The Colorado Attorney General’s pre-rulemaking document for the ADMT Act (SB 26-189) is a study in regulatory inertia. Released in June 2026, the document poses a series of questions centered on the mechanics of human-in-the-loop systems: how to define materially influence, what constitutes meaningful human review, and how to distinguish consequential decisions from de minimis ones. Yet, across the entire inquiry, there is a conspicuous absence of any mention of autonomous agents, multi-agent systems, or software that operates independently of human direction.
This omission reveals a profound structural gap between the current regulatory framework and the reality of modern software deployment. The ADMT Act is built on the premise that automated technology is a tool used by a human deployer to assist in making decisions. The law mandates pre-use notices, post-adverse-outcome notifications, and human review processes that assume a human is at the helm.
The materially influence standard, which replaces the previous significant factor test, is broad enough that it could theoretically be applied to autonomous agents retroactively. However, the specific obligations attached to this standard are fundamentally incompatible with systems that operate autonomously. If an agent takes an action that results in an adverse outcome, the requirement for a human to review that decision is a category error. The system did not influence a human decision; it executed an action.
This disconnect is not unique to Colorado. It reflects a cross-jurisdictional pattern where every emerging framework treats AI as a system that makes recommendations to a human. Regulators are attempting to map 20th-century administrative law onto 21st-century agentic systems.