INESC TEC, Faculty of Engineering of the University of Porto & LTPlabs
INESC TEC & Faculty of Engineering of the University of Porto
The popularity of Artificial Intelligence (AI) has been growing rapidly, due to multiple scientific and technological breakthroughs, as well as their potential application to several areas. The most recent disruption was in generative AI, with large language models (e.g., ChatGPT1) accumulating vast amounts of knowledge, and using it to generate text and images in response to human prompts. Those models can be used in different applications, from information retrieval to content generation. However, ChatGPT does not solve all the problems in the world, like those that came before did not. From our experience in interacting with managers across different business sectors, all the glare that came with the emergence of ChatGPT is often undermining the basics of employing AI to improve and transform business processes.
In this article, we review the five main pillars that managers ought to know to be able to convert the promise of AI into reality. These pillars are linked to the need to properly match the existing tools to the tasks to be tackled – this attention is particularly relevant for prescriptive tasks (Pillar #1); the importance of combining different AI approaches to fully address the complex challenges of organisations (Pillar #2); the relevance of focusing on AI methods that are explainable to promote trustworthiness and collaboration (Pillar #3); the possibilities that different human-machine interaction modes open (Pillar #4); and the fundamental activity of ensuring basic AI knowledge across the whole organisation (Pillar #5).
AI has been increasing in companies, supporting decision-making at different levels, especially with descriptive and predictive tasks. Descriptive methods, such as clustering and association rules, allow for instance to segment customers or detect consumer patterns. Predictive methods, such as those used in supervised learning, as the name suggests, can provide predictions, such as sales forecasts or customer churn. However, those methods should not be directly applied to prescriptive problems – problems in which the decision to be made is the core issue to be tackled, such as deciding the best route for a vehicle or the assortment to keep in a store. What we have often witnessed in practice is that managers want to use descriptive and predictive tools for problems that require a different approach. As Abraham Maslow once said, “if the only tool you have is a hammer, you tend to see every problem as a nail”.
Prescriptive problems require a different family of AI methods: the likes of reinforcement learning. These methods were in the news almost ten years ago, when AlphaGo2 was able to defeat the world champion in the ancient game of Go. These are the same methods used in autonomous driving and robot navigation, which are still emerging, due to the complexity and issues involved. Nevertheless, there is an enormous untapped potential for those methods to step into many prescriptive applications, such as order allocation in online retail3, dynamic scheduling in manufacturing, and dynamic routing4 in internal or external logistics.
With the new waves of AI methods emerging from time to time (for example, some years ago it was machine learning, now we have generative AI), it is easy (and lazy!) to think about these technologies as replacements. An equivalent to having the new revamped version of a car substituting its predecessors. Although tempting, this analogy is not accurate. Of course, there are methodological evolutions that render previous algorithms obsolete, but often what we end up having are new tools that can be used in concert to solve ever more complex problems that emerge across society. Consequently, instead of thinking of using the car versions analogy, one can use a LEGO analogy in which new blocks can be added to build more interesting creations.
Let us take the example of large language models – known for their natural language capabilities – which can be combined with more traditional predictive (machine learning) and prescriptive (optimisation) algorithms to unlock some of the current challenges with these methodologies5. We see opportunities for generative AI to tackle challenges within advanced analytics throughout the development and deployment phases. Large language models can be particularly useful in helping users incorporate unstructured data sources into analyses, translate business problems into analytical models, and understand and explain models’ results. This last potential synergy between large language models and advanced analytics is connected to our next pillar.
AI models, although performing well in certain tasks, tend to be hard to trust, due to their complexity and, hence, lack of interpretability. This is a major issue hindering adoption in many critical applications, in areas like healthcare, finance, and operations. Explainable AI (XAI) is a growing field, proliferating with multiple research streams. Some propose to use local explanators on top of the AI black-box models (e.g., a decision tree mimicking a neural network). An alternative is to involve human intelligence in the discovery process, resulting in AI and humans working together, in a human-centred, ‘guided empirical’ learning process. This is made possible by employing ‘explainable-by-design’ symbolic models and learning algorithms6.
Symbolic models can be learned by Genetic Programming algorithms. The idea is often to learn a compact model, whose performance does not result from its complexity and fine-tuning of several constants, but rather from learning the structure of the model, by freely combining the problem features with user-defined operators (e.g., arithmetic, logical, etc.). The final model is compact and inspectable and thus does not require an explanator; it is explainable-by-design. In some cases, such a model can have an even higher performance, as it might generalise better. In other cases, the performance may not reach the same level as that of a black-box. In that case, it can be beneficial to keep the black-box and add an explanator on top, to obtain insights into its inner workings.
Having humans involved in the discovery process of decision models implies a deep human-machine interaction. However, that level of interaction is not always feasible, e.g., in settings where decision speed is paramount. Also, humans can interact in different ways, such as at the decision level. Decision support systems are exactly that: systems based on advanced methods, suggesting decisions that may or may not be adopted by decision-makers. In some cases, it is important that every decision is evaluated by the human decision-maker, such as when suggesting medical procedures. In other cases, such as fraud detection in credit card transactions, having human agents evaluating all of the thousands of daily predictions is not economically viable.
Despite these constraints, and for companies to get the most out of AI, it is fundamental to grasp that there is no silver-bullet type of solution to set an ideal human-machine interaction. On the contrary, research has identified that companies that are able to be more versatile in the configuration of these interactions are the ones poised to reap more benefits7. Sometimes AI decides and implements; other times AI decides and the human implements; but there are variants, such as AI recommends and human decides, or AI generates insights that the human uses in a decision process, or even the human generates solutions, and the AI evaluates. Figuring out the right model can only be (probably) done by humans!
To be able to manoeuvre a company along the four pillars described above and get the most out of AI, it is key that the organization, from top to bottom, understands the basics of these technologies. Day in and day out we have interactions with managers from different sectors that demonstrate how far from understanding the basics of AI they are. This is, of course, a major roadblock to making good decisions about the use of these technologies, to either improve current business processes or find opportunities to revamp existing business models. As we have heard someone rightfully saying: “If you throw technology (AI) into an ‘outdated organisation’, the only thing you get is an ‘expensive outdated organisation’”.
To revert this situation, companies must invest heavily in education initiatives that go up and down the ranks. Interestingly enough, according to Gartner , this year, 40% of all organisations will offer or sponsor specialised data science education to accelerate upskilling initiatives. This is a 35 percentage points jump from what we have witnessed in 2021. This must be a continuous effort at the corporate and individual levels, as the pace of evolution of these technologies shows no signs of slowing down.