A machine may make the decision. A machine may not explain it.
That is the load-bearing sentence in Colorado's interim redline for the Automated Decision-Making Technology Act, and almost nobody building agents has read it. Released by the Colorado Department of Law on October 6, the draft closes a gap federal agencies have left open since the spring. The relevant clause sits inside the human review provision: the ADMT itself may not assist in the review. Not to draft the summary. Not to retrieve the record. Not to flag the error.
Every agent architecture I have reviewed this quarter violates that by default. That is the finding.
Context
SB 26-189 reaches any automated decision-making technology that materially influences a consequential decision about a Colorado consumer, employee, or applicant โ employment, financial services, healthcare, housing, insurance, education, government services, legal services. Proposed Rule 5 splits duties.
Developers owe documentation: intended uses, training data categories, known limitations, known harmful uses, monitoring and meaningful-human-review instructions, and methods to control or instruct the ADMT to mitigate risk. They must also notify deployers of material updates and intentional modifications.
Deployers carry the operationally heavier load. Pre-decision notice, clear and conspicuous. Post-adverse-outcome disclosure within 30 days, through two communication channels, containing a plain-language description of the decision and the ADMT's role, the principal reasons, the ADMT's name, version and developer, the categories of personal data used, and how to request more information and human review. Then: a meaningful review within 45 days of request, by an independent reviewer with relevant expertise and genuine authority to override.
Enforcement is not decorative. The Colorado AG holds exclusive authority. A 60-day cure period applies to actions initiated before January 1, 2030. And the rules void contractual indemnification for an entity's own ADMT-related anti-discrimination violations as against public policy. The effective date, January 1, 2027, is statutory. Rulemaking cannot move it. A hearing sits on October 26 in Denver; comments close 11:59 PM MST that night. The AG is also soliciting input on three open questions: the precise definition of the material influence standard, how obligations apply when a deployer relies on a third-party vendor to operate covered ADMT, and further developer-obligation clarifications.
Core
Three numbers reprice this. 45 days. 30 days. Two channels.
I have spent the last year modeling machine-to-machine payment viability โ agents using ZK proofs to settle without human intervention, a market I put near $2 trillion if latency and cost barriers fall. Forty-five days is 3,888,000 seconds. Against a settlement rail that clears in milliseconds, that is not a delay. It is an option written against every consequential decision, and the deployer is short it.
My audit background matters here. In 2017 I read fifteen ICO whitepapers and learned that the document a builder is forced to write is the document that gets subpoenaed. "Known harmful uses" is not a disclosure field. It is an admission. A developer who files it accurately hands a plaintiff the map; a developer who files it vaguely fails Proposed Rule 5 on its face. That is a genuine squeeze, and it is the first time a US state has made model documentation a discoverable artifact by design.
The indemnification clause is the underrated weapon. Cross-border payments scale because liability is sliced, priced, and sold downstream. Void that transfer for anti-discrimination violations and the risk never leaves the originator's balance sheet. It must then be capitalized, and capitalized risk must be priced into the fee. Liability is not a legal line item; it is a cost of capital. Colorado just widened the spread on agent-mediated consequential decisions, and the market has not marked it.
Contrarian
The consensus read is that this is a US compliance tax and agent innovation will route around it. Wrong on both counts.
The federal vacuum is doing the opposite of what people assume. DOJ signaled in September โ per Bloomberg Law's October reporting on Attorney General Todd Blanche โ that it will not pursue standalone AI enforcement through prosecution, consistent with the hands-off lineage traced to an April 2025 crypto memorandum. The Warner AI AGENT Act (S. 5051), introduced July 21, has no markup. The Hawley-Murphy AI Agent Accountability Act, announced October 1, has not reached committee. States are filling by default, and fifty frameworks is not a fragmentation problem โ it is a pricing problem. You price to the strictest state and sell everywhere. Colorado becomes the US baseline by arithmetic, not by authority.
The sharpest question in the docket is the one nobody is bidding on: how obligations apply when a deployer relies on a third-party vendor. If obligations flow through to vendors, the agent supply chain gets re-intermediated โ and the winner is whoever can certify the deepest chain, not whoever ships the best model. I watched this exact dynamic in rollups. The OP Stack and ZK Stack divergence was never really technical; it was who could convince more projects to deploy first. Agent infrastructure is about to run the same play.
Takeaway
The pivot was not a retreat, but a recalibration โ and it belongs to the builders, not the regulators. The first enforceable agent-liability framework in the United States is no longer hypothetical. Yields are not gifts; they are risks wearing suits, and the 60-day cure period is the most gift-shaped risk in the document. In a tape where protocols are bleeding LPs for want of a narrative, compliance infrastructure is the only story with a statutory deadline attached.
We do not predict the wave; we engineer the vessel. Between now and October 26, every agent team selling into Colorado has one decision: map your decision chains to Proposed Rule 5 now, or discover in January 2027 that your architecture was non-compliant by construction. The open question I would fund research on is narrower than the AG's list โ does "material influence" capture payment routing? Because if it does, every cross-border agent rail is in scope, and the $2 trillion machine-economy thesis needs a 45-day asterisk.
Behind every transaction is a map of human greed. Colorado just made the map public.