Most of what has been written about the "Colorado AI Act" describes a law that will never take effect in its original form. The 2024 statute, Senate Bill 24-205, with its risk management programmes, impact assessments and reports to the Attorney General, was delayed and then repealed and reenacted in 2026 by Senate Bill 26-189, a narrower law about automated decision-making technology (ADMT) used in consequential decisions. This guide describes the law as of 23 October 2026: scope, developer and deployer duties, timeline, enforcement, open rulemaking questions, and what it means for AI agents and non-human identities in decision workflows. It is for security, privacy, legal and AI governance teams that build or use scoring, ranking or recommendation systems affecting Colorado residents, and it says plainly what is not yet settled.
Key takeaways
- The 2024 law has been replaced. SB 26-189, signed on 14 May 2026 according to the Colorado Division of Real Estate, repeals and reenacts Part 17 of Article 1 of Title 6 of the Colorado Revised Statutes. The Attorney General states that "This new law and its provisions go into effect January 1, 2027."
- The test is now "materially influence", not "high-risk AI". The law covers ADMT whose output is a non-de minimis factor in a decision about a consumer in seven domains: education, employment, residential real estate, financial or lending services, insurance, health care, and essential government services and public benefits.
- Duties are about transparency and redress. Developers document uses, limitations and training data categories for deployers; deployers give notice, explain adverse outcomes within 30 days after the decision, allow correction, offer human review and keep records for three years.
- Impact assessments and the risk management programme are gone. The enrolled text has no impact assessments, risk management programme, Attorney General reporting or framework-based defence.
- Rules are not final. As of 23 October 2026 the Attorney General's ADMT rules are at the proposed stage, with a public hearing on 26 October 2026. Expect detail on disclosures and human review to change.
Scope: who and what is covered
A system is regulated only if every link in the C.R.S. 6-1-1701 chain holds: it is ADMT, it materially influences a decision, and the decision is a consequential decision in a covered domain.
| Term (C.R.S. 6-1-1701) | Meaning in the enrolled text | Practical reading |
|---|---|---|
| Automated decision-making technology, (2) | Technology that processes personal data and uses computation to generate output, including predictions. Listed exclusions include firewalls, spreadsheets without machine learning and certain tools used only to summarise or route information for human review. | Function, not technique: models, rules engines and LLM tools alike. |
| Consequential decision, (3) | A decision, determination or action made about a consumer relating to a covered domain; routine, low-stakes and fraud-prevention activities are excluded. | Individual decisions only. |
| Covered ADMT, (5) | ADMT used to materially influence a consequential decision. | The inventory unit for compliance. |
| Materially influence, (13) | The ADMT output is a non-de minimis factor used in making the decision; incidental, trivial or clerical uses are excluded. | A human final say does not remove a weighed score from scope. |
| Adverse outcome, (1) | A decision that denies, terminates, revokes or materially reduces or restricts access, or results in materially less favourable pricing or terms than similarly situated consumers receive. | Triggers the 30-day explanation duty. |
| Developer, (8) | A person doing business in Colorado that develops, offers, sells, leases, licenses or otherwise makes commercially available a covered ADMT; internal-use and research developers are excluded. | Vendors; in-house builders are usually deployers. |
| Deployer, (7) | A person doing business in Colorado that deploys a covered ADMT. | Lenders, employers, landlords, insurers, agencies. |
In the covered domains (6-1-1701(6)), insurance includes underwriting, pricing, coverage and claims, and government services include eligibility and renewal determinations.
Exemptions and carve-outs (6-1-1708): insurers subject to C.R.S. 10-3-1104.9 are deemed compliant; HIPAA covered entities and business associates are excluded except for employment-related decisions, with specific provisions for certain health-care providers; FDA-regulated devices and related research are excluded; and nonpublic personal information need not be disclosed where the Gramm-Leach-Bliley Act would prohibit it. The enrolled text read for this guide contains no small-business exemption, unlike the 2024 law, which exempted many deployers with fewer than 50 full-time employees.
Security tooling: firewalls and fraud prevention are excluded, so ordinary security tooling is outside the law. A behaviour analytics score that feeds an HR disciplinary decision may not be: map tools by the decision they influence.
