Colorado’s Revised ADMT Draft Rules Didn’t Post. Here’s the Document to Use Now.

Colorado’s revised ADMT draft rules have not appeared, despite trade-press signals that a second draft was coming by September 23, 2026. As of this writing, the Colorado Attorney General’s rulemaking page still lists only the original August 11, 2026 draft of the Automated Decision-Making Technology and Conversational AI Service rules, with no distinct revised version filed or posted.

If your compliance or L&D team paused reviewer-training work to wait for that revision, the fix is simple: unpause it. The August 11 draft is still the operative document, the formal comment deadline is still October 26, and the underlying obligations for employers using AI in hiring, performance and training decisions have not changed.

Did Colorado post a revised ADMT draft in September 2026?

No. As of late September 2026, the Colorado Attorney General’s AI rulemaking page lists only the August 11, 2026 draft of the ADMT and Conversational AI Service rules, its Notice of Hearing, and its Statement of Basis, Authority and Purpose. No second, revised draft has been filed or referenced there.

That gap matters because several legal alerts published in September described a revised draft as expected, and a few readers appear to have telescoped “expected” into “released.” A direct fetch of the Attorney General’s own comment page shows a single ruleset on file: the one from August 11. Nothing dated later has replaced it.

The Colorado Department of Law had signaled it would circulate a second draft incorporating comments submitted by an early, informal deadline of September 4, timed so stakeholders could react to it before the formal October 26 hearing. That plan has not produced a document yet. Rulemaking offices routinely slip on internal drafting timelines without changing the statutory deadlines around them, and that appears to be what happened here.

Why did law firm alerts imply a revision had already happened?

Several September client alerts described a revised draft as scheduled for September 23 rather than as delivered, and later summaries compressed that into present tense. Cross-referencing dated alerts against the Attorney General’s page is the only reliable way to tell the two apart.

Legal-update newsletters are often drafted days before publication and then lightly edited before release, which is how a sentence like “a revised draft is expected by September 23” turns into “the revised draft” in a follow-up piece two weeks later without anyone intending to mislead. A September 9 update from Consumer Finance Monitor is explicit that only the original August draft exists, which is a useful anchor point if you are trying to date-check other sources.

The practical lesson for anyone tracking state AI rulemaking: treat “expected” and “revised” as different words, and check the regulator’s own comment site before changing a training build plan based on a law firm’s forward-looking sentence.

What is the operative document for compliance and training work right now?

The operative document is the August 11, 2026 draft of Colorado’s ADMT and Conversational AI Service rules, plus its accompanying Statement of Basis, Authority and Purpose. Build your reviewer-training curriculum, your notice templates and your impact-assessment process against that text, not against a revision that has not been published.

This is the draft that defines which employment tools count as “automated decision-making technology,” what counts as a “consequential decision,” and what deployers and developers must do before and after using such a system. Everything downstream in this article, including the reviewer-competency requirements, is drawn from that August text, cross-checked against the detailed employer-focused breakdown from Fisher Phillips.

What are the real dates: comment period, hearing, and effective date?

The formal written-comment period runs through October 26, 2026 at 11:59 p.m. MST, the public hearing is set for October 26 at 10:00 a.m., and both underlying laws, the ADMT Act (SB 26-189) and the Chatbot Safety Act (HB 26-1263), take effect January 1, 2027. None of those dates has moved.

Note the last row. The January 1, 2027 effective date comes from statute, not from the rulemaking calendar, so a delayed or missing revised draft does not push it back. Teams still need to be operating compliant hiring, performance and training programs on that date regardless of when, or whether, a second draft appears.

What hasn’t changed: the human-review requirements to build training against

The August 11 draft requires “meaningful human review” of any consequential employment decision materially influenced by AI, performed by a reviewer who is independent of the original decision, trained on the subject matter, and authorized to override the system’s output. None of that has been revised or softened.

Specifically, the reviewer cannot be the person or system that generated the original AI-influenced recommendation, and cannot be that person’s direct subordinate. The rule requires documented, subject-matter training for reviewers, management independence from the outcome being reviewed, and explicit authority to override the AI’s recommendation. Reviews must be completed within 45 days, and the draft is direct on one point operations teams tend to miss: AI tools may not assist in performing the review itself. A human review that quietly runs through a second AI system to save time is not a human review under this draft.

This is exactly the kind of requirement that turns into an audit finding if it lives only in a policy document and never in a training compliance management software record showing who was trained, when, and against what curriculum version.

What must adverse-decision notices to employees and candidates include?

Under the August draft, an employer who takes an adverse action materially influenced by ADMT must send a written notice within 30 days, in plain language and at least 12-point font, that discloses the AI’s role, confirms a human reviewer was involved, and states the principal reasons for the outcome with enough specificity to identify the actual scores or data points that triggered the decision.

A generic line like “an automated system was used as part of this decision” will not satisfy that standard. The notice has to name what the tool measured, roughly how the individual scored against it, and that a qualified human reviewer looked at the case and reached the same conclusion. Candidates and employees also retain the right to request the personal data the tool used about them and to demand correction of anything inaccurate, which is a separate workflow from the review itself and needs its own owner.

