Decision Custody: Who Answers for a Decision After the Machine That Made It Is Gone

Abstract

If someone challenges an AI-made decision three years from now, could your organisation still show how it was made? For most, the honest answer is no. Models are replaced within a year or two, suppliers change hands, reference data is overwritten — while the decision itself lives on: a refused licence follows a person for a career; a land title outlives everyone who touched it. In Europe, the logs of a high-risk AI system may legally be kept for as little as six months. This paper names the missing obligation. Decision custody: an institution's demonstrated ability to reproduce, explain, defend, and answer for every consequential decision made in its name — human or machine — for as long as the consequences last. Grounded in a century of archival science and in ISO 14721, the paper introduces the custody gap, the four tests of custody, and five questions for a board. It is the institutional counterpart to decision sovereignty — and the question the accountability literature has not asked: not whether a decision can be examined, but for how long, and by whom. The covenant is the promise; custody is the proof.