Amazon is challenging one of the most widely accepted ideas in AI governance: that putting a human in the approval chain automatically makes AI systems safer.
Speaking to The Register, Amazon Security Vice President and Distinguished Engineer Eric Brandwine argued that humans are often treated as a reliable safeguard against AI mistakes, but in practice they are far less consistent than many organizations assume. According to Brandwine, people tend to overestimate their ability to continuously review and validate AI-generated decisions, especially when operating at the speed and scale required by modern AI agents.
The traditional “human-in-the-loop” model requires a person to review AI outputs and approve actions before they are executed. For years, this approach has been promoted as a cornerstone of responsible AI deployment. However, Amazon argues that repeated approval tasks create fatigue, causing reviewers to gradually become less effective over time. What begins as careful oversight can quickly turn into routine button-clicking, reducing the value of the human review process.
Brandwine’s position reflects a broader shift occurring across the technology industry. As AI agents become capable of handling larger volumes of work, companies are increasingly moving toward models where AI performs routine operations autonomously while humans supervise the overall system rather than reviewing every individual action. Similar views have recently been expressed by leaders at major cloud and AI providers who believe machine-speed operations cannot realistically be governed through constant human approvals.
Rather than relying on continuous manual intervention, Amazon advocates for stronger guardrails, predefined policies, automated controls, and governance frameworks that constrain what AI systems can do. Under this approach, humans remain responsible for setting rules, defining acceptable behavior, and monitoring outcomes, but they are not expected to approve every action generated by an AI agent.
The debate comes at a critical moment for the AI industry. Regulators in many jurisdictions continue to view human oversight as an important safeguard, particularly in high-risk areas such as healthcare, finance, critical infrastructure, and law enforcement. At the same time, organizations deploying large-scale AI systems are discovering that manual review processes may become impractical as agentic AI systems execute thousands or millions of decisions per day.
The discussion ultimately highlights a growing tension within AI governance. The challenge is no longer simply whether humans should oversee AI, but how that oversight can remain meaningful when machines operate faster than people can realistically evaluate. As enterprises continue adopting autonomous AI agents, governance models may evolve from “human-in-the-loop” toward “human-over-the-loop” systems, where people supervise the framework rather than every individual decision.