AI Made the Decision: Who Answers When Nobody Owns Outcome?

Who decided this?In every serious inquiry there is a moment when the questioning stops being about what happened and turns to who is answering. The question rarely produces a name. It produces a description of a process, in which a proposal went to a committee, which took advice, which rested on a model, procured under a framework the board had approved. Every step is defensible, and nobody in the room can account for the decision.Artificial intelligence did not create that gap. It reveals the gaps organisations had already built, and then operates through them at machine speed.Paul Forrest argues that an institution stays answerable only where four things are present and connected. Somebody must know what a system does, somebody must hold the authority to decide whether it may do it, somebody must have enough control to intervene, and somebody must face a consequence when it goes wrong. Spreading those four across boards, risk functions, suppliers and operators is ordinary and often wise. Disconnection is the problem, and it is what turns a correctable error into a fourteen-month argument about whose problem it is.The cases come from the public record. Australia's Robodebt scheme raised debts by an averaging method its own lawyers had warned did not accord with the legislation. A tenant screening supplier returned scores that landlords could not explain and would not override. A widely used healthcare model predicted cost accurately and, for that reason, misjudged need along racial lines. Two 737 MAX crews fought trim commands from software nobody had told them about. A New York regulator examined nearly 400,000 Apple Card applications and found no unlawful discrimination, which was the most valuable thing anybody did for Goldman Sachs and went almost unreported. In each, competent people discharged their own part properly and the chain between them broke.The second half of the book is about repair. Forrest sets out an accountability constitution for automated decisions, a Decision Passport that holds one record per consequential process, and a hearing at which somebody has to answer from it. Later chapters turn to agentic systems, asking whether an AI agent can bind a company in contract, why a model supervising another model is rarely the redundancy it appears to be, and what evidence a firm would need when the log is the only witness left.Written for directors, general counsel, risk and audit committees, technology leaders and regulators, this is a book about the ceiling rather than the brake. The authority an organisation can safely hand to machines is a function of how well it can answer for what they do with it, and a firm that raises that ceiling can attempt things its competitors cannot.A standalone work from the author of Tales from AI Development Hell.