AI & Industrial Planning : The impact beyond the trend

Artificial intelligence is now positioning itself as a major and global transformative force, a reality confirmed by data and recent advances. This article focuses on the impact of AI in the specific field of industrial planning, highlighting its increasing integration into decision-making processes. We propose a simple approach to classify the uses of AI in planning, based on the level of delegation granted by decision-makers, while illustrating these concepts with concrete examples. We also emphasize the crucial importance of the explainability of AI models, a relevant issue at all levels of delegation. In addition, we outline PlaniSense's commitment to developing adaptable and high-performance AI solutions to meet the changing needs of our customers in the supply chain sector.

AI: a fundamental trend and not just a fad

The question could’ve still legitimately been asked ten years ago, but it is now clear that the ingredients are there to affirm that AI is a real underlying trend with a global dimension. Four ingredients are there to remind us of this. 

• The market: The AI market has grown steadily in recent years, currently exceeding $300 billion, with the prospect of doubling in the next five years ^[1].

• Knowledge: AI research has accelerated dramatically, as evidenced by the doubling of the number of scientific publications in this field over the past decade, far outpacing the growth seen in other fields ^[2].

• Data: the fuel of AI algorithms, data has never been so abundant, and its volume is only increasing. To give you an idea, the most advanced current generative AI models are fed by datasets that exceed 100 trillion observations ^[3].

• Computing capacities: the supercomputers used to train large models are capable of processing more than a billion gigaflops, or 10 to 18 operations per second ^[4]. At the time of writing, NVIDIA announced the release of its Blackwell B200 chip with 25 times higher inference performance than previous-generation GPUs ^[5].

In the supply chain space, this underlying trend is evident in key areas such as demand forecasting. For more than 40 years, M-Competitions have been pitting the best forecasting models against each other on the basis of extremely rigorous reliability criteria. Until the early 2000s, smoothing methods were unbeatable despite the emergence of sophisticated autoregressive methods [6]. It was only in 2020, on the fourth edition of the M-Competitions, that learning methods began to appear in the rankings without outclassing the so-called "classic" methods [7].  In 2022, barely two years later, the situation is changing radically in favor of learning approaches, and particularly gradient boosting techniques, which clearly outperform traditional methods and all their combinations [8].

It is therefore no longer time to ask whether AI is a fad, but rather to anticipate the way in which it will change certain practices or even bring out new ones in the field of industrial planning, which is of particular interest to us.

AI and Decision-Making in Planning 

In the context of decision support, AI uses algorithms and models to analyze large amounts of data, extract relevant information, detect underlying trends or patterns, and make recommendations or even propose solutions to help decision-makers.

From this point of view, AI encompasses a multitude of methods, ranging from traditional expert systems to newer and more sophisticated machine learning techniques, including combinatorial optimization. AI thus offers a panoply of tools and possibilities to effectively address complex decision-making challenges in various industrial and organizational contexts. 

In the rest of this article, we present a simple approach to classify the uses of AI in planning, based on the degree of delegation granted by the decision-maker or, in other words, the level of autonomy granted to AI in the decision-making process. 

To put it simply, there are three levels of delegation: assistance, recommendation and finally autonomous decision-making. They each correspond to a type of use that can be observed in practice, which we will try to illustrate with a few examples. 

Assistance

This is the minimum degree of delegation. AI collects, selects, and formats information that is useful for decision-making. Customizable dynamic dashboards or intelligent conversational agents are forms of assistance that we develop at PlaniSense

The challenges associated with this first degree of delegation are the contextualization and readability of the analyses.

• Contextualization: Providing relevant analytical elements for informed decision-making in a particular context
• Readability: Facilitatethe understanding of complex data with configurable visualization tools.