Key obligations
| Duty | Who | Section | Timing and detail |
|---|---|---|---|
| Documentation to deployers: intended uses, known harmful uses, categories of training data, known limitations, instructions for use and human review, information deployers need to comply | Developer | 6-1-1702(1) | From 1 January 2027; deployer must be told if information is withheld |
| Notice of material updates and intentional and substantial modifications | Developer | 6-1-1702(2) | "Within a reasonable time"; public release notes count only if direct notice is also given |
| Compliance records | Developer | 6-1-1702(4) | At least three years |
| Records of each consequential decision | Deployer | 6-1-1703 | At least three years after the decision, or longer where other law requires |
| Clear and conspicuous notice that covered ADMT is used | Deployer | 6-1-1704(1)-(2) | Before use; a prominent public notice at points of consumer interaction can satisfy it |
| Post-adverse-outcome disclosure: plain-language description of the decision and the ADMT's role, how to request more information, consumer rights | Deployer | 6-1-1704(3) | Within 30 days after the decision; content to be shaped by Attorney General rules |
| Withholding trade secrets or protected information | Deployer | 6-1-1704(5) | Permitted, but the consumer must be told information was withheld |
| Correction of factually incorrect or materially inaccurate personal data | Deployer | 6-1-1705(1) | On request after an adverse outcome; opinions, predictions and scores need not be "corrected" |
| Meaningful human review and reconsideration | Deployer | 6-1-1705(1) | To the extent commercially reasonable; Attorney General rules required |
Liability for discrimination: 6-1-1707 makes developers and deployers potentially liable under Colorado anti-discrimination law for consequential decisions materially influenced by covered ADMT, allocates fault by relative responsibility and voids certain indemnification clauses. The enrolled text also states that compliance with Part 17 is not itself a defence. Bias testing therefore still matters: the exposure has moved from a regulatory checklist to discrimination law.
Human review that counts: Holland & Knight's commentary describes trained reviewers with authority to change the outcome who consider primary evidence rather than defaulting to the output. The Attorney General's rules will set the standard.
How the 2026 law differs from the 2024 law
Section references for the 2024 law are to the signed act.
| Topic | SB 24-205 (2024, never in force) | SB 26-189 (in force 1 January 2027) |
|---|---|---|
| Trigger | "High-risk artificial intelligence system" that makes or is a substantial factor in a consequential decision | Covered ADMT that materially influences a consequential decision |
| Core duty | Reasonable care to protect consumers from algorithmic discrimination | Transparency, explanation, correction and human review; discrimination handled through existing anti-discrimination law (6-1-1707) |
| Deployer governance | Risk management policy and programme (6-1-1703(2)) | No equivalent duty |
| Impact assessments | At deployment, at least annually, and within 90 days of a substantial modification (6-1-1703(3)) | No equivalent duty |
| Reporting to the Attorney General | Discovered algorithmic discrimination to be disclosed | No equivalent duty |
| Framework-based defence | Affirmative defence for compliance with a recognised AI risk management framework (6-1-1706(3)) | None; the enrolled text names no framework |
| Small organisations | Deployer exemption for fewer than 50 full-time employees, with conditions (6-1-1703(6)) | No small-business exemption found in the enrolled text |
| Enforcement | Attorney General only | Attorney General only, as a deceptive trade practice, with a 60-day cure period until 1 January 2030 |
The NIST AI RMF and ISO/IEC 42001 remain useful for governance and for showing diligence, but carry no statutory defence in Colorado. See the NIST AI RMF Guide and the ISO/IEC 42001 Guide.
Timeline and current status
All dates are as stated by the named primary source, checked on 23 October 2026.
| Date | Event | Source |
|---|---|---|
| 2024 | SB 24-205 passed and signed; signed text applied "on and after February 1, 2026" | Colorado General Assembly, signed act; Colorado Attorney General |
| 28 August 2025 | SB25B-004 signed, delaying SB 24-205 to 30 June 2026 (effective 25 November 2025) | Colorado General Assembly, SB25B-004 |
| 1 May 2026 | SB 26-189 introduced | Colorado Division of Real Estate |
| 7 and 9 May 2026 | Passed the Senate (34-1) and the House (57-6) | Colorado General Assembly, SB26-189 bill page |
| 14 May 2026 | Signed; specified sections took effect on that date | Colorado Division of Real Estate |
| 11 August 2026 | Proposed ADMT and Chatbot Safety rules filed with the Secretary of State | Colorado Attorney General |
| 6 and 21 October 2026 | Interim draft rules released; regulatory analysis released | Colorado Attorney General |
| 26 October 2026 | Public rulemaking hearing, 10:00 AM, 1300 Broadway, Room 1D, Denver (virtual by registration); written comments until 11:59 PM MST that day, or the last day of the hearing if longer | Colorado Attorney General |
| 1 January 2027 | Law in effect; applies to consequential decisions made on or after this date; rules on post-adverse-outcome disclosures and human review due by this date (6-1-1704(4), 6-1-1705(3)) | Enrolled text; Colorado Attorney General |
| 1 January 2030 | Cure period provision repealed | Enrolled text, 6-1-1706(3) |
Not yet settled as of 23 October 2026:
- Final rules: the Attorney General describes the current text as proposed draft rules and says written comments will be considered "In drafting the final set of rules". Commentary on the interim draft (Akin Gump) describes a rebuttable presumption that qualifying output materially influences a decision, and notice requirements covering accessibility and the languages ordinarily used with consumers. Treat both as draft.
