June 22, 2026

Big Tech is Increasingly Moving Beyond ‘Human in the Loop’: Can Agentic AI Oversee Its Own Governance?

The human reviewer at the end of your AI pipeline may be less useful than you think. That, at least, is the argument being made by one of Amazon’s most senior security engineers. The rest of Big Tech appears to be reaching the same conclusion, potentially signposting a major change in agentic AI governance.

In a candid interview with The Register, Eric Brandwine, VP and distinguished engineer at Amazon Security, took aim at one of enterprise AI’s most trusted safety assumptions: that keeping a human in the loop is the responsible way to govern automated systems.

Brandwine said:

“When you actually get down to it, humans are not terribly consistent. So human-in-the-loop isn’t necessarily the gold standard.”

His argument draws on a concept from safety science called the normalisation of deviance. This is the gradual process by which shortcuts become accepted practice because nothing has gone catastrophically wrong yet. Brandwine first raised the idea at AWS re:Invent in 2017. His illustration is possibly a useful one. An emergency department nurse who arrives on day one, jumping at every alarm, and within months has learned to tune most of them out. The patients are fine, until one isn’t.

His argument is that the parallel to enterprise AI oversight is direct. At any meaningful volume, a human approving AI-generated decisions isn’t really reviewing them. They’re rubber-stamping. The governance model built around that approval is providing assurance rather than protection.

Amazon’s proposed alternative isn’t the removal of human responsibility, but a redistribution of it. “If my agent writes a script that they then run, and it causes an outage,” Brandwine explained, “that’s still my responsibility.” The model centres on end-to-end accountability. Human identity tracks through every action an agent takes, even when no human is present at each step.

In practice, that means layered automated guardrails, tightly scoped agent permissions, and a complete audit trail, rather than a person stationed at the end of a conveyor belt.

The Industry Is Moving in One Direction

What makes Brandwine’s comments significant isn’t the argument alone, but the consensus forming around it.

Francis deSouza, COO and President of Security Products at Google Cloud, told reporters in April:

“It is very clear that we have moved from a human-led defence strategy to a human-in-the-loop defence strategy, to an AI-led defence strategy that’s overseen by humans. Our model for the future is an agentic fleet that does a lot of the routine cybersecurity work at a machine pace and then is overseen by humans.”

Microsoft’s Satya Nadella has made a similar case, pushing for AI systems that learn continuously through use rather than pausing at each step for human sign-off. IBM, meanwhile, has called for human accountability, not human presence, at every stage of agentic AI governance.

“We know how humans fail. We’re comfortable with it. So human-in-the-loop isn’t necessarily the gold standard,” Brandwine added.

Four of the largest business tech companies on the planet reaching the same position in the same month tends to embody a seismic change in the architectural consensus. Naturally, it’s one that CX leaders and tech buyers with existing governance frameworks built around legacy assumptions should take note of.

The commercial signals point the same way. 1Password’s reported acquisition of access-governance startup Apono, estimated at between $250 million and $300 million, reflects the growing urgency around managing what AI agents can access inside organisational systems. Agentic identity management is moving from a security niche to an infrastructure-level requirement.

What This Means for CX Operations

The governance question is already very, very live for CX teams. Agentic AI is handling complaint resolution, outbound communications, account queries, and escalation decisions in production deployments across financial services, retail, and utilities. The accountability question, who owns it when an agent gets it wrong, is all too real.

Brandwine notes that the fundamental difference between humans and AI agents is that humans fear consequences, such as job loss and legal liability. Agents do not, and that asymmetry is already being exploited.

Arguably, the practical implication for CX leaders is that presence shouldn’t be confused with governance. A supervisor processing a queue of AI decisions at pace is not a safety net. A clearly defined accountability structure, automated guardrails, and a reliable audit trail are. They also scale in a way that a fatigued human reviewer never will.

What Amazon is describing repositions the human from approver to architect. Its vision is that people will be responsible for how the system is designed, configured, and constrained. This is instead of nominally checking outputs that arrive faster than any reviewer can meaningfully assess.

Customer Experience Employee Experience