Summary From warehouses to recruitment platforms, algorithms are quietly becoming the new managers of modern work. Increasingly, decisions about hiring, scheduling, produc­tivity and even dismissal are shaped by automated sys­tems that analyse vast amounts of data. These technolo­gies promise efficiency and innovation, but they also raise important questions about fairness, transparency and worker rights. Algorithms are often presented as neutral tools, but they reflect the choices, assumptions and data used to build them. Past data is likely to reflect inequalities – such as gender or racial bias in hiring – and algorithms may reproduce and even amplify such patterns. At the same time, many systems operate as »black boxes«, which makes it difficult for workers, un­ions or regulators to understand how decisions, which are likely to affect people’s livelihoods, are really made. This paper takes readers inside the process of how al­gorithms are designed, trained and deployed in work­places. It highlights the key stages at which bias, exclu­sion or unfair treatment may enter the system, from definition of the problem to selection of data and test­ing of models. It also examines the growing importance of algorithmic audits and stronger governance to ensure accountability. The message is clear: algorithms are not destiny. With stronger oversight, worker participation and ethical design, these technologies can support more transparent, equitable and humane workplaces rather than reinforce existing inequalities. Introduction – the algorithmic turn Imagine standing on the floor of a massive warehouse, around which hundreds of workers move swiftly under the gaze of screens and scanners. Orders flash across handheld devices, dictating every motion, pick, scan, deliver, repeat. However, there is no visible supervisor shouting instructions. The authority sits elsewhere, em­bedded in lines of code. To capture this reality, it is useful to speak not only of »AI« but more broadly of algorithmic systems. This in­cludes traditional rule-based software(for example, scheduling tools or scoring systems), as well as ma­chine-learning and foundation models. Algorithms oper­ate across nearly all sectors, from payroll systems and shift schedulers to recruitment platforms, credit scoring tools and content recommendation engines. Not all al­gorithmic systems are»artificial intelligence«, properly speaking; many rely rather on rule-based software that follows predefined instructions. AI systems are a subset of algorithmic systems, built on algorithms but en­hanced with learning techniques that detect patterns and make probabilistic predictions. In short, not every algorithm is AI, but every AI system relies on algorithms. For workers and unions, the key issue is not the label, but the impact of these systems on decisions about work, pay and opportunities. While AI-driven digital management and monitoring systems promise greater efficiency, flexibility and re­sponsiveness for the workforce, the devil lies in the de­tails of how these systems are designed and governed. Without careful attention to context, consent and ac­countability, such technologies risk amplifying work­place inequalities and eroding trust(Capasso et al. 2024). Anchoring their deployment in an ethics of care (Arora, Raman and König 2023) – one that values rela ­tionality, transparency and worker well-being over mere productivity – can ensure that digital oversight empow­ers rather than disciplines, and that technological pro­gress aligns with the broader vision of humane and in­clusive work. This paper takes readers behind the curtain to trace the technical and human process of algorithm creation. It explains, step by step, how algorithms are designed, trained and deployed, but also how choices at each stage can lead to exclusion or inequality. It explores how audits are conducted, where accountability falters and what can be done to ensure that these systems serve people rather than the other way around. Part I: the making of an algorithm At its core, an algorithm is simply a set of instructions, a recipe for solving a problem. But as any cook knows, the outcome depends on what ingredients you use, who wrote the recipe and what they intended it to taste like. Developing an algorithm starts with clearly defining the problem and the desired outcome. The task is broken down into logical steps, specifying inputs, outputs and constraints. These steps are translated into code, tested with representative data and refined to handle errors and improve performance. After validation and optimi­sation, the algorithm is deployed. This sequence – defi­nition, design, coding, testing and optimisation – forms the practical»recipe« that turns an idea into a function­ing system. In the digital world, the recipe begins with a problem statement . A company might say: We want to predict which job applicants will perform best, or we want to op­timise delivery times. From there, data scientists gather information to train a model, such as past employee re­cords, CVs, performance metrics and customer ratings. The system then learns patterns, identifying what kinds of people have done well in the past. The trouble starts here: if the past reflects inequality, the model will faithfully reproduce it. This is known as historical bias . For instance, when Amazon developed a hiring algorithm in 2018(Pathak 2025), it trained the system on ten years of past resumes, most of which Part I: the making of an algorithm 3