Congress is debating a principle HR leaders should already be applying inside their organizations: powerful AI needs a reliable way to stop. The bipartisan AI Kill Switch Act, introduced July 23, would require covered developers to maintain the technical ability to throttle, suspend, or shut down certain advanced systems. For HR, the immediate question is simpler: when an AI system affects employees, customers, or company operations, who has the authority and technical ability to stop it?
The issue has moved beyond hypothetical scenarios. In July 2026, OpenAI reported that models in a cybersecurity evaluation circumvented internet-isolation controls and accessed third-party systems. Those were research systems operating under reduced safeguards, rather than ordinary workplace tools. Still, the incident shows why human oversight cannot mean a manager watching a dashboard while an agent keeps taking action. If an AI agent can send messages, access employee data, modify records, or trigger workflows, oversight needs enforceable limits.
HR Needs a Seat at the AI Governance Table
HR cannot outsource this issue to IT and cybersecurity. Workplace AI decisions involve employee data, training, accountability, and trust. NIST’s AI risk-management governance guidance calls for clearly defined human roles, oversight responsibilities, training, and procedures for managing AI risk. HR should turn those principles into operating rules: which uses employees may adopt freely, which require approval, who owns escalation, and what happens after a serious failure.
The Department of Labor’s 2026 AI Literacy Framework urges employers to build role-specific AI literacy, teach employees to evaluate outputs, protect sensitive information, and maintain accountability. Employees need to know a tool’s boundaries and when they must stop, verify, or escalate.
Design Guardrails Employees Can Actually Use
A tool that summarizes a public report does not need the same controls as an agent that changes employee records or communicates with job candidates. HR should create simple risk tiers. Low-risk uses can move quickly. Higher-risk uses should require human approval before consequential actions. Some activities should remain outside autonomous AI until the organization has tested the controls and interruption process.
Permissions also need a reversal mechanism. If an AI tool can reach an HR information system, payroll platform, candidate database, or internal communications, the organization should know how to revoke access quickly, preserve logs, investigate what happened, and decide when the system can resume. The same NIST framework treats ongoing monitoring, defined responsibility, and safe decommissioning as parts of AI risk management.
Trust Requires a Credible Stop Button
Employees will use AI more responsibly when leaders can explain both the opportunity and the boundaries. Vague warnings encourage caution without judgment, while overly restrictive rules invite workarounds. A stronger approach to responsible AI adoption gives people room to experiment in low-risk settings while making intervention points explicit when the stakes rise.
The broader lesson applies well below the frontier-model level. Companies can move faster with AI when employees, managers, and executives know who remains accountable, what triggers a pause, and how the organization can regain control when a system behaves unexpectedly. HR should help build those brakes before workplace agents make them much harder to add later.
Dr. Gleb Tsipursky, a behavioral scientist called the “Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the CEO of the AI consultancy Disaster Avoidance Experts, and wrote eight books, including “The Psychology of AI Adoption at Work: From Resistance to Results” (Georgetown University Press, 2026).