How will artificial intelligence (AI) transform government? Stemming from a major study commissioned by the Administrative Conference of the United States (ACUS), we highlight the promise and trajectory of algorithmic tools used by federal agencies to perform the work of governance. Moving past the abstract mappings of transparency measures and regulatory mechanisms that pervade the current algorithmic accountability literature, our analysis centers around a detailed technical account of a pair of current applications that exemplify AI’s move to the center of the redistributive and coercive power of the state: the Social Security Administration’s use of AI tools to adjudicate disability benefits cases and the Securities and Exchange Commission’s use of AI tools to target enforcement efforts under federal securities law. We argue that the next generation of work will need to push past a narrow focus on constitutional law and instead engage with the broader terrain of administrative law, which is far more likely to modulate use of algorithmic governance tools going forward. We demonstrate the shortcomings of conventional ex ante and ex post review under current administrative law doctrines and then consider how administrative law might adapt in response. Finally, we ask how to build a sensible accountability structure around public sector use of algorithmic governance tools while maintaining incentives and opportunities for salutary innovation. Reviewing and rejecting commonly offered solutions, we propose a novel approach to oversight centered on prospective benchmarking. By requiring agencies to reserve a random set of cases for manual decision making, benchmarking offers a concrete and accessible test of the validity and legality of machine outputs, enabling agencies, courts, and the public to learn about, validate, and correct errors in algorithmic decision making.