Last reviewed: April 27, 2026
Key Takeaways
- The Colorado AI Act (SB 24-205) takes effect June 30, 2026, after a date push from the original February 1, 2026. Penalties run up to $20,000 per violation, and each affected consumer counts as a separate violation.
- It is the first comprehensive state AI law in the United States and the closest US analogue to the EU AI Act’s high-risk regime — but the obligations are distributed differently, with deployers carrying the heavier load.
- The law splits duties between developers (who build or substantially modify high-risk systems) and deployers (who use them to make consequential decisions in eight domains: education, employment, financial services, government services, healthcare, housing, insurance, legal services).
- Compliance with the NIST AI Risk Management Framework or ISO/IEC 42001 creates a rebuttable presumption of reasonable care — the cleanest affirmative defense in the statute.
- A replacement framework — the “KILO draft” / Proposed ADMT Framework — was released in April 2026 and could supersede SB 205 with a disclosure-driven model effective January 1, 2027. As of April 21, 2026 it has not been formally introduced as a bill. Companies face a two-deadline planning problem.
What Does the Colorado AI Act Require?
The Colorado AI Act requires developers and deployers of “high-risk” AI systems to use reasonable care to protect Colorado consumers from algorithmic discrimination, and to back that care up with documentation, impact assessments, consumer notices, and an annual review. Enforcement is handled exclusively by the Colorado Attorney General, with civil penalties up to $20,000 per violation under the Colorado Consumer Protection Act (leg.colorado.gov).
The statute, formally titled “Concerning Consumer Protections in Interactions with Artificial Intelligence Systems,” is codified at C.R.S. Title 6, Article 1, Parts 1701–1706. Governor Jared Polis signed it on May 17, 2024, after a Senate floor amendment rewrote the bill to its current form. He signed with explicit reservations about compliance costs and the absence of a federal standard, and he urged the General Assembly to revisit the law before its effective date (Seyfarth Shaw).
The law was originally scheduled to take effect on February 1, 2026. A special legislative session in August 2025 failed to deliver the substantive overhaul Polis wanted; what passed was SB 25B-004, signed August 28, 2025, which substituted “June 30, 2026” for “February 1, 2026” throughout the developer- and deployer-obligation sections and changed nothing else (Akin Gump tracker).
Two pieces matter for any compliance program. First: there is no private right of action. Consumers cannot sue under SB 205. The Attorney General is the sole enforcer, with civil investigative demand authority and the ability to designate violations as per se unfair or deceptive trade practices under the Colorado Consumer Protection Act. Second: a 60-day cure period lets developers and deployers correct discovered violations and avoid liability — provided they notify the AG and document the fix.
Interactive tool
Test your AI system against Colorado SB 24-205. Six questions, one verdict, cited to the statute. Free, no login, results stay in your browser.
What Counts as a “High-Risk AI System” Under the Colorado Definition?
A system is high-risk if it makes, or is a substantial factor in making, a consequential decision in one of eight statutory domains. The Act defines AI broadly: “a machine-based system that, for any explicit or implicit objective, infers from inputs how to generate outputs, including content, decisions, predictions, or recommendations” (C.R.S. 6-1-1701).
A “consequential decision” is a decision with a material legal or similarly significant effect on the provision, denial, cost, or terms of any of the following:
| # | Domain | Examples |
|---|---|---|
| 1 | Education | Enrollment, scholarship eligibility, academic placement |
| 2 | Employment | Hiring, promotion, termination, compensation, task allocation |
| 3 | Financial / lending | Loan approval, credit limits, interest rates, account access |
| 4 | Government services | Benefits eligibility, licensing, permit approvals |
| 5 | Healthcare | Treatment recommendations, coverage decisions, clinical prioritization |
| 6 | Housing | Rental applications, mortgage approvals, property valuations |
| 7 | Insurance | Underwriting, claims processing, premium setting, coverage denial |
| 8 | Legal services | Case outcome predictions, legal aid eligibility, sentencing tools |
Three categories are excluded by the statute: narrow procedural tasks that do not replace human decision-making; anti-fraud or anti-malware technology, unless it relies on facial recognition; and communication platforms with acceptable use policies that prohibit discriminatory content (DWT).
