On Saturday, October 24, 2026, Combs Intelligence Group LLC filed a public comment with the Colorado Attorney General on the proposed rules at 4 CCR 904-6 — the rules that implement Colorado’s Automated Decision-Making Technology Act (SB26-189). The filing went in two days before the comment deadline, under our name, on the public record.

This is the work this firm does: we read regulatory text at the operative level, form a practitioner’s position, and put it in front of the regulator while the rules can still change. This article lays out what Colorado is building, the four points our comment argued, and what any business running automated decisions should be doing before the rules take effect on January 1, 2027.

What Colorado’s ADMT Rules Cover

ADMT is automated decision-making technology: software that produces a score, ranking, recommendation, or decision about a person. The Colorado framework attaches duties when that output materially influences a consequential decision — the decisions that shape a person’s livelihood and access, like employment, housing, credit, education, and health care.

For the businesses deploying these systems, the proposed rules build a working set of obligations: assess the risks of the system before and while it runs, tell consumers when automated decision-making is in play, give a person who gets an adverse decision a real path to appeal it to a human, and keep records that show how a decision was actually made. The Attorney General holds exclusive enforcement, with a 60-day cure period.

Every statehouse in the country is watching rulemaking like this. What Colorado finalizes will be read in Oklahoma, Texas, and everywhere else a legislature is drafting its own version.

How We Built the Comment

We argued from the rule text itself. The easily found version of 4 CCR 904-6 is the August proposal, and it does not contain the material-influence presumption that sits at the center of this rulemaking. The presumption lives in the Attorney General’s October 6 interim redline. Our comment cites that redline — Rule 6.1, Rule 6.1B, Rule 6.2 — because a comment aimed at text the regulator has already replaced persuades nobody.

That close read changed the comment itself. Our early draft treated the rebuttal path as missing. The redline shows Rule 6.1B does enumerate evidence a deployer can use to rebut the presumption. So we revised the point to credit what the rule does — and to press on the gap that remains. Regulators can tell when a commenter read the current draft. The revision is the receipt.

The Four Points We Filed

1. Judge “materially influence” by the decision record. Rule 6.1 presumes that an ADMT output materially influenced a consequential decision whenever a reviewed score, ranking, or recommendation is consistent with the outcome. Read that plainly: the human review itself becomes the trigger. A deployer who routes outputs through a qualified reviewer lands in the same position as one who rubber-stamps them. That penalizes the oversight the Act exists to encourage. Our comment asks the rule to judge the output’s actual role by the contemporaneous decision record — what the decision-maker saw, weighed, and did.

2. Give the rebuttal path a sufficiency standard. Rule 6.1B lists evidence that can rebut the presumption: the deployer did not consult or incorporate the output, a hard cut-off was consistently applied, the decision-maker conducted an independent review of substantial information, the decision was not consistent with the output, and other contemporaneous evidence of independent judgment. The list is welcome. It also never says how much of that evidence is enough, or what makes evidence “persuasive” under Rule 6.1B.5. We asked for three clarifications: a stated sufficiency standard for rebuttal evidence; confirmation whether partial satisfaction of Rule 6.1B.3 — independent review and real decision authority, without a formal subject-matter showing — can still rebut; and worked examples in Rule 6.2 of successful rebuttals, matching the presumption examples the rule already provides. A deployer who cannot know in advance whether its decision record will hold will treat the presumption as a conclusion, because in the cases that matter most, it is one.

3. Create a safe harbor for documented human review. Deployers who maintain a named reviewer with genuine override authority, who log overrides, and who keep a decision record for each consequential decision should receive a defined compliance benefit in exchange for that real oversight. Certainty for documented review is the mechanism most likely to produce meaningful human review in practice — oversight with a paper trail, performed by a person who can actually say no.

4. Confirm the process if the effective date moves. Enforcement is currently stayed in xAI v. Weiser. If the January 1, 2027 effective date shifts, commenters and deployers should receive a supplemental comment window on any revised text. Businesses are building compliance programs against this date; the process should keep them informed if the target moves.

Where the Rulemaking Stands

Honesty about status: these rules are proposed, the rulemaking is still open, and no comment — ours included — has been adopted into final text. Anyone who tells you a filed comment changed a rule is selling you something. What a comment does is put a practitioner’s argument on the record, in front of the people drafting the final version, with citations to the sections that need work. That is the most any filing can claim, and it is worth doing well.

What Businesses Should Do Now

You do not need a Colorado address to use this checklist. If your business uses software that scores, ranks, screens, or recommends people, the habits below are the substance of every ADMT regime now being drafted:

1. Inventory your automated decisions. List every tool whose output touches a hiring, lending, housing, pricing, or eligibility decision. Most companies find a system nobody remembered connecting.
2. Name a human reviewer with real authority. The reviewer needs the power to override the output, the information to judge it, and a name on the workflow. A reviewer who cannot say no is decoration.
3. Keep a decision record for every consequential decision. What the output said, what the reviewer saw, what was decided, and why. That record is your rebuttal evidence, your audit trail, and your defense — build it as you go.
4. Log overrides. Every time a human overrules the system, record it. A healthy override log proves the review is real; an empty one raises the question the presumption answers for you.
5. Build the appeal path before you need it. A consumer who receives an adverse decision should reach a human who can reconsider it. Write the path down, staff it, and test it.

Do this work now and the final rules — Colorado’s or your own state’s — arrive as a documentation exercise. Skip it, and the same work arrives as an enforcement response.

Where to Start

The checklist above runs on documents: an AI use policy, a risk register, a vendor scorecard, an approval workflow for human review, and an incident playbook for the day a decision gets challenged. The AI Governance Starter Kit from Combs Intelligence Group packages those working documents as fill-in templates built for small and mid-sized teams. It is $197, and it is the same documented-review system our Colorado comment argues regulators should reward.

Get the kit here: The AI Governance Starter Kit

From Combs Intelligence Group, published on Big Jerm Says.

Big Jerm