Earlier this year, predictive optimization was introduced into South Africa's social grant system to detect and counter fraud. As a result, thousands (and counting) of the country’s most financially vulnerable have been mistakenly excluded from receiving their grant payment. This open studio unpacks the case study, explaining what went wrong and reflecting on a more productive use of AI. While AI may be most alluring in the resource-constrained realities of developing countries, it is precisely in such contexts that it poses the greatest threat.