No One Owns This Answer: Designing the Chain from an AI's Output to an Accountable Person

When a product lets an AI answer, rank or act, the design problem is no longer the prompt. It is ownership: who owns the answer once a person acts on it, and who fixes it when it is wrong. Four public incidents — Moffatt v. Air Canada (2024), Amazon's recruiting engine (2014–17), Knight Capital (2012) and the Replit agent incident (2025) — each show the same missing link, at a different level of AI authority: an output reached a person with no source, no reason, no preview, no owner and no way back. From the incidents we derive four frameworks a team can apply in order (map the interaction, set the authority, design the trust cues, gate the release) and seventeen design patterns, grouped in four suits by what the AI does, each paired with one metric to keep score. We map the patterns to four moments of a product cycle and show all seventeen inside one worked example, a meeting assistant. The contribution is practical: a vocabulary and a review method for teams shipping AI today, grounded in what has already gone wrong.