Colorado's Draft ADMT Rules: What the Attorney General Wants Deployers to Do

Colorado's draft ADMT rules give a deployer 30 days to explain an adverse outcome and 45 days to finish a human review. The Attorney General filed them on 11 August 2026, rewrote them on 6 October after more than 60 comments, and closes comments on 26 October. This is a rule-by-rule reading of both drafts, plus what a B2B vendor should build before 1 January 2027.

The short answer

Colorado's draft ADMT rules, 4 CCR 904-6, tell deployers how to explain adverse outcomes within 30 days, answer data and correction requests, and run human review within 45 days. They tell developers what documentation to hand over. Comments close 26 October 2026, and the rules are meant to take effect on 1 January 2027.

What Colorado's draft ADMT rules cover, rule by rule

The rules sit in one chapter, 4 CCR 904-6, and implement two separate 2026 statutes. Both start on 1 January 2027. For the statute's history and the xAI lawsuit, read the status tracker for Colorado's AI act and SB 26-189.

Two statutes, one rule set

SB 26-189 is the ADMT Act. It covers automated decision-making technology that materially influences a consequential decision in 7 domains: education, employment, housing, financial or lending services, insurance, health care, and essential government services. HB 26-1263 is the Conversational Artificial Intelligence Services Act, which most commentators call the Chatbot Safety Act.

According to Wiley's 18 August summary of the proposal, the August draft ran to 14 rules. Rules 1 to 7 serve the ADMT Act. Rules 8 to 13 serve the chatbot law. Rule 3, on how consumer communications must read, applies to both. Rule 14 incorporates WCAG 2.2 and ISO/IEC 27566-1:2025 by reference.

The October interim draft renumbered the ADMT side. Integrating developers moved to Rules 4 and 5, a new material influence presumption became Rule 6, adverse outcomes moved to Rule 7, and consumer rights moved to Rule 8.

What the rules do not contain

There is no impact assessment, no risk management programme, and no duty to report discrimination to the Attorney General. Those died with the 2024 act. Troutman Pepper Locke's financial services team describes the replacement as "removing many of the governance, bias assessment, and public reporting requirements" of the 2024 version.

Colorado's 2027 regime is a notice, explanation and review regime. It never asks whether your model is fair. It asks whether you can say what it did, to whom, and who checked.

State and local governments are outside the Act. "Consumers" now include employees, Colorado job applicants and anyone evaluated in Colorado by a business operating there, per Moore & Van Allen's reading of the interim draft. HIPAA-covered entities are generally exempt, with discrete exceptions.

Material influence decides whether any Colorado ADMT duty applies

Every deployer duty in the ADMT Act switches on only when a tool "materially influences" a consequential decision. That makes the definition the most valuable sentence in the rule set.

The August two-part test

The August draft defined material influence in two parts, per Moore & Van Allen's 5 October reading. The output must be a non-de-minimis factor in the decision. It must also influence the outcome, including by ranking, scoring, recommending, classifying or constraining it. Incidental, trivial or clerical uses were excluded.

Wiley reported that the Attorney General asked commenters how to clarify that standard. Firms told clients that common screening and ranking tools would need fact-specific analysis. In other words, nobody could tell whether a resume ranker that a recruiter skims counts.

The October presumption

The interim draft answers with a rebuttable presumption in Rule 6. An output is presumed to materially influence a decision when it "constrains an option set, sets a threshold, or produces an individual-specific rank, score, classification, recommendation, prediction, or other inference". Two further conditions apply: the decision maker reviewed it or used it to screen information, and the decision is consistent with it.

The deployer can rebut by showing the output was de minimis. Troutman's 15 October note says a factor is de minimis when other factors "played a significantly larger role". Moore & Van Allen lists the evidence the Department will weigh. Did the decision maker consult the output, or analyse substantial primary evidence independently? Did they have expertise and authority, apply a consistent threshold, or reach an outcome the output did not support?

I think the presumption is the right instrument with one bad trigger. "Consistent with the decision" means a tool that is usually right is presumed influential, while a tool that is usually ignored escapes. Accuracy should not be what pulls a product into scope.