Every one of those notices, along with the reviewer’s underlying documentation, is exactly what a regulator or plaintiff’s counsel will ask to see first if a decision is challenged. Storing it the way you would audit evidence for compliance training programs, tied to a case ID and a timestamp, is far easier to produce on request than reconstructing it from email threads after the fact.

How should L&D teams design a reviewer-training curriculum from the August draft?

A compliant reviewer-training curriculum needs five components: what counts as a consequential decision, how to verify reviewer independence, how to read and challenge an AI system’s output, how to document an override decision, and how to draft a compliant adverse notice within the 30-day and 45-day windows. Build it as a standalone module, not a slide bolted onto general AI-awareness training.

A five-module outline you can build today

Step 1: Scope and triggers
Teach reviewers which employment decisions in your organization actually qualify as “consequential” under the draft’s definition, and which recruiting or scoring tools are presumed to materially influence those decisions because a human reviews or acts on their ranked output.

Step 2: Independence and conflict checks
Give reviewers a short checklist to confirm they are not the original decision-maker or that person’s subordinate, and require them to log the check before opening a case.

Step 3: Reading and challenging AI output
Train reviewers to ask the vendor’s own tool for its principal reasons on a specific case, not just its overall accuracy claims, since the draft requires vendor explainability at the level of an individual outcome.

Step 4: Override authority and documentation
Walk through what a documented override looks like: what was disagreed with, why, and what evidence supported the reviewer’s alternative conclusion, completed inside the 45-day window.

Step 5: Notice drafting
Have reviewers draft a sample 30-day adverse notice using a real (anonymized) case, then check it against the plain-language and specificity requirements before it goes anywhere near a template library.

Rebuilding assessment questions for each of these modules is where most teams lose time. A scenario-based assessment design approach, where reviewers work a realistic case file rather than answer abstract policy questions, is a better fit here than a multiple-choice quiz, because the draft’s standard is about what a reviewer actually does with a case, not what they can recite about the rule.

What should reviewer-competency assessment measure before go-live?

Reviewer-competency assessment should confirm three things before anyone reviews a live case: they can correctly classify which decisions require review, they can identify a conflict-of-interest disqualifying them from a specific case, and they can produce a notice draft that would pass a plain-language and specificity check without help.

Pass or fail should be tracked against the curriculum version each reviewer trained on, not just a generic completion date, since the underlying rule text may still change once a revised draft eventually posts. That version-tracking is a job for your LMS governance framework, where course ownership, versioning and retirement already need to be defined for exactly this kind of regulation-driven content.

What happens if no revised draft ever appears before the hearing?

If no revised draft appears before October 26, the Attorney General’s office can still proceed with the hearing and finalize rules based on the August 11 draft plus the written comments received, since Colorado’s Administrative Procedure Act does not require a second draft before adoption. The comment period and hearing date do not depend on a revision being published first.

That means the compliance risk of waiting is asymmetric. If you wait for a revision that never comes, you have lost weeks of training-build time against a January 1, 2027 deadline that hasn’t moved. If you build against the August draft now and a revision later changes specific thresholds, such as which coverage standard applies to ranking tools, you adjust a curriculum you already have rather than starting from nothing with ten weeks left.

How does this fit alongside California’s parallel ADMT rules for employers?

Colorado is not the only state putting a January 1, 2027 compliance date on employer use of automated decision-making tools. California’s finalized CCPA regulations on employer ADMT carry the same effective date, with their own notice, opt-out and risk-assessment requirements that run in parallel to, not in place of, Colorado’s.

Read more: California SB 947 and AB 1883: What Changes When Newsom Decides by September 30

Multi-state employers building one reviewer-training track should design it around the stricter of the two standards where they overlap, most obviously human-review independence and individualized explanation of an adverse outcome, and then layer state-specific notice language and timelines on top. Building two disconnected training tracks now, one per state, is the more expensive path once a third state’s rulemaking lands in 2027.

What should compliance and L&D teams actually do this week?

This week, three actions matter more than waiting for a revised draft: confirm your reviewer roster against the independence requirement, start module one and two of reviewer training against the August text, and calendar a written comment for submission before October 26 if your organization has a coverage-standard preference to register.

Set a recurring check of the Attorney General’s page rather than relying on secondhand alerts, since that is the only way to catch a real revision the moment it posts rather than weeks later. And keep whatever training you build this month inside a system that can show an examiner exactly who was trained, on which curriculum version, and when, because that audit trail is the actual deliverable here, not the training session itself.

Conclusion

Colorado’s revised ADMT draft is not out yet, and the smart move is to stop waiting on it. Build reviewer-competency training against the August 11 rules, keep the October 26 comment deadline on your calendar, and treat January 1, 2027 as fixed regardless of what the rulemaking calendar does between now and then.

If a revision does post before the hearing, you will be adjusting an existing curriculum rather than starting one from a standing stop, which is a far better position to be in with three months left on the clock.