- Federal litigation: law-firm commentary reported a federal challenge to the 2024 law and an April 2026 stipulation deferring enforcement until after rulemaking; its current status was not confirmed for this guide.
- Signed text: this guide relies on the enrolled bill. Check the signed act for final differences.
The Chatbot Safety Act (HB 26-1263), in the same rulemaking, is out of scope here.
Enforcement and penalties
- Who enforces: the Attorney General, through the Colorado Consumer Protection Act (6-1-1706(1)).
- What a violation is: "A violation of this Part 17 is a deceptive trade practice" under 6-1-1706(2)(a), and C.R.S. 6-1-105(1)(uuuu) adds it to the statutory list.
- Cure period: where a cure is possible, the Attorney General must issue a notice of violation and allow 60 days to cure before acting. The cure period does not apply to knowing or repeated violations, and the provision is repealed on 1 January 2030.
- No new private right of action: 6-1-1706(4) and 6-1-1709(1). Rights under other laws are preserved.
- Penalty amounts: SB 26-189 sets no amount of its own. Remedies come from the Colorado Consumer Protection Act; take advice on penalty levels.
Privacy law: the Colorado Privacy Act already gives consumers rights over profiling for decisions with legal or similarly significant effects. See the US State Privacy Laws Guide.
What it means for AI agents and non-human identities
SB 26-189 does not mention AI agents; what follows is this guide's reading of its functional definitions.
- Agents can be covered ADMT: an agent that screens CVs, triages insurance claims, ranks tenancy applicants or recommends credit limits is covered ADMT if its output is a non-de minimis factor.
- The summarise-and-route exclusion is narrow: an agent that only summarises or routes information for human review may fall within the 6-1-1701(2)(b) exclusion; one that adds a score, a ranking or a recommendation probably does not.
- Each agent needs an owner and a declared use: developer liability turns on intended, advertised or contracted use, so register every agent with a human owner, a documented intended use and permitted domains. See the Agentic AI Identity Guide.
- Attribution in records: records and explanations must show which agent, model version and configuration produced an output, so each agent needs its own identity, not a shared service account. See the AI Agent Observability and Incident Response Guide.
- Least privilege on personal data: Scope agent credentials and API tokens to the data the declared use needs, so a correction can be traced to every source read.
- Change control: developers must notify deployers of material updates and substantial modifications. A silent upgrade of a third-party model is a change to detect: pin model versions and record them per decision.
Deciding your position
- Developer or deployer: a vendor selling a hiring or underwriting tool to Colorado businesses is a developer. A business building its own model for internal use is normally only a deployer, as internal-use developers are excluded.
- In scope or not: start with the decision, not the technology. If no decision about an individual in one of the seven domains is affected, stop. Then ask whether the output is weighed or merely incidental.
- Small organisations: with no small-business exemption in the enrolled text, a small landlord or employer using a vendor screening tool still owes every deployer duty; require the 6-1-1702 documentation in the contract.
- Large organisations: fold Colorado into an enterprise AI inventory and a single consumer-rights workflow rather than a state-specific process. See the EU AI Act Compliance Guide and the Global AI Regulation Map.
Compliance steps
- Inventory (now): list every system, model and agent that produces output about individuals in the seven domains. Record owner, vendor, data sources and decision supported.
- Classify (now): apply the 6-1-1701 chain and the exclusions. Record the reasoning, since the draft rules suggest a presumption of material influence.
- Collect developer documentation (by December 2026): request the 6-1-1702(1) package from each vendor and add update-notification and three-year record clauses to contracts.
- Build notices (after final rules): draft the pre-use notice and the adverse-outcome disclosure template, then finalise them against the Attorney General's final rules.
- Stand up the request workflow: route requests for information, correction and human review to a single queue with a 30-day clock.
