Human Review Capacity Is the Constraint Nobody Budgeted For

The productivity case for generative systems usually compares the cost of producing a draft with and without automation, and the ratio is genuinely dramatic. The comparison is incomplete wherever the output cannot be used without someone approving it — which describes most of the settings where the work matters: clinical documentation, legal drafting, financial reporting, marketing claims, code entering a shared codebase, communications sent under an organisation's name. In those workflows the total throughput is not set by generation at all. It is set by how fast a qualified person can read, check and either approve or correct, and that rate has not improved. Worse, it can get slower per item. Reviewing text that is fluent and plausible but occasionally wrong is a harder cognitive task than reviewing text that is obviously rough, because the errors do not announce themselves through surface cues. The reviewer cannot skim, because skimming is exactly what a fluent error survives. Two failure patterns follow, and both are visible in practice. The first is queue growth: production capacity increases tenfold, review capacity does not move, and the backlog absorbs the entire benefit while adding delay that did not exist before. The second is worse and quieter — approval degrades into rubber-stamping. When a reviewer is measured on items processed and the last two hundred items were fine, attention decays, and the control that the whole arrangement depends on stops functioning while continuing to produce approvals. Nothing in the metrics distinguishes a careful yes from a tired one. The designs that hold up stop treating review as a uniform gate. They route by risk so that scrutiny is spent where consequences are, they surface uncertainty from the generating step so the reviewer knows where to look rather than starting cold, they sample approved output independently to measure whether review is still catching anything, and they resist the temptation to set throughput targets that make careful reading impossible. Any deployment plan that scales generation without a matching plan for checking has not identified its own bottleneck.