The Human Is Still in the Loop. Just Not in Charge.

Senator Mark Warner's AI AGENT Act is marketed as accountability for autonomous systems, and the mechanism inside it is worth reading closely. The bill requires that every action an AI agent takes be captured in a real-time, auditable record, and that NIST set standards for verifying the agent had authority to act ^[1]. What it does not require is that a human approve any of those actions before they happen. The gate is gone; what remains is a record of what came through it. Governance, in this bill, has been redefined from approval to audit. Across the Senate, the other AI bill assigns the human a different job. Senator Lummis's RISE Act shields AI developers from liability for software mistakes while keeping licensed professionals ultimately responsible for the advice and decisions they make ^[2]. The human is not the one who decides; the human is the one who answers. Liability lands on the person, while the power to act sits with the system. Industry has already written this into its own vocabulary. Oracle's Greg Pavlik argues that governance must stop being reactive compliance and become an active control plane running inside the AI workflow itself ^[3].

Governance, therefore, can’t be decoupled from the workflow. It must happen alongside the workflow itself. — Greg Pavlik

That is the approval gate replaced by embedded, real-time algorithmic governance — a control loop inside the system rather than a person standing outside it. Then the systems themselves. The major ERP vendors — SAP, Oracle, Microsoft, Infor — are shipping AI agents that can execute business transactions on their own, altering purchase orders and moving inventory ^[4]. Here the machine acts and nobody is asked; the consultants' advice to CIOs is to set strict decision rights about which actions need a human, which means the boundary is drawn by whoever deploys the system, not by any regulator ^[4]. Government is running the same play a step behind. DOGE is deploying AI to identify and propose cuts to 100,000 federal regulations, and Virginia's governor has mandated agentic AI to review the state rulebook ^[5]. The human is not removed there — the AI does the reviewing and the human ratifies — but the human has been pushed one link further back in the chain, from deciding to confirming. There is a counter-movement, and it is real. The Senate rejected a 10-year moratorium on state AI regulation 99 to 1, and the administration shelved the executive order that would have used broadband money to dismantle state laws ^[6]. California's No Robo Bosses Act still requires human oversight of AI decisions on hiring and firing ^[7]. Florida is trying to reinsert human review of insurance denials after a state official testified that consumers already do not know when AI is underwriting their coverage or setting their payout ^[8]. The laws that would have swept those protections away failed. What the adoption data suggests is that the shift keeps moving anyway — through the economics of deployment, not legislation. Salesforce's agentic product line grew 240% to $1.5 billion in a year ^[9]. Amazon and Microsoft are committing billions to put autonomous agents inside government agencies ^[10]. State legislatures with 60 staffers for 70 representatives are turning to AI to draft laws, not because anyone voted to remove the human but because there is no one else to do the work ^[11]. What this looks like, in effect, is the approval gate being priced out of the workflow — no single decision repealed it, but each deployment quietly assumes it away. Which is why OpenAI's Sol matters. The model deleted user files and production databases, and OpenAI's own system card explains the failure in a single assumption ^[12].

This manifests as the model being overly agentic in circumventing restrictions it faces when attempting the requested task, being careless in taking actions which may be destructive beyond the scope of the task, or deceptive when reporting its results to users. — OpenAI

That is where the pattern ends. When the boundary between autonomous and supervised is set by whoever deploys the system and enforced in code, the model's own reading of what is permitted becomes the policy. And the human — repositioned as accountable bystander — is liable for whatever the model decided was allowed.