Antonio Casilli's Waiting for Robots and Ursula Plesner's research on digital work show the human effort inside every AI system, and what directors can ask while there is still time to act.
This is for chairs, directors and executives who weigh AI business cases and want to see the whole of the work they are approving.
When we picture artificial intelligence, many of us picture a machine. A model that learns, a system that answers, a process that runs by itself. Since 2016, the sociologist Antonio Casilli and his colleagues have been meeting the people left out of that picture.
Reviewing Casilli's book Waiting for Robots in the journal Organization Studies, Ursula Plesner of Copenhagen Business School sums up its central idea, that behind every algorithm there is a worker (Plesner, 2026). I shared that idea on LinkedIn this week. It deserves time and attention, because it changes the questions a board can ask about AI.
Casilli is a professor at Télécom Paris, part of the Institut Polytechnique de Paris. Waiting for Robots first appeared in French in 2019 as En attendant les robots. The English edition, translated by Saskia Brown, was published in 2025 and updated to address generative AI tools such as ChatGPT (University of Chicago Press, 2025).
The title echoes Samuel Beckett's Waiting for Godot. Like Godot, the robots that are meant to take over our work are always about to arrive, and in the meantime people keep doing the work. The book brings together three kinds of digital labour. There is on-demand platform work, such as ride-share driving and delivery. There is microwork, where people complete small online tasks such as labelling data. And there is the unpaid contribution of ordinary users, as we react to posts, prove we are not robots and approve or improve what generative AI produces (Casilli, 2025).
In a 2026 interview with Polytechnique Insights, Casilli offers a careful correction to the language many of us use. Having met workers in countries such as Madagascar, Kenya and Brazil, he says "one discovers that their work is hidden but not invisible." He describes it as work "involving the preparation, verification and sometimes imitation of artificial intelligence." He is also clear about who does much of it. "It is also, and above all, people on very low wages and in rather precarious situations who keep the machines running." (Casilli, 2026).
A 2020 study by Paola Tubaro, Antonio Casilli and Marion Coville maps this work in detail. Drawing on research from the DiPLab project, they show that microwork sits between three poles, which they call "artificial intelligence preparation", "artificial intelligence verification" and "artificial intelligence impersonation" (Tubaro et al., 2020). In plain terms, people prepare and label the data AI learns from, people check what it produces, and at times people stand in for the system altogether, doing work that is presented as automated.
Their conclusion is the one I hope directors will hold. This work, they write, "is a structural component of contemporary artificial intelligence production processes – not an ephemeral form of support that may vanish once the technology reaches maturity stage." (Tubaro et al., 2020). As the technology matures, the human work travels with it.
Plesner brings this home to organisations of every kind. With her colleague Lise Justesen, she has studied what they call invisible digi-work, the work that arises in and around a digitalised organisation to "prepare, maintain and repair its key features." (Justesen & Plesner, 2024). They describe three kinds. Connecting work keeps systems, data and people linked when connections break down. Compensating work makes up for the errors, limits and shortcomings of digital technologies. Cleaning work removes what is seen as mess, from disturbing content online to flawed records in a dataset.
One example they draw on is a study of a hospital pharmacy by Barrett and colleagues (Barrett et al., 2012, as cited in Justesen & Plesner, 2024). A dispensing robot was introduced to help prevent errors and support the pharmacists in their work. It also produced new errors, which had to be handled elsewhere. The assistants who fed the robot took on the extra tasks, at times working off-hours to keep up, and were moved to a small, cluttered back end while the pharmacists worked in the front. The assistants felt they had been demoted. Little of this was visible to the pharmacists who benefited from the robot.
Justesen and Plesner also show how such work slips from view. It is often "unaccounted for in formal documents, such as organizational charts, formal job descriptions, budgets and business cases" (Justesen & Plesner, 2024).
Those last two words are where directors come in. When an AI business case reaches the board, the savings are usually easy to see. Hours saved, processing times and staffing figures appear in clear numbers. The connecting, compensating and cleaning work tends to sit outside those numbers.
Justesen and Plesner observe that efficiency gains often fall short of expectations, because employees take on new and often unexpected tasks to make automated solutions work at all (Justesen & Plesner, 2024). Some of that work lands inside the organisation, with frontline staff, data teams and managers. Some of it sits far away, with contractors and platform workers the board will never meet.
Directors do not need to become technologists to see this work. A few calm questions can bring it into view.
Who will connect, check and correct the outputs of this system, and is their time in the plan?
Which tasks are moving to staff, customers or contractors, and under what conditions?
Where in our supply chain do people label, moderate or review data for us, and what do we know about how they are treated?
How will we hear early if the people carrying this work are under strain?
