural language, code, even music).

• Machine learning - clustering, regression and classification
• Machine Learning - model
• Machine Learning - embeddings

It is useful to think of AI use cases in terms of the business task which they help to accomplish. A good way to categorise AI use cases is in three buckets, each with their own approaches to design and implementation:

• Insight and analytics - supporting decision-making by producing deep or novel insights (e.g. propensity models, segmentation algorithms, summarisation tools)
• Engagement and personalisation - augmenting customer experiences by adapting based on data (e.g. message personalisation, chatbots, best-next-action models)
• Automation - automating repetitive business processes which involve judgement calls and/or unstructured processing (e.g. outcome triage bots, metadata tagging, text drafting)
  
  

As a team scales their usage of AI technology, they will begin to have different requirements from other data disciplines. They will need different technology, working with big data, pre-trained models, embeddings, and feature processing pipelines; and different operations, including deployment systems, model performance monitoring, explainability and interpretability frameworks, and dedicated risk management. This is a journey with multiple steps, but companies can already begin to:

• Develop a strategy and roadmap towards highly impactful use cases
• Experiment with proof-of-concept models and foundational capabilities
• Nurture existing talent and data management, analytics and privacy processes