The legal trigger to memorize: algorithmic discrimination. The Act defines it as a condition where AI use produces unlawful differential treatment or impact disfavoring individuals based on age, color, disability, ethnicity, genetic information, limited English proficiency, national origin, race, religion, reproductive health, sex, veteran status, or other protected classifications. Self-testing for discrimination, intentional diversity initiatives, and private clubs are explicit exceptions.
If your system touches a consequential decision in any of the eight domains and a real person’s outcome is affected, assume it is high-risk until you have a documented analysis to the contrary.
What Are Developer Obligations?
Developers — the entities that build or substantially modify high-risk AI systems — owe a duty of reasonable care to prevent algorithmic discrimination, plus a documentation hand-off to deployers. Five specific obligations apply.
| Obligation | Detail |
|---|---|
| Reasonable care | Use reasonable care during system development to prevent algorithmic discrimination. Compliance with a recognized risk-management framework (NIST AI RMF, ISO/IEC 42001) creates a rebuttable presumption of reasonable care. |
| Documentation to deployers | Provide model cards, dataset cards, known limitations, harmful or inappropriate use cases, evaluation methods, bias mitigation measures, and data governance practices — sufficient to let deployers complete their own impact assessments. |
| Public website statement | Maintain a summary on the developer’s website describing the categories of high-risk systems developed and how discrimination risks are managed. |
| AG + deployer notification | Within 90 days of discovering algorithmic discrimination — or credible evidence of it — notify the Colorado Attorney General and all known deployers of the affected system. |
| Proprietary protections | Documentation may be marked as trade secret or proprietary; attorney-client privilege is preserved. The AG cannot use the documentation hand-off as a back door to discoverable trade secrets. |
The hand-off documentation matters most in practice. Deployers cannot complete their impact assessments without it. Developers who treat the documentation requirement as a perfunctory PDF leave their customers exposed and their own contracts open to indemnification claims. Pacific.ai’s compliance walkthrough is blunt about this: the developer-deployer split is a contracting problem before it is a compliance problem (pacific.ai).
For the reasonable-care duty itself, the NIST AI Risk Management Framework is the safe-harbor reference most US developers reach for first. Colorado names it explicitly. ISO/IEC 42001 is the international alternative. Either provides the rebuttable presumption. For the broader landscape of how AI vendors get pulled into US litigation as developers, see our AI liability in the United States overview.
What Are Deployer Obligations?
Deployers — the entities that use a high-risk AI system to make or substantially influence a consequential decision — bear the heavier obligation set. Six categories apply.
Risk Management Program
Deployers must implement an iterative risk management policy and program that specifies the principles, processes, and personnel used to identify, document, and mitigate algorithmic discrimination. The program must align with the NIST AI RMF, ISO/IEC 42001, or another framework designated by the AG. It must scale with the deployer’s size, the system’s complexity, and the sensitivity and volume of the data the system processes (Schellman).
Impact Assessments
Impact assessments are the centerpiece of the deployer obligation set. They are required:
- Before initial deployment of any high-risk system
- Annually thereafter
- Within 90 days of any intentional and substantial modification
Each assessment must document the system’s purpose and intended use cases, known and reasonably foreseeable risks of algorithmic discrimination plus mitigation strategies, the categories of input and output data, performance metrics and known limitations, transparency measures, and post-deployment monitoring processes. Deployers must retain assessments for three years after final deployment.
Consumer Notifications
Deployers owe consumers a pre-decision notice: that a high-risk AI system is being used, the purpose and nature of the decision, a plain-language description of the system, contact information for inquiries, and information about the right to opt out of profiling. The Act provides one exception: no notice is required where AI use is obvious to a reasonable person.
Adverse Decision Handling
When the system contributes to an adverse decision, deployers owe the consumer five things. The principal reasons for the decision. The degree of the AI system’s contribution. The data types and sources processed. An opportunity to correct inaccurate personal data. And, where technically feasible, the right to appeal the decision and obtain human review.
Website Disclosures and Annual Review
Deployers must publish a website statement describing deployed high-risk systems, the discrimination risk-management approach, and data collection practices. They must also conduct an annual review of each deployed system to confirm it is not causing algorithmic discrimination.