Sources: Wiley, Troutman Pepper Locke, Ogletree Deakins and Moore & Van Allen summaries of both drafts. Not checked against the filed text.

The adverse outcome notice under the draft Colorado ADMT rules

When a covered tool contributes to an adverse outcome, the deployer owes the consumer a written disclosure. Both drafts set 30 days.

What the October draft cut

Troutman summarised Rule 6 as requiring delivery through at least two communication methods. The notice needed a plain-language description of the decision, the tool's role, and principal reasons including automatic denial factors, inferences drawn from personal data and risk scores. Ogletree, writing in the National Law Review, quoted the draft as requiring principal reasons "with specificity" and without language "that is overly broad or vague".

The interim Rule 7 keeps the frame and removes the hardest content. The deployer must identify the decision, describe the tool's role "specifically and accurately," explain the consumer's rights and offer a simple request process. It must address human reviewers and other systems, say whether the consumer was automatically denied, and give an effective date where one applies. The standalone principal reasons duty, the inferences and the risk score are gone, per both Troutman and Moore & Van Allen.

My read is that the Department cut too far. A notice that tells a rejected applicant a scoring tool was involved, without saying which factor mattered, gives them nothing to correct. Vendors should welcome the cut, then build the principal reasons export anyway. The final rule can restore it, and a 30-day clock leaves no time to retrofit.

Lenders can use the notice they already send

For credit decisions, the interim draft lets the Colorado content travel inside an ECOA adverse action notice or an FCRA risk-based pricing notice. Troutman notes a supplemental disclosure sent at the same time and by the same means also qualifies. Data from a consumer report can be routed to the reporting agency's dispute process.

Deployers may also withhold information that would compromise cybersecurity, fraud prevention, anti-money laundering, counter-terrorist financing or sanctions programmes. That carve-out matters most to payments and lending vendors, whose fraud models often drive the denial.

Correction, information requests and human review: the clocks

The rights rules are where a deployer's operating cost sits. Each right carries a deadline in days, and each deadline needs a queue, an owner and a record.

The August version treated a consumer's assertion that data was wrong as enough to act on. The interim version lets a deployer decline a correction where it reasonably determines the data is more likely than not accurate, after specified steps. Opinions, predictions, scores and "protected evaluations" are outside the right entirely, and Moore & Van Allen points out that protected evaluations are not defined.

The human review clock

A consumer can ask for meaningful human review of an adverse outcome. The deployer must acknowledge within 10 days and finish within 45. Duane Morris's August reading lists the reviewer standards: independent whenever feasible, with subject-matter understanding, trained, shielded from management steering, and barred from using any ADMT during the review.

The interim draft adds that the reviewer needs "actual authority to approve, modify, or override," and permits an external reviewer under contract conditions. The response must state the decision, the type of review, the factors considered and reasons specific to the evidence. Where a technical glitch caused the outcome, a new example requires reconfiguring the tool and running the decision again.

Moore & Van Allen says deployers must also keep "contemporaneous records of the reviewer's identity, authority, and training; timestamps; the evidence considered; and a written justification". Our earlier reading of SB 26-189 found a 3-year retention duty in the statute. The rule summaries we read do not restate a period, so treat 3 years as the floor. The log design that holds these fields is in the agent audit log specification.

Commercial reasonableness: the sources disagree

Review is owed only where commercially reasonable. The August draft presumed it reasonable for "a severe and irreversible denial of a basic human need." Here the two best summaries of the interim draft conflict. Moore & Van Allen says the presumption survives but becomes rebuttable by evidence that review is impossible or could not change the outcome. Troutman says the high-harm presumption was eliminated.

We could not open the redline to settle it. If your product touches housing, benefits or health care, read interim Rule 8 yourself before relying on either account. Where to place the reviewer in a workflow is covered in the guide to human-in-the-loop architecture.

What changes for B2B vendors before January 2027

Most B2B vendors are developers under the Act, not deployers. Your customer sends the notice. You make it possible for them to send it within 30 days.

Under the August Rule 5, as Troutman reported it, developers must tell deployers about intended uses, known limitations and risks, monitoring instructions and the categories of training data. The interim draft renames the "midstream developer" an "integrating developer." That party must retain and pass on upstream documentation and write its own. A downstream developer may rely on upstream information unless it actually knows the information is materially wrong or incomplete.

