---
type: "synthesis"
sources: ["agentic"]
tags: ["human-in-the-loop", "architecture", "escalation"]
id: "cross-hitl-escalation"
sourceVaultSlug: "hbr-seg-agentic"
originDay: 6
articleStem: "hbr-seg-agentic"
sourceUrl: "(unified vault: 9 sources)"
sourceTitle: "HBR — Firm Ⅱ-A · Agentic operating model / brand code / workstreams"
---
Almost every article, regardless of topic, lands on the same architectural pattern: **agents handle routine volume autonomously; a designed threshold escalates hard cases to humans.** It is the closest thing to a shared implementation standard in the corpus.

- A018 names it directly: [[concept-human-in-the-loop-escalation]] — [[entity-servicenow]] resolves 80% autonomously, escalating 20%; [[entity-ag1]] and [[entity-vuori]] are further examples. It is the moat [[concept-brand-agents]] hold over [[concept-consumer-agents]].
- A026 engineers the *trigger*: [[action-design-hesitation]] (confidence thresholds, anomaly detection). [[entity-ramp-d26]] escalates the toughest 10–15% of edge cases, and the humans act as teachers ([[contrarian-humans-teach-implicit-rules]]).
- A016 formalizes the *rules*: [[framework-accountability-rules]] and [[action-define-decision-rights]] specify what agents may do autonomously vs. what needs approval.
- A002 embeds it in architecture: the [[concept-orchestration-layer]] "escalates to humans when judgment is needed."
- A058 makes it a job duty: [[framework-ai-orchestration-responsibilities]] includes managing handoffs.

**The unifying insight:** escalation design *is* the mechanism that keeps [[cross-oversight-paradox|the oversight paradox]] at bay — the ratio of autonomous-to-escalated work must keep residual human load inside [[concept-oversight-capacity]]. A026 adds the subtlety that even escalation must be *sampled* correctly ([[action-govern-system]]) to avoid rubber-stamp fatigue, and encodes the claim that human oversight is [[contrarian-human-oversight-permanent|permanent]], not transitional.