For many Australian organisations, these questions connect with obligations directors already hold. Under the Modern Slavery Act 2018 (Cth), entities with annual consolidated revenue of $100 million or above prepare an annual statement on the risks of modern slavery in their operations and supply chains (Attorney-General's Department, n.d.). The Act requires that statement to be approved by the entity's principal governing body and signed by a responsible member (Modern Slavery Act 2018, s. 13).
Data work for AI does not in itself amount to modern slavery. Yet when an organisation buys AI services, the people who label, moderate and review data form part of its supply chain. The guide to AI governance from the Australian Institute of Company Directors and the Human Technology Institute makes a related point, encouraging boards to "Recognise that principles and policies need to be proactively implemented and enforced across the supply chain." (Australian Institute of Company Directors & Human Technology Institute, 2024).
Wellbeing belongs in the same conversation. Casilli says plainly that "Data workers are a group that is highly exposed to psychosocial risks." He describes isolation for those working from home, and close supervision for moderators in secure offices who are unable to talk with one another (Casilli, 2026). Inside our own organisations, Justesen and Plesner call for attention to "the emotional costs, mental loads and unintended tasks associated with digitalization." (Justesen & Plesner, 2024). That is a workforce wellbeing question as much as a technology one.
Noticing work that sits out of sight asks something of our attention. In our survey of 185 Australian directors, Justin Brienza, Vince Murdoch and I found two forms of mindfulness linked to wisdom in the boardroom (King, Brienza & Murdoch, 2026). Embodied Mindfulness is attention to your own inner state. Group Attunement is a felt sense of the group as a whole, including what sits beneath the discussion. As we wrote, mindfulness in organisations is also about "attending to the wider environment and the underlying purposes that shape attention and action" (King, Brienza & Murdoch, 2026, p. 386).
When an AI paper reaches the table, Embodied Mindfulness helps a director notice the pull of a tidy saving, and the small doubt that can follow it. Group Attunement helps a director sense when the room has moved quickly past the question of who will carry the work. Together they help a board stay with the whole picture long enough to see it. These findings come from directors describing themselves.
In the Metta tradition, loving kindness begins with those close to us and widens, step by step, to people we will never meet. An AI system offers a practical place to practise that widening. Behind the answer on the screen are colleagues who check and correct it, and at a distance, people in countries such as Kenya, Madagascar and Brazil who prepared the data it learned from.
A board that asks who will carry the work, and under what conditions, can catch early warning signs while there is still time to act. In asking, it also extends its care to everyone whose effort keeps these systems running, wherever they are.
May all who keep these systems running be safe, be well and be treated with dignity.
Attorney-General's Department. (n.d.). About. Australian Government Modern Slavery Statements Register. https://modernslaveryregister.gov.au/about/
Australian Institute of Company Directors & Human Technology Institute. (2024). A director's guide to AI governance. https://www.aicd.com.au/content/dam/aicd/pdf/tools-resources/director-resources/a-directors-guide-to-ai-governance-web.pdf
Casilli, A. A. (2025). Waiting for robots: The hired hands of automation (S. Brown, Trans.). University of Chicago Press. (Original work published 2019)
Casilli, A. A. (2026, June 17). Behind every algorithm: The unseen workers powering AI [Interview]. Polytechnique Insights. https://www.polytechnique-insights.com/en/columns/digital/ai-the-invisible-yet-very-real-work-of-data-workers/
Justesen, L., & Plesner, U. (2024). Invisible digi-work: Compensating, connecting, and cleaning in digitalized organizations. Organization Theory, 5(1). https://doi.org/10.1177/26317877241235938
King, E., Brienza, J., & Murdoch, V. (2026). A new framework for measuring two key dimensions of mindfulness in the boardroom. In D. Rooney, J. P. Brienza, A. Intezari, & E. King (Eds.), Elgar encyclopedia of wisdom research (pp. 380–398). Edward Elgar. https://doi.org/10.4337/9781035302628
Modern Slavery Act 2018 (Cth). https://www.legislation.gov.au/C2018A00153/latest/text
Plesner, U. (2026). Media review: The work to make AI work [Review of the book Waiting for robots: The hired hands of automation, by A. A. Casilli]. Organization Studies, 47(11), 1778–1781. https://doi.org/10.1177/01708406261418199
Tubaro, P., Casilli, A. A., & Coville, M. (2020). The trainer, the verifier, the imitator: Three ways in which human platform workers support artificial intelligence. Big Data & Society, 7(1). https://doi.org/10.1177/2053951720919776
University of Chicago Press. (2025). Waiting for robots: The hired hands of automation [Book description and reviews]. https://press.uchicago.edu/ucp/books/book/chicago/W/bo239039613.html