Sixty days. On January 1, 2027, Colorado's rewritten AI law reaches businesses whose software helps decide who gets hired, who gets a loan, and who gets priced. The rules behind it: the Attorney General's office held the public rulemaking hearing October 26 and closed written comments at 11:59 that night; the final set is being written, aimed at adoption before January 1. The statute's duties are already fixed.

In June, we walked through the reset itself (our June 15 piece: https://www.socosystems.net/post.php?slug=2026-06-15-colorado-ai-reset). Colorado replaced the nation's first comprehensive AI statute with a narrower one, after xAI's lawsuit had already suspended the old law's enforcement. What changed since: the rules defining "notice" and "explain" are drafted, and the effective date is 60 days away.

Here is what Senate Bill 26-189 asks of a business that runs AI.

What the rewrite dropped

The 2024 law required companies whose "high-risk" AI touched hiring, lending, and similar decisions to run formal risk programs, produce annual impact assessments, and report problems to the state. The replacement drops all three.

Early drafts floated lighter treatment for small employers. The final statute carries no small-business exemption: if you heard that businesses of a certain size are off the hook, it did not survive drafting.

Where the law applies

The law reaches AI that materially influences a consequential decision in one of seven domains: employment and hiring, financial and lending services, residential real estate, insurance, health care, education enrollment, and government benefits. "Materially influences" means the system's output is more than a trivial factor in a decision that matters to someone.

Exclusions are practical: spam filters, scheduling and workflow tools, spreadsheets a person actually analyzes, security systems, and fraud prevention sit outside. One wrinkle: a business operating in Colorado owes the same notices even when the person evaluated lives elsewhere.

The four duties for the business running the tool

The statute calls a company using one of these systems a "deployer." Four duties:

Say it up front. Before AI materially influences a covered decision, the business owes the affected person clear and conspicuous notice, met by a prominent public notice at the point of interaction or a link near the transaction.

Explain the outcome within 30 days. When a covered decision goes against a person, they receive a plain-language description of the decision and the AI's role in it, a simple process for requesting more detail (which system, which version, which vendor, what data), and an explanation of their rights.

Fix the facts. Run a real human check. On request, they can get instructions for correcting factually incorrect or materially inaccurate personal data used in the decision, and human review and reconsideration, to the extent commercially reasonable. The statute defines "meaningful human review" precisely: a reviewer with authority to change or override the decision, who considers the underlying evidence, is trained for the work, and does not default to the system's output. A rubber-stamp step will not qualify.

Keep the records. Compliance records, kept at least three years after each covered decision: version identifiers, change logs, mitigation documentation.

Corrections cover inaccurate inputs, not the system's opinions, predictions, or scores.

Your vendors owe you paperwork

Anyone who builds, sells, or licenses a covered system owes its customers documentation: intended and inappropriate uses, training-data categories where known, known limitations and risks, instructions for use, monitoring, and human review, and notice of material updates. The draft rules add a "midstream developer" role for vendors embedding someone else's AI in their product; they must pass the documentation along.

If you run software that screens or ranks candidates, prices quotes, or routes applications, that documentation feeds your notices and explanations. A vendor who cannot produce it is a conversation to have before January.

A note for clinics and practices

Practices covered by HIPAA get a partial carve-out: patient-care decisions largely sit outside this law. Employment decisions do not. If AI screens applicants or evaluates staff, the practice is covered there. HIPAA facilities also owe patients a general notice about advanced technologies, plus disclosures when AI helps decide eligibility for financial assistance. For rural clinics and practices across Southern Colorado, that is the checklist.

Where the rules stand

The October 26 hearing took up the fight over the open questions: how much influence counts as "material" (the department has proposed a more-than-minimal-factor standard alongside a presumption of material influence), and what a business owes when a vendor's AI sits behind the product it bought (Law Week Colorado, October 2026). Business groups, including the U.S. Chamber of Commerce, pushed back in written comments; the Center for Democracy and Technology argued for more worker protection (comments, October 2026). The same rulemaking also covers the state's separate chatbot law and its annual reporting.

The enforcement gap

January 1 is the statutory date. Enforcement can begin later, under the April court order: the attorney general can take no enforcement step until the final rules are adopted, then xAI has 28 days to file its injunction motion, the court rules, and 14 more days pass. That sequence has not started; no injunction filing has surfaced, and the case stays paused with nothing decided.

Only the attorney general enforces this law. Violations are deceptive trade practices; an accused business generally gets a 60-day window to cure, which knowing or repeated violations forfeit; there is no private right of action. The statute also keeps liability under Colorado's anti-discrimination law for the companies that built and used the system, allocated by relative fault (Moore & Van Allen, October 2026). The transparency regime is lighter. Discrimination claims were never part of the rewrite.

The bottom line

The work this law asks for is an inventory problem. You cannot notify anyone about tools you do not know you are running, and as we wrote in "Your Team Is Already Using AI," unapproved tools are common. The sequence: list which systems touch hiring, lending, housing, insurance, health care, education, or benefits decisions; collect your vendors' documentation; draft the point-of-use notice and the adverse-outcome explanation while the final rules settle. The January date is fixed even if enforcement is not.

If that inventory is longer than expected, or your vendors' paperwork thinner, have a conversation with someone who does this for a living.

Reporting sources: SB 26-189 session law text; Colorado AG rulemaking record, status November 2, 2026; U.S. Chamber of Commerce and CDT comments, October 2026; Jones Walker via Mondaq, October 26, 2026; Moore & Van Allen, October 20, 2026; Cloud Security Alliance note, October 20, 2026; Law Week Colorado; x.AI v. Weiser docket.