The Colorado AI Act you prepared for no longer exists. SB 24-205, with its duty of reasonable care, risk management program and impact assessments, was repealed and replaced by SB 26-189, signed on May 14, 2026, and the new law starts on January 1, 2027, although a federal court stay holds back enforcement for now. It is narrower, built around notice, explanation and human review, and it applies to any business using automated decision-making technology (ADMT) to materially influence a decision about someone’s job, loan, housing, insurance, health care, education or public benefits.
Here is the scene the new law is really about. A hiring manager opens a shortlist that a screening tool has already sorted, interviews the top three, and hires the candidate at the top. Under the Colorado Attorney General’s draft rules, it is presumed that the score materially influenced the hire anyway, and the burden of showing otherwise sits with the employer. That presumption is why we think this is a records problem as much as a legal one, and why security and data teams will end up owning much of the work.
Colorado AI Act repealed and replaced: what changed
The original statute set duties that would have kicked in on February 1, 2026. The enrolled SB 24-205 required developers and deployers of high-risk AI systems to use reasonable care against algorithmic discrimination. Deployers also had to run a risk management policy and program (the text pointed at the NIST AI RMF and ISO/IEC 42001), complete impact assessments, and tell the Attorney General within 90 days of discovering algorithmic discrimination. After industry pushback the start date moved to June 30, 2026, and in April 2026 xAI brought a federal constitutional challenge, with the US Department of Justice intervening on its side.
That lawsuit still shapes the timeline. On April 27, 2026, a magistrate judge granted the parties’ joint motion, and the minute order bars the Attorney General from starting enforcement, investigations included, under SB 24-205 “or any legislation replacing or amending” it. The bar covers violations up to 14 days after the court rules on xAI’s preliminary injunction motion, which xAI files within 28 days after final adoption of the implementing rules. In plain terms, the replacement law inherits the stay.
SB 26-189 repeals and reenacts part 17 of article 1 of title 6, the same place SB 24-205 lived. Reading the two texts side by side, the reasonable-care duty, the risk management program, the impact assessments and the 90-day discrimination report are all gone. A disclosure regime replaced them, with teeth in a few specific places.
Colorado AI Act requirements: who is covered
Coverage turns on three definitions. The signed act defines ADMT as technology that processes personal data and computes an output (a prediction, recommendation, classification, ranking or score) used to make, guide or assist a decision about an individual. It becomes a covered ADMT when it materially influences a consequential decision, meaning the output is a non-de minimis factor that affects the outcome, including by constraining, ranking, scoring or recommending.
A consequential decision touches seven domains, which are education, employment, residential housing in Colorado, financial or lending services, insurance, health-care services, and essential government services and public benefits. “Consumer” explicitly includes employees and Colorado-resident job applicants, so internal HR tooling counts.
The exclusions are generous. Firewalls, anti-malware, spam filtering and databases are not ADMT, and decisions made with technology used for cybersecurity, fraud prevention (including identity verification), anti-money laundering and sanctions screening are not consequential decisions. A tool used solely to summarize, organize, translate, draft or route information for human review is out, and so is a general chatbot, provided it is not intended for consequential decisions and an acceptable use policy forbids that use. That last condition is worth checking, because an AUP that is silent on hiring or credit decisions does not earn the exclusion.
What deployers must do from January 1, 2027
A deployer is any person doing business in Colorado that deploys a covered ADMT. Its obligations come in three parts, and each one produces a record.
Before the decision, the deployer gives a clear and conspicuous notice that a covered ADMT is or will be used, with instructions for getting more information. The act lets a prominent public notice placed close to the point of interaction satisfy this, so a link beside the application form can do the job.
After an adverse outcome (a denial, a termination, or materially worse price or terms), the deployer has 30 days to send a plain-language description of the decision and the role the ADMT played. The same disclosure explains how to request more detail, including the ADMT’s name, its version number, its developer, and the types, categories and sources of personal data used. Creditors whose ECOA and Regulation B adverse action notices also meet these requirements do not have to send a second notice.
Then come the consumer rights. The person can ask how to obtain and correct factually incorrect personal data, and can request meaningful human review and reconsideration “to the extent commercially reasonable”. Opinions, predictions and scores are not subject to correction. The act is specific about meaningful human review, too. The reviewer is designated by the deployer, has authority to approve, modify or override, is trained, considers the relevant primary evidence, does not default to the system output, and understands the output’s intended use, limitations, input categories and principal factors.
All of this sits on a retention rule. Section 6-1-1703 has deployers keep records reasonably necessary to show compliance for at least three years after each consequential decision, and names version identifiers, changelogs and records of mitigation changes as examples.
What developers owe deployers
Developers carry a documentation duty rather than a risk program. From January 1, 2027, a developer whose system is marketed or contracted for consequential decisions must give each deployer a statement of intended and known harmful uses, plus the categories of training data and the known limitations. The package also covers when the system should not be used, and how to use, monitor and review it appropriately. Material updates (new versions, changed parameters or default settings likely to change outputs) need notice to deployers within a reasonable time, and public release notes count only if each deployer gets direct notice of the release. Developers keep those records for three years.
The liability section changes how procurement should read AI contracts. In discrimination claims, fault is allocated between developer and deployer by relative fault, and a contract clause that indemnifies either party for its own discriminatory acts or omissions is void as against public policy. Our advice is to have legal read every ADMT vendor agreement for broad hold-harmless language before renewal, because a clause the act voids is protection you only think you have.
The draft rules make “material influence” the hard part
So far this reads like a notice exercise, and the Attorney General’s rulemaking is what turns it into an evidence exercise. Proposed rules were filed on August 11, 2026, and an interim redline followed on October 6, with comments and the public hearing both set for October 26.
Draft Rule 6.1 in the interim redline covers any output that constrains the option set, sets a threshold, or produces a rank, score, classification, recommendation or prediction about the individual. Such an output is presumed to materially influence the decision when the decision-maker saw it (or it screened what they saw) and it is consistent with the final outcome. Our hiring manager is close to one of the draft’s own examples, where a manager sees ADMT scores for internal candidates, interviews each one and promotes the highest scorer. The score was in front of the decision-maker and matched the outcome, so the presumption applies. Ranking candidates without showing a number does not escape it either.
The presumption can be rebutted, and the draft lists what rebuts it. The deployer can show it never consulted the output, that it applied a consistent hard cut-off, or that the decision went against the output. It can also show that someone with authority and subject-matter understanding independently reviewed substantial other information and reached the same result. Orrick’s summary adds that the draft keeps the deployer responsible when a third party runs the system on its behalf, and allows internal or external reviewers as long as they have actual authority to override.
Every one of those rebuttals is a fact you prove with records created at decision time, which is the gap we expect most organizations to have. Interview notes, a reviewer’s rationale and the model version behind the score are rarely captured together, let alone kept for three years. These are draft rules, and the final text may move after the October 26 hearing, but the statute’s own retention and versioning language points the same way. The stay does not change the advice. Eckert Seamans reads it as making enforcement unlikely before late 2027, and we agree a documented record still matters whenever it starts.
What to do now: preparing for the Colorado AI law
Most of this work is reusable whatever the final rules say.
- Inventory decisions, not models. List every process in the seven domains where software produces a score, rank or recommendation about a person, including vendor tools inside HR, lending and claims platforms.
- Test each one against the exclusions, and write down why. Note the AUP condition for chatbots and the fraud and identity carve-outs.
- Request developer documentation now, under section 6-1-1702, and add material-update notice terms to renewals. Check indemnity clauses against section 6-1-1707.
- Log the decision, not just the model. For each consequential decision, capture the ADMT name and version, the output the reviewer saw, the reviewer’s identity and authority, and their rationale, with three-year retention and access controls that fit data this sensitive.
- Build the adverse outcome path, meaning the 30-day disclosure template, the correction request route, and a human review queue staffed by people trained to disagree with the system.
- Track the final rules after the October 26 hearing and adjust templates, not architecture.
If you already run an AI management system, much of this maps onto controls you have, and our ISO 42001 certification guide covers the supplier controls that make step 3 easier. For steps 1 and 4, an AI governance and assurance review is how we usually help teams build something an auditor or regulator can test.
Frequently asked questions
When does the Colorado AI Act take effect?
SB 26-189 takes effect on January 1, 2027, and applies to consequential decisions made on or after that date. Its rulemaking provisions took effect on passage, and the Attorney General must adopt rules on adverse outcome disclosures and consumer rights by January 1, 2027. Enforcement is a separate question, because the federal stay in the xAI case covers replacement legislation until the court rules on a preliminary injunction motion.
Who enforces SB 26-189, and is there a private right of action?
Only the Attorney General enforces it, under the Colorado Consumer Protection Act, and a violation is a deceptive trade practice. The act creates no new private right of action. Before January 1, 2030, the Attorney General must give notice and 60 days to cure where a cure is possible, though knowing or repeated violations get no cure period.
Does the Colorado AI law apply to companies outside Colorado?
It applies to developers and deployers “doing business in Colorado”. Consumers include Colorado-resident job applicants and anyone whose access to an opportunity in Colorado is evaluated in a consequential decision. A company headquartered elsewhere that hires, lends or insures in Colorado should assume it is in scope and test its tools against the definitions.
Bottom line
Treat January 1, 2027 as the date your decision logs must be able to answer one question, which is whether a trained person with authority actually judged this outcome, or just agreed with the score. Inventory the decisions this month, get vendor documentation into contracts, and start capturing version, output and reviewer rationale per decision before the final rules land.