---
id: "action-govern-system"
type: "action-item"
source_timestamps: ["§ Designing from the Real Organization"]
tags: ["governance", "quality-assurance"]
related: ["concept-machine-speed-compounding", "claim-multi-agent-failure"]
action: "Assign aggregate outcome responsibility and mix weak-confidence AI cases with normal cases for human review."
outcome: "Maintains human reviewers' baseline calibration of 'normal' and prevents rubber-stamping fatigue."
speakers: ["K. Sudhir"]
sources: ["agentic"]
sourceVaultSlug: "hbr-seg-agentic"
originDay: 6
articleStem: "hbr-new-26-agentic-systems-implicit-rules"
sourceUrl: "https://hbr.org/2026/06/how-to-design-agentic-systems-around-the-implicit-rules-that-govern-your-company"
sourceTitle: "How to Design Agentic Systems Around the Implicit Rules that Govern Your Company"
---
# Govern the multi-agent system and sample human review deliberately

**Action:** Assign aggregate outcome responsibility and mix weak-confidence AI cases with normal cases for human review.

**Outcome:** Maintains human reviewers' baseline calibration of 'normal' and prevents rubber-stamping fatigue.

Assign **operational responsibility for the aggregate outcomes** produced by multiple interacting agents — watch for [[concept-machine-speed-compounding]], where each agent looks blameless but the system decays (e.g., retention drops six months later). Govern the *system*, not just each agent.

**Human-review sampling discipline:** When routing cases to reviewers, deliberately *mix* cases where agent confidence is weakest with a sufficient volume of 'normal' cases. If reviewers only see rare errors, their attention decays into **rubber-stamping** and they lose their baseline sense of what normal looks like.

This is Step 3 of [[framework-design-real-organization]] and the operational answer to the 40–80% failure risk in [[claim-multi-agent-failure]].