A European logistics company had deployed an AI-assisted procurement system that managed supplier communications, generated purchase orders, and flagged invoicing anomalies at a volume no human team could match. It was working well, until the operations director looked closely at one specific category: orders above €50,000, which under company policy required explicit human authorisation before execution.
The system was flagging these correctly. The problem was the approval queue. Between the operations director and his two deputies, they were authorising an average of thirty-eight high-value orders a day, on top of their other work. The dwell time per order was approximately ninety seconds: a purchase order committing the firm to a six-figure supplier relationship was getting less than two minutes of senior attention.
The governance policy was being complied with in form while being violated in substance, and there was no visible mechanism to fix it, because slowing the AI down would eliminate the productivity gains that justified having it at all.
This is the exact problem the Three Control Taxonomies solve, and it's the second proprietary framework in The Verification Economy. The answer the logistics firm eventually found wasn't full human review, and it wasn't full automation. It was a tiered architecture: one of three distinct oversight models, each suited to a different combination of risk and speed.
HITL: Human-in-the-Loop. The full gate. Every output is reviewed and explicitly authorised by a qualified human before release. This is the correct architecture for high-risk, client-facing, or legally sensitive work, and it's the one most organisations default to for everything, which is exactly how you end up with a director giving six-figure decisions ninety seconds of attention.
HILE: Human-in-the-Loop Enforcement. The AI operates autonomously inside pre-defined boundaries (a spending threshold, a category exclusion, a variance ceiling) and only escalates to a human when a boundary is approached or breached. The human isn't reviewing every output. They're placed exactly at the points where their judgement is actually required.
HOTL: Human-on-the-Loop. Full AI autonomy for low-risk, high-volume, reversible work, with a human monitoring aggregate performance and retaining kill-switch authority. Critically, this is not "no oversight." The human is governing the system rather than reviewing individual outputs, which is a structural distinction most organisations get wrong.
The goal of governance architecture is not maximum human involvement. It is optimal human involvement: placing human judgment precisely where it creates genuine value and removing it from the tasks where it adds friction without adding protection.
These three models govern human oversight of what an AI system produces. The Green Mitigation's seventh pillar, Transparency & Verification, is the same discipline applied one layer down, to the infrastructure the system runs on, not the outputs it generates. A firm that can prove HITL/HILE/HOTL discipline on its outputs but can't prove what its AI actually costs in energy, carbon, or resilience terms has only verified half the system.
A boutique investment advisory firm in Singapore rebuilt its research workflow around this exact logic, sequencing AI generation, structured junior verification, and senior authorisation as discrete, timed, documented stages, every document carrying a verification identifier a client could query to see exactly who reviewed it, for how long, and what they changed. Within six months, two institutional clients cited that infrastructure specifically as their reason for deepening the engagement. One moved its entire research mandate from a larger competitor, in writing, because that competitor couldn't demonstrate equivalent oversight.
The firm's managing director put it precisely: "We didn't build the Human Loop because we were worried about errors. We built it because we understood that in a world where every firm has an AI, the firms that survive are the ones that have something the AI doesn't have. We have accountability. That's the product."
The book gives you the full Governance Spectrum decision model, the four boundary types that make HILE actually work (with the design principles that separate a real boundary from a false one), and the complete HITL Architecture Blueprint: the four components that turn "someone reviewed it" into a defensible, documented, commercially valuable governance system.