Recommendation

This is the intermediate degree of delegation.  The AI submits one or more proposals to the decision-maker who remains sovereign in the final choice even if he may be influenced by the options recommended to him.  In planning, the recommendation may concern a more or less extensive scope (a network, a factory, a workshop, an item, etc.) and/or a more or less targeted aspect (a parameter, a quantity to be produced, a production order to be moved, a capacity to be adjusted, etc.). The automatic generation of optimized schedules is one of the features of PlaniSense that is clearly part of a recommendation logic. The issues associated with this second degree of delegation are the recommendations' effectiveness and adaptability.

• Effectiveness: Provide recommendations that are achievable and effective in terms of the objectives pursued. 
• Adaptability: Facilitate the adjustment of proposals and measure the consequences.

Autonomous decision-making

This is the maximum degree of delegation. AI makes decisions without prior validation by a human operator. Recommender systems, when they are endowed with explicit criteria for choice, can be transformed into autonomous decision-making systems. In planning, autonomous decision-making is usually limited to simple and/or non-critical cases. A typical example of autonomous decision-making is demand forecasting for items with high predictability and low impact.

Looking at the rapid progress made in AI in recent months, it is reasonable to expect autonomous forms of decision-making to gain ground. The issues associated with this ultimate degree of delegation are explainability and sovereignty. 

• Explainability: Being able to understand and interpret a posteriori the reasons for a choice made by AI.
• Sovereignty: Maintain control over the scope of delegation and be able to adjust it at any time to regain control if necessary.

The dilemma of explainability and the requirement for efficiency 

We have introduced the issue of explainability as a major challenge related to autonomous decision-making. While this challenge is more pronounced in this specific context, it is just as crucial for uses where the level of delegation is lower.  

Explainability refers to our ability to understand and interpret the decisions made by an AI model. This is an essential element for the adoption of AI by users who must be able to trust before delegating decision-making. 

Explainability is an intrinsic feature of AI techniques used to build models. Generally, the more complex a technique is, the more likely it is to become a black box and less explainable it is. 

The dilemma of explainability is that we need it most precisely where it tends to be lacking. This is because the more sophisticated AI models are, the more effective they are. This increased efficiency paves the way for an extension of the scope and level of delegation in decision-making. However, as this area and level of delegation expands, the need for explainability becomes more and more pressing. The problem is that, generally, the more sophisticated the models, the less explainable they are. 

In this context, the principle of parsimony remains an important benchmark for reconciling the effectiveness of AI in decision support and its acceptance by users. When multiple models offer comparable efficiency, it is best to choose the least complex ones. These less complex models are usually more explainable, making them more likely to be accepted by users. 

Conclusion

AI is now firmly entrenched in the industrial planning landscape, paving the way for significant transformations in everyday uses. However, to ensure its long-term adoption, it is imperative to address the challenges related to the explainability of models, an essential prerequisite for extending the scope of autonomous decision-making. In order to maximize the benefits of AI in industrial planning, it is paramount to carefully choose the appropriate use cases, choose the technologies best suited to the level of delegation envisaged, and ensure that AI models are explainable enough to inspire confidence in users. 

At PlaniSense, we attach high importance to three fundamental principles in the development of our AI solutions for industrial planning. Firstly, we strive to modulate the delegation level according to the specific needs of each use case, while maintaining a rigorous control over the degree of delegation granted. Secondly, we strive to avoid 'black boxes', ensuring that our models are transparent and understandable, allowing users to understand the underlying decision-making processes. Finally, our goal is to provide efficient algorithms, designed to ensure operational excellence, optimizing planning processes and improving decision-making. By combining these three principles, we are committed to offering reliable, adaptable and high-performance AI solutions to meet the evolving needs of our customers in the supply chain space.

Sources

[1] Statista Market Insights[2] Center for Security and Emerging Technology[3] epochai.org[4] Top500 supercomputer database[5] Nvidia[6] The M3-Competition: results, conclusions and implications[7] The M4 Competition: 100,000 time series and 61 forecasting methods[8] M5 accuracy competition: Results, findings, and conclusion