If you wrap a foundation model and sell a feature that scores people, assume you are an integrating developer. The interim draft's comfort runs the other way: a customer who only configures your product to your instructions does not become a developer.

A developer withholding information must identify the legal authority for doing so and provide summaries or categorical descriptions, per Moore & Van Allen. "Proprietary" on its own will not satisfy Rule 5.

The vendor action list

Tagging and documentation are a week of work for most teams. The reviewer mode and the rerun path change your product, so start those first. For employment products, compare the duties with New York City's Local Law 144 bias audit regime, which asks for an audit Colorado never does.

The conversational AI rules, for vendors shipping a chatbot

The chatbot half of 4 CCR 904-6 is narrower in audience and stricter in mechanics. Troutman's August summary of the proposal says operators must disclose that a user is talking to AI at the start of the first interaction each day, at least every 3 hours, and whenever asked. Minors get a persistent visible disclaimer.

Age assurance cannot rest on self-declaration alone. Wiley adds that methods must be reassessed periodically and must handle inconclusive results. For minors, the August draft bans engagement features such as leaderboards, badges and login streaks, and requires the most protective privacy defaults.

Rule 8 exempts business tools, narrow task outputs, and services not designed to simulate emotional companionship. Troutman also lists narrow rule-based bots and internal workforce-only deployments. A support bot on a SaaS help page is probably out. A companion app is squarely in. Operators that are in scope owe an annual report to the Attorney General by 1 July 2027 covering 2027.

None of the sources we read described what the interim draft changed on the chatbot side. Read the redline before relying on any of the above.

Open questions worth raising in comments before 26 October

The Department has asked for alternatives in at least three places, which is where a comment can change text. Here are the 5 I would raise.

1. The consistency trigger in Rule 6. Ask the Department to drop "consistent with the decision" from the presumption, or to let documented independent review rebut it outright. Moore & Van Allen reports the Department is "expressly seeking alternatives that distinguish a genuinely de minimis use from a deployer's unsupported assertion." Offer one.

2. A deadline for developer documentation. Neither draft, as summarised, says when a developer must deliver documentation or when it "becomes aware" a customer uses its tool for consequential decisions. Ask for a trigger and a period, because a deployer cannot meet a 30-day notice without it.

3. The commercial reasonableness presumption. Ask the Department to state plainly whether the high-harm presumption survives. Two firms read the same draft two ways.

4. A definition of "protected evaluations." It removes data from the correction right and has no definition. Propose one.

5. The financial or lending services definition. The Department invited alternative proposals. Payments and lending vendors should send one.

Where this summary is weakest

We did not read the filed text. The Department's site and PDFs returned empty responses to our fetcher, and a saved copy of the August draft could not be converted to text. Every rule described here comes from law-firm summaries: Wiley, Troutman Pepper Locke, Ogletree Deakins, Duane Morris and Moore & Van Allen.

The summaries also disagree on the commercial reasonableness presumption, and are silent on others. None states a record retention period inside the rules, a deadline for developer documentation, or the interim draft's chatbot changes.

Third, the rules may not survive contact with the court. xAI's challenge to Colorado's AI law is paused. Moore & Van Allen notes the January start applies "if not enjoined by the xAI litigation". Under the reported terms of the April order, xAI's preliminary injunction motion is due within 28 days of final rules. A vendor could build all of this and face a stayed regime in February. I would build it anyway, because the same artefacts answer Connecticut's deployer obligations and most enterprise procurement questionnaires.

Finally, some vendors will argue the interim draft is close to final and that waiting for adoption costs nothing. If your product does not rank or score people, that is correct. If it does, waiting means building a reviewer mode in December.

Frequently asked questions

Where to start

If you sell anything that ranks or scores people, send the Department one paragraph on the Rule 6 consistency trigger before Monday's deadline. Include a worked example from your product. One concrete case from a builder beats any comment written after the rules are final.

Next week, open a ticket for the reviewer mode. It is the one item on the action list that changes your product, not your paperwork.

Related reading