An agent that constantly pauses for human approval isn't automation — it's a new notification queue. This aggregates the audit log by task type and surfaces the override hotspots: where the AI gets stuck most. A small-sample fluke won't be flagged, and an outcome it can't prove is clean is counted as friction.
Honest scope. A deterministic aggregator + ranker over the events you feed it (task type + outcome); it surfaces where overrides cluster so a PM can fix the prompt or add context. It does not collect the log or diagnose the cause. FAIL-CLOSED — an unknown outcome counts as friction (a state not proven clean isn't assumed clean), an untyped event lands under "(unspecified)" rather than vanishing, and a type needs minimum volume before it's flagged a hotspot (no 1/1 false alarms). Source: flowdesk-friction-kernel.js.