Outside the Doughnut: A Socio-Economic Framework for the True ROI of Human-Centric AI Supply Chains

The current emerging AI economy is being built on a measurement problem wherein AI systems are evaluated by accuracy, speed, and cost, while the human labour behind them is invisibilized. Millions of data annotators, content moderators, and labellers work under conditions of cognitive strain, unstable employment classification, and a lack of social protection - none of which are factored into standard ROI calculations. This paper argues that this omission is not accidental, but a built-in feature of how tech supply chains have been designed and governed so far. 

To explain this, we use Kate Raworth's Doughnut Economics, which defines a safe and fair space for economic activity between two limits: a social foundation and an upper ceiling. The social foundation represents the minimum working conditions, including fair wages, contracts, social protection, and psychological safety. The upper ceiling marks the point where extractive harms intensify and hidden costs begin to accumulate. We argue that AI supply chains operate outside this space on both sides: below the foundation, data workers lack basic protections; above the ceiling, firms maximise value while shifting costs onto workers.

This is enabled by a narrow conception of ROI prioritizing productivity gains and enterprise value through metrics such as operational velocity, inventory costs, linehaul freight diversion, cost-per-unit, and time saved. While these metrics account for infrastructure, engineering, and deployment costs, they exclude cognitive load, long-term psychological harm, and the cost of informal labour. Many annotation workers in regions such as Sub-Saharan Africa and Southeast Asia earn less than 2$ per hour for work that supports highly profitable AI systems (Perrigo, 2023; Gray & Suri, 2019). These costs do not disappear, they are shifted onto workers, public systems, and eventually back to firms as regulatory pressures increase.

Our aim is not to calculate a revised ROI figure. Instead, drawing on the political economy of AI, we show how current ROI frameworks systematically externalise costs onto workers, and exclude the human cost of AI. Further, we offer a policy roadmap to redefine ‘value' in AI supply chains.

The roadmap follows three phases linked to the S-curve of technology adoption. First, mandate disclosure and social dialogue to make labour conditions visible early. Second, introduce supply chain liability, including transparency in subcontracting, cross-jurisdictional enforcement, and clear employment classification. Third, develop international certification frameworks based on ILO standards, so that responsible AI practices gain tangible market value.