Small Business Exemption
Deployers with fewer than 50 full-time equivalent employees are exempt from the risk-management policy, impact assessments, and website disclosures. Three conditions must be met. The deployer does not use its own proprietary data to train or fine-tune the system. The system is used for its intended purpose as specified by the developer. And the deployer makes the developer’s impact assessment available to consumers. The exemption does not waive consumer notification duties; small businesses still owe pre-decision and adverse-decision notices (Skadden).
There is no revenue threshold. Headcount is the sole exemption test.
What Changed in the 2026 Amendments — and What’s Coming with the KILO Draft?
Two events bracket the 2026 amendment story. The first happened in late summer 2025 and only moved the deadline. The second started in March 2026 and could rewrite the whole law.
SB 25B-004 — The Date-Only Amendment
Signed August 28, 2025, SB 25B-004 substituted “June 30, 2026” for “February 1, 2026” in every relevant section of the Colorado AI Act. It changed nothing else. A separate bill, SB 25-318, would have delayed the effective date to January 2027, redefined algorithmic discrimination, and broadened exemptions; it failed. The date-only amendment was the only legislation to emerge from the special session (Akin Gump).
Working Group Consensus — March 2026
On March 17, 2026, the Colorado AI Policy Working Group — convened by Governor Polis and made up of consumer groups, hospitals, school districts, and large and small technology companies — reached unanimous consensus on a framework to replace the Colorado AI Act. The Governor’s office announced the consensus the same day and promised legislative text to follow (Colorado Governor’s Office).
The KILO Draft — April 2026
The framework was formalized into a draft titled “Concerning the Use of Automated Decision Making Technology in Consequential Decisions” — the Proposed ADMT Framework, codenamed the “KILO draft” (DCI Consult, April 13, 2026; Mondaq / Mayer Brown, March 25, 2026).
The KILO draft is a fundamental rewrite. It moves Colorado away from the EU AI Act-style risk-based governance model and toward a disclosure-and-transparency model closer to state data privacy laws. The shift, in plain English, is “less about how you build and monitor your systems and more about disclosing to consumers what is happening” — as one practitioner described it in Law Week Colorado (April 21, 2026).
What the KILO draft removes:
- The defined duty of care for developers and deployers
- Mandatory risk management policies
- Mandatory pre-deployment, annual, and post-modification impact assessments
- The 90-day algorithmic-discrimination notification to the AG
What it adds:
- Up-front consumer notice when AI or ADMT is used in a consequential decision
- Plain-language disclosures within 30 days of an adverse outcome
- An opportunity to correct inaccurate inputs
- Human review on request
- Three-year recordkeeping of compliance records
- Explicit liability provisions that did not exist in SB 205 — attributed in Law Week Colorado to attorney Marisa Baum
What it keeps: AG-only enforcement and the absence of a private right of action.
If the KILO draft is enacted as written, it would take effect January 1, 2027 — replacing the SB 205 obligations on the books for June 30, 2026. As of April 21, 2026, however, the KILO draft has not yet been formally introduced as a bill in the Colorado General Assembly. The legislative session ends in mid-2026, leaving a narrow window for introduction, debate, and passage before the SB 205 effective date.
The practitioner consequence is a two-deadline planning problem. Build a compliance program for SB 205 as written. Track the KILO draft’s progress in parallel. If KILO is introduced and passes before June 30, 2026, your SB 205 obligations may never bind. If KILO stalls, SB 205 binds in full. The compliance checklist below is built to work under both scenarios.
How Does Colorado Compare to the EU AI Act?
Colorado’s law is the most EU-like state AI law in the United States, but the obligations are distributed differently. The cleanest way to see the difference is dimension by dimension.