- Design human review: name reviewers, give them authority to override, train them, and give them access to the primary evidence, not only the score.
- Logging and retention: capture decision, inputs or references to them, model version, agent identity and reviewer action; retain for at least three years after each decision. See the Log Management and Retention Guide.
- Pilot and test: run the end-to-end flow on one domain before 1 January 2027 with synthetic adverse outcomes, and keep the manual process as fallback until the workflow reliably meets the 30-day clock.
- Monitor the rulemaking: assign an owner to track the final rules after the 26 October 2026 hearing and to update templates.
Framework mapping
An aid to reusing existing work, not legal equivalence; AI RMF and ISO/IEC 42001 entries are at function or topic level.
| Colorado duty | NIST AI RMF | ISO/IEC 42001 | EU AI Act (high-risk systems) |
|---|---|---|---|
| Inventory and scope classification | MAP | AI system inventory and scope of the AI management system | Classification under Article 6 |
| Developer documentation (6-1-1702) | GOVERN, MAP | AI system documentation and information for interested parties | Article 13 transparency and information to deployers |
| Notice, explanation and records (6-1-1703, 6-1-1704) | GOVERN, MANAGE | Documented information and communication | Article 12 record-keeping; Article 26 deployer obligations; Article 86 right to explanation |
| Human review (6-1-1705) | MANAGE | Human oversight controls | Article 14 human oversight |
| Anti-discrimination liability (6-1-1707) | MEASURE | AI system impact assessment (voluntary in Colorado) | Article 10 data governance |
Where a decision is solely automated and the data subject is in the EU or UK, GDPR Article 22 applies as well. See the AI Governance Guide.
Detection and measurement
- Coverage: percentage of systems in the AI inventory classified against 6-1-1701, with recorded reasoning.
- Vendor documentation: percentage of covered ADMT with a current 6-1-1702(1) package and an update-notice clause.
- Explanation timeliness: median and 95th percentile days from adverse outcome to disclosure, against the 30-day limit.
- Human review effect: share of reconsiderations that changed the outcome; near zero suggests rubber-stamping.
- Unattributed decisions: count of logged decisions without a model version or agent identity, which should be zero.
Common failure patterns
- Working to the 2024 law: spending effort on impact assessment calendars and Attorney General reporting that SB 26-189 removed, leaving the 30-day explanation workflow unbuilt.
- "Human in the loop" as an exemption: assuming a human sign-off takes a system out of scope. Materially influence turns on whether the output is a non-de minimis factor.
- Shared service accounts for decision agents: several agents under one credential, so no record shows which produced an output.
- Vendor silence: no contractual route to the developer documentation or to update notices, leaving the deployer unable to explain decisions.
- Treating draft rules as final: no plan to revise templates after the hearing.
Practitioner checklist
- Brief legal and governance owners that SB 24-205 has been repealed and reenacted by SB 26-189, in effect 1 January 2027.
- Inventory every system and AI agent producing output about individuals in the seven covered domains.
- Classify each against the 6-1-1701 definitions and record the reasoning, including exclusions relied on.
- Obtain 6-1-1702 documentation and update-notice commitments from every developer.
- Run each decision agent under its own identity with least-privilege access to personal data.
- Build the pre-use notice, adverse-outcome disclosure and 30-day request workflow; finalise after the Attorney General's final rules.
- Give human reviewers authority, training and access to primary evidence.
- Retain decision records, model versions and reviewer actions for at least three years.
Standards and references
- Colorado General Assembly: SB 26-189 Concerning the Use of Automated Decision-Making Technology in Consequential Decisions (enrolled text)
- Colorado General Assembly: SB26-189 bill history
- Colorado Attorney General: Artificial Intelligence and ADMT rulemaking
- Colorado Division of Real Estate: SB 26-189 summary
- Colorado General Assembly: SB 24-205 Consumer Protections for Artificial Intelligence (signed act)
- Colorado General Assembly: SB25B-004 bill history
- Holland & Knight: Colorado Governor Signs SB 189
- Akin Gump: Colorado Attorney General Releases Interim Draft Rules Implementing ADMT and Chatbot Safety Acts
Related NHI Mgmt Group resources: Global AI Regulation Map · EU AI Act Compliance Guide · NIST AI RMF Guide · ISO/IEC 42001 Guide · US State Privacy Laws Guide. This guide summarises Colorado SB 26-189 and its rulemaking as of 23 October 2026 and is not legal advice.