Anthropic's New Usage Policy Puts a Human in the Loop on High-Risk Claude Work

Anthropic’s 2026 Usage Policy takes effect November 12, and the sharpest change for companies deploying Claude is about accountability, not capability. Uses that affect health, legal rights, finances, or livelihood still require a qualified human in the loop and disclosure to the people affected. The annual update adds examples for Claude’s longer, more independent work, drawing on misuse patterns in Anthropic’s September 2026 threat intelligence report.

Several specifics matter for delivery teams. When Claude controls hardware, an operator must be able to observe and stop the equipment, and it must hold a safe state if Claude disconnects. Claude cannot be used to decide who to investigate, arrest, or charge, and tracking people without consent is barred. Rules on fraud, privacy, disinformation, and elections are consolidated into a new section on deceptive campaigns and artificial activity, covering fake accounts and fabricated news sites. The Supported Regions Policy now extends to entities controlled from unsupported regions, so check your ownership structure before the effective date.

Reading this as a legal checklist misses the point. Gartner’s survey of 223 data and analytics leaders found that cultural resistance now outweighs funding constraints as the main reason governance initiatives fail, by 60% to 40%. Gartner predicts that 60% of organizations that ignore governance culture will fail to govern AI successfully by 2027. A human-in-the-loop requirement only holds if a named person has the authority, the context, and the time to say no. Otherwise it is a checkbox on a workflow that ships anyway.

The practical move before November 12: list every Claude workflow that touches a decision about a person, name the human responsible for each, and confirm that person can see what Claude did and stop it. Outcomes depend on the review step being real.