| Dimension | Colorado AI Act (SB 24-205) | EU AI Act |
|---|---|---|
| Approach | Risk-based, focused on algorithmic discrimination | Risk-based, four-tier (unacceptable / high / limited / minimal) |
| High-risk trigger | Substantial factor in a consequential decision in 8 domains | Annex III use-case list + Annex I product-safety integrations |
| Primary obligation bearer | Heavier on deployers (impact assessments, consumer notice, appeals) | Heavier on providers (conformity assessment, technical documentation, CE marking) |
| Territorial reach | Developers and deployers doing business in Colorado | Global — any provider or deployer whose systems affect EU users or produce EU-used outputs |
| Enforcement | Colorado AG; civil penalties up to $20,000 per violation | National supervisory authorities + EU AI Office; fines up to EUR 35M or 7% of worldwide turnover |
| Risk-management safe harbor | NIST AI RMF / ISO/IEC 42001 (rebuttable presumption) | Harmonized standards under EU standardization bodies |
| Prohibited uses | None (no “unacceptable risk” tier) | Social scoring, real-time biometric ID (with exceptions), manipulation |
| Private right of action | None | None for the AI Act specifically (possible via GDPR or product liability) |
For dual-jurisdiction practitioners, three points carry real weight. One: if you have a working EU AI Act conformity-assessment program for Annex III high-risk systems, you already have most of the documentation Colorado deployers need from their developers — model cards, dataset cards, evaluation methods, bias mitigation. Two: the EU AI Act’s CE-marking provider focus is not directly transferable; under Colorado law, the impact-assessment burden sits on deployers, and deployers cannot delegate it. Three: the EU AI Act has prohibited-use categories Colorado does not. A system that is unacceptable in the EU can still be deployed in Colorado, subject to discrimination liability.
For the federal context, the White House’s March 2026 AI framework proposes overriding state AI laws via a federal preemption statute, and the AG-led AI Litigation Task Force was created in part to challenge them. Colorado’s law, like Texas TRAIGA, is squarely in that preemption target zone.
The Compliance Checklist
The checklist below assumes you may be a developer, a deployer, or both. It is structured so the same actions cover SB 205 as written and position you for the KILO draft if it passes.
1. Inventory your AI systems. List every AI system that touches a consequential decision in the eight Colorado domains. Tag each as developer-built, deployer-deployed, or both.
2. Classify each system. Apply the consequential-decision test. If a system makes or substantially factors into a consequential decision, treat it as high-risk until proven otherwise.
3. Adopt a risk-management framework. NIST AI RMF or ISO/IEC 42001. Pick one and document the choice. Compliance with either gives you the rebuttable presumption of reasonable care.
4. Run impact assessments. For every high-risk system: pre-deployment, annually, and within 90 days of substantial modification. Document purpose, risks, mitigation, data categories, performance metrics, and monitoring. Retain three years.
5. Build the consumer-notice plumbing. Pre-decision notice, adverse-decision notice, opt-out of profiling, data correction, appeal with human review. Build it into the product, not into a compliance memo. If KILO passes, you will need a 30-day plain-language disclosure path for adverse outcomes — overspec the notice infrastructure now.
6. Hand off documentation between developers and deployers. Developers: model cards, dataset cards, limitations, foreseeable risks, evaluation methods. Make this a contractual deliverable, not a marketing PDF.
7. Publish website disclosures. Both developers and deployers owe public statements describing the high-risk systems involved and the discrimination-risk approach.
8. Stand up the 90-day discovery-to-notification workflow. If you are a developer and you find evidence your system caused or is likely to cause algorithmic discrimination, you have 90 days to notify the Colorado AG and known deployers. Train someone to recognize the trigger.
Our recommendation. If you are a small developer or a sub-50-FTE deployer, do not skip the consumer-notice steps. The small-business exemption covers the heavy operational lift (risk-management policy, impact assessments, website disclosures), but consumer notification still applies. If you are larger, treat the impact-assessment template as your single source of truth — every other deliverable falls out of it.
Sources
Official Sources
- Colorado General Assembly — SB 24-205 bill page: https://leg.colorado.gov/bills/sb24-205
- Colorado General Assembly — enrolled bill text (PDF): enrolled bill text
- Colorado Attorney General — AI rulemaking page: coag.gov/ai/
- Colorado Governor’s Office — Working Group consensus announcement (March 17, 2026): colorado.gov
- Akin Gump — SB 25B-004 postponement tracker: akingump.com
- NIST AI Risk Management Framework (AI 100-1): nist.gov
- ISO/IEC 42001 — AI management system standard: iso.org
Analysis & Commentary
- Mondaq / Mayer Brown — KILO draft analysis (March 25, 2026): mondaq.com
- Law Week Colorado — KILO draft reporting (April 21, 2026): lawweekcolorado.com
- DCI Consult — KILO draft overview (April 13, 2026): dciconsult.com
- Skadden — Colorado AI Act analysis: skadden.com
- DWT — risk-based regulation analysis: dwt.com
- Reed Smith — EU AI Act and Colorado AI Act comparison: reedsmith.com
- Center for Democracy and Technology FAQ: