Beyond the Hype: The Enduring Value of Predictive AI in a GenAI World

Generative AI has revolutionized the tech landscape since OpenAI's ChatGPT burst onto the scene in late 2022. In response, companies, across sectors, are swiftly pivoting their strategies and redirecting significant investments to capitalize on this revolutionary technology. In a recent survey of around 1,500 executives published by the Boston Consulting Group, about 89% ranked AI and generative AI as one of their top-three tech priorities for 2024, with 51% of them putting it at the top of their list.

This shift towards generative AI is reshaping the broader artificial intelligence field. While generative AI builds on the principles of machine learning used in predictive AI, its ability to create content and generate new data instances has captured widespread attention. Companies are now allocating significant resources to incorporate generative AI into their businesses and products. Consequently, projects using predictive AI are being overshadowed. According to McKinsey's technology trends outlook for 2024, GenAI-related job postings have increased by 111% from 2022 to 2023, whereas job postings in applied AI, which encompasses predictive AI, have decreased by 29% during the same period. Despite this widespread attention and investment from companies, generative AI technology has yet to live up to its lofty expectations. Many organizations are struggling to demonstrate clear ROI from their generative AI projects. Predictive AI, on the other hand, has been a cornerstone of artificial intelligence in the pre-GenAI era and continues to deliver substantial value across myriad use cases and industries.

\In this essay, we will explore why predictive AI still remains a powerhouse and continues to deliver value for businesses in today's generative AI revolution.

Understanding the Distinction: Predictive AI vs. Generative AI

Artificial intelligence is often used as a catch-all term to mean any type of smart machine. Generative AI and predictive AI, the two subsets of artificial intelligence, have unique capabilities appropriate for different use cases. Predictive models excel at learning from domain-specific data to classify information or predict future outcomes based on historical patterns. For instance, Walmart uses an ML-driven inventory management system that leverages proprietary sales data, online shopping patterns, weather forecasts, and local demographics to predict consumer demand and strategically place items across its supply chain.  Generative AI models expand these capabilities to create summaries, uncover complex hidden correlations, or generate new content—like text, images, or videos—that reflect the style and patterns within the training data. They are built to converse with or follow instructions from humans. As such, they are trained on vast amounts of diverse data so that they can capture a broad understanding of language, concepts, and human-like reasoning.

Technical Considerations and Constraints

Generative models tend to follow scaling laws – as model size and training data increase in scale, performance improves smoothly and predictably. Theoretically speaking, with more investments in compute power and data, one can make these models increasingly powerful. For predictive models, however, there is a clear limit to how well things like future sales can be predicted, and investments in more complicated models and compute power beyond a certain point likely lead to diminishing returns. As such, custom-trained predictive models are many orders of magnitude smaller in size, rarely exceeding a million parameters as opposed to billions in their generative counterparts. The cost of training these models from scratch is exorbitantly high – it is estimated that OpenAI spent over $100 million to train their latest generation of large language model, GPT-4. Providers of general-purpose LLMs such as OpenAI, Anthropic, and Google provide fine-tuning APIs to adapt the general-purpose language models to perform better on specialized tasks by training them on domain-specific data. However, given the nature of these models, they can primarily be fine-tuned only on language-specific tasks and not on tasks that consume structured data, which predictive AI excels at.

The Explainability Factor

Besides training costs, explainability – the ability to understand and interpret how an AI model arrives at its outputs –  is a key factor that differentiates predictive and generative models. Predictive models are generally more transparent and interpretable, while generative models often function as black-boxes with less explainable decision-making processes. Explainability is a prerequisite for certain use-cases, such as in lending and credit decisioning where the Fair Credit Reporting Act (FCRA) requires lenders to be able to explain decisions taken by AI models. While research being undertaken by companies like Anthropic and OpenAI to understand inner workings of their flagship models, this is still in nascent stages and unusable in the use cases such as the aforementioned. This makes predictive models easier to implement, specifically in industries and use-cases that are highly regulated.

The Rush to Adopt GenAI: Challenges and Pitfalls

Companies are scrambling to integrate generative AI into their products and services – 75% of company leaders report feeling increased pressure to incorporate generative AI into their business strategy. However, only 1 in 4 companies have successfully launched generative AI initiatives in the past year. To keep up with the pressure, many companies are resorting to 'solutionism', a phenomenon prevalent in the tech industry for the tendency to adopt a trendy technology and then seek the problems it might solve. As an example, Lattice, a popular people management software, announced a feature that would enable organizations to treat AI entities as digital workers with official employee records. The initiative aimed to integrate generative AI-based workers into the organizational framework alongside human employees, involving processes such as onboarding, training, goal setting, and performance evaluation. However, the plan faced significant backlash and was suspended just a few days later, with critics concerned about the implications of treating AI as equivalent to human employees, arguing that it could undermine the value of human work and disrupt traditional workforce dynamics. In comparison, predictive AI-based use cases are more developed and established in many business contexts and have demonstrated clear ROI.

A Balanced Approach to AI Integration

Generative AI undoubtedly has its own set of advantages. Its ability to generate creative content, such as text, images, and even music, opens new avenues for innovation in industries like entertainment, marketing, and design. It can power intelligent chatbots for customer service, generate marketing materials and code, and uncover insights from vast datasets. These capabilities are compelling, but also present a challenge: how to implement generative AI effectively within existing business structures. Despite the allure of these innovations, businesses should resist the urge to overhaul their entire product roadmap or strategy merely to keep up with trends. Instead, a more measured approach is advisable. Companies should start by identifying specific GenAI-related use cases that align with their existing product strategy and overall business goals. This targeted approach allows for strategic integration of generative AI where it can provide the most value.

Making Informed Technology Choices

In addition, business and product owners should collaborate closely with their data scientists and ML engineers to carefully evaluate whether a predictive or a generative model is more appropriate for their particular use case. It is crucial not to default to using a generative AI-based approach simply because it's the latest trend. Each technology has its strengths, and the choice should be guided by the specific needs and constraints of the business problem at hand. To aid in this decision-making process, resources are available. Google Cloud, for instance, provides a comprehensive guide to help practitioners identify when generative AI or predictive AI might be more appropriate for their business use case. Such tools can be invaluable in navigating the complex landscape of AI technologies and ensuring that implementations are both effective and aligned with business objectives.

The Power of Integration

Though both technologies have their own set of unique advantages, the true power of AI can be unlocked by integrating predictive and generative approaches. Predictive models can provide structured data insights that enhance the context for generative AI applications, leading to more informed outputs. A compelling example of this synergy is content moderation, a critical task for online platforms aiming to maintain safe and respectful environments. Predictive models, with their robust pattern recognition capabilities, can use metadata about the content and the user, such as user engagement patterns and historical violation records, to act as an initial layer of defense for handling large volumes of data. Meanwhile, generative models can delve deeper into the nuances of language and context, identifying subtle forms of inappropriate content that might evade simpler filters. By combining these approaches, platforms can achieve a more comprehensive and effective moderation system, ensuring both efficiency and accuracy.

In conclusion, the current frenzy surrounding generative AI, while exciting, should not overshadow the enduring value of predictive AI. Predictive AI continues to deliver tangible ROI across diverse sectors, offering explainability and a unique ability to incorporate deep domain expertise that generative AI currently struggles to match. Rather than blindly chasing the latest trend, businesses must adopt a discerning approach. By leveraging the strengths of each, organizations can unlock the true potential of AI, driving innovation and value. The future of AI is not about choosing sides, but about harnessing the power of both to build a more intelligent future.