This article is by Professor Enrico Zio, Scientific Director of Datrix Group and Full Professor at the Centre for Research on Risk and Crises (CRC) at Ecole de Mines, ParisTech, and at the Politecnico di Milano. He is the author and co-author of seven books and over 500 articles in international journals (with an H-index of 103 on Google Scholar), president and co-president of several international conferences, associate editor of several international journals, and a reviewer for more than 20.
In addition to holding academic roles in China, he is a board member of the Fondazione Politecnico di Milano and a Distinguished Lecturer for IEEE and Sigma Xi. In 2020, he received the Humboldt Research Award for his contributions to risk and reliability assessment. In 2023, he was elected a Fellow of the Asia-Pacific Artificial Intelligence Association as “the leading scientist with outstanding achievements in the area of reliability engineering and risk assessment.” In 2024, he was recognized with the Lifetime Achievement Award from the Society for Reliability and Safety for pioneering the application of AI and genetic algorithms to support risk and resilience assessment for complex engineering systems. Recently elected Fellow of the Industry Academy of the Artificial Intelligence Industry Academy.
Opportunities and risks of Industry 5.0. A particularly relevant theme today in the debate on artificial intelligence (AI) because it represents an evolution centered on human-machine interaction, emphasizing collaborative and human-centric AI.
Unlike Industry 4.0, which focused primarily on automation and efficiency, Industry 5.0 aims to leverage the potential of AI to enhance human creativity, personalize production, and make work more meaningful. Furthermore, Industry 5.0 promotes sustainability and resilience, crucial themes for addressing contemporary environmental and social challenges, emphasizing the coexistence of advanced technology and human values.
The context
Industry 5.0 refers to a business model characterized by strong human-machine cooperation and great sensitivity toward environmental sustainability themes, whose main objective is to add value to production through extreme product personalization and thus respond to the peculiar needs of consumers. This model represents the natural evolution of Industry 4.0, enabled by the development of increasingly advanced technologies, particularly in the sectors of AI, ICT, and robotics, which are leading to the creation and dissemination of increasingly powerful IoT devices and more advanced Cyber Physical Systems. In accordance with the European vision of supporting the development of Industry 5.0, the Italy 5.0 Transition Plan has been defined in Italy, which offers various subsidized investment opportunities along three directions:
- innovative technologies (robotics, IoT, 3D printing, cloud computing): improve productivity, optimize processes, and remain competitive in the global market;
- energy efficiency: ensure more sustainable production by reducing operating costs;
- sustainability: reduce the environmental impact of activities with renewable energy and eco-friendly production practices.
AI plays a fundamental role in achieving the plan’s objectives and this offers great prospects for scientific and technological development; however, this evolution also entails new challenges to the competitiveness of companies offering solutions in this area, with consequent new risks that must be estimated, evaluated, and properly managed. For this reason, it is important to identify some elements of competitiveness in proposing AI solutions for Industry 5.0, starting from the current technological context and also considering the risks associated with this type of commercial offering.
What is happening
Over the last 5 years, we have witnessed a proliferation of algorithm libraries, developed also with the direct intervention of the “over the tops” (Google, Meta, Microsoft, etc.). For example, TensorFlow and PyTorch, among the most popular software frameworks for AI development, are maintained and updated by Google and Meta, respectively. In many cases (above all, generative and multimodal AI), algorithms are made available on the market as ready-to-use models, pre-trained using huge computational and economic resources. The size of these models (such as the number of parameters they are based on) makes both their retraining and personalization complex from both a technical and economic point of view. This situation has, on the one hand, generated a widespread capacity to develop solutions based on these algorithms and models: all players in the AI market have roughly access to the same models and technologies, generally under the same economic conditions. On the other hand, it has made it difficult to clearly express the differentiating elements of each company when proposing AI-based solutions on the market, particularly for Industry 5.0. In fact, the AI market recognizes value not so much in the algorithm as in the ability to use appropriate models and run them on quality data, securely. A significant example is given by the explosion of Large Language Model (LLM) solutions, even for Industry 5.0 applications (such as expert systems for maintenance based on conversational chatbots trained on manuals or maintenance reports). There are several pre-trained models available for free (Alpaca, OpenLLama, ChatGPT3). The most high-performing ones – those with a much larger number of parameters – are however linked to commercial offers (among others, ChatGPT4, Gemini, Claude). In this context, there are many companies offering solutions based on pre-trained models, such as the fine-tuning of lighter LLM models – i.e., with fewer parameters – or Retrieval-Augmented Generation (RAG) solutions, which present an approach that extends LLM capabilities to specific domains and internal organizational knowledge, without the need to retrain the model and ensuring that the output remains relevant and accurate. In the development of these solutions, however, it is difficult to differentiate oneself from other players from a technical point of view, thus leaving competitiveness to commercial aspects, as demonstrated by the increasingly cheaper offers with which these services are provided.
At the same time, we have witnessed an exponential growth in technical and scientific literature, with articles easily available online and increasingly for free, given the broad success of publication models based on open repositories (for example, arxiv.org). This, on the one hand, brings evident benefits for the dissemination of knowledge and the technical and scientific development of AI; on the other hand, it increases the difficulty in selecting scientifically relevant contributions, especially when outside the mainstream. This is the case for niche industrial applications, to address which there are few reference experiences to compare with. The process of dissemination and sharing of theoretical aspects and models, defined as the democratization of AI, also has as a consequence the stagnation, if not the decline, in AI research diversity: some of the largest and most prestigious universities (MIT; University of California, Berkeley; Carnegie Mellon; Stanford University) show thematic diversity levels in AI research much lower than expected considering their volume of activity and public nature. These influential universities tend to be strong collaborators of large private companies, leading to a certain homogenization at the top of AI research. This means that some thematic areas and application domains, particularly in Industry 5.0, are not sufficiently covered. In this scenario, it is therefore important to combine considerations arising from university scientific research with observations arising from concrete application projects in the business and industrial field. Thanks also to the scientific direction activities of the Datrix Group, we can distinguish 3 distinctive elements that companies can adopt to offer competitive solutions to Industry 5.0:
- systemic vision: to give value to an AI algorithm-based solution, it is fundamental to keep in mind the industrial context in which it is placed. Therefore, a multidisciplinary approach, in which AI skills are corroborated by software development, specialist engineering, and business skills, is a key factor for succeeding in meeting industrial needs in the 5.0 paradigm.
- Experience: the availability of many alternatives would require, to consider them all, development, adaptation, and performance analysis times for each algorithm that are too long compared to the time and budget needs typical of industrial contexts. Although AutoML approaches help accelerate this decision-making process, it is nonetheless an undoubted added value to know how to evaluate the appropriateness of available algorithms and methodologies with respect to the knowledge, information, and data available, as well as scientific rigor and the specific needs of the individual use case. For example, the data to use certain solutions is not always available, just as algorithms do not always produce the results reported in literature when applied to other data. In general, the most innovative algorithm is not always the most suitable for the context.
- Identification of AI research, development, and innovation areas capable of responding promptly, and with the best technologies, to the specific needs of Industry 5.0 not yet fully intercepted by commercial solutions. This requires very advanced technical and scientific capabilities, constant attention, and investment in research and innovation themes.
What risks do we face
Although there are many opportunities emerging in the landscape of possible AI applications for Industry 5.0, it is fundamental to also consider the associated risks.
In this regard, it can be noted that while traditional software development uses structured methodologies that allow for detailed planning, AI model development is characterized by an iterative and experimental approach, in which “trial and error” is an integral part of the process leading to the finalization of the AI-based solution. Although this does not involve a disruption of the fundamental phases of any software development (design, development, and release into production), it does pose risks related to the execution timing of development projects and, above all, makes it difficult to offer the industrial client guarantees on the solution’s results: these will depend on the available data, in terms of quantity and quality, the type of problem considered, etc. For this reason, the development of AI solutions in the industrial field usually starts too timidly with Proof of Concept (PoC) activities, for which methodological guarantees are offered, generally under economic conditions that make this phase inexpensive from the point of view of investment by all stakeholders. The commercial risks regarding this model are evident: a lot of effort can be spent on low-value-added activities that then remain suspended without continuation.
Another relevant risk can be recognized in the gap between the advanced mathematical, computer science, and engineering skills necessary to develop effective AI solutions, and those typically made available by industrial decision-makers who, especially in SMEs, are very vertical on the business but often with gaps in the theoretical aspects of AI. This gap can make commercial relationships between the parties ineffective and full of misunderstandings, with the risk that these do not result in success, even in the presence of excellent technical results. In fact, one must consider that traditional software development is based on algorithms and rules that produce responses that are in principle repeatable and easily interpretable, while AI models generate a probabilistic and in some cases counter-intuitive response. It is fundamental to consider this aspect when the users of the solution do not have data scientist skills and, therefore, expect deterministic results, as well as ones that are easily interpretable and intuitive. This makes the risk very high that the value of the developed tool is perceived by the user as inferior to an expert assessment. It is therefore necessary to also work on the training of end users, to increase their awareness and familiarity in using these systems.
A further aspect linked to the speed of algorithm evolution is that related to the risk of product obsolescence, even when customized. In this regard, consider how the entry of LLMs has in a short time made some functionalities of many process management softwares (for example those based on bots, OCR, etc.) obsolete. The dynamism of skills, therefore, is reflected in the need to invest in solutions that have, on the one hand, a very short return on investment and, on the other, an ability for continuous adaptation to algorithmic evolutions. This strong need for evolutionary maintenance undoubtedly also represents an opportunity for AI solution development companies.
Finally, regulatory compliance can be particularly insidious given the diversity of regulatory frameworks, the abundance of vague principles, and the rapid pace of change characterizing the current phase. The regulatory evolution expected in the coming years certainly represents one of the risks to be managed.
What lessons to draw (at least for now)
The spread and speed of AI evolution bring many implications, not only technical but also economic, social, ethical, and anthropological. These aspects, typical of technologies that have not yet reached a high degree of maturity, make entrepreneurial and technical activity very complex, called upon to solve complex trade-offs between the opportunities and risks lurking behind the development of AI solutions, particularly for Industry 5.0. The objective of this contribution was to identify and discuss some of the opportunities and risks, with the awareness of the fundamental value of these analyses: the alternative to the approach of identifying, evaluating, and managing these risks is to make poorly informed decisions and, therefore, to risk not being part of this incredible moment of change.
The tools identified to be able to compete in this field are those of multidisciplinarity, experience, and continuous training, which, after all, are precisely those that guarantee the capacity for analysis and mitigation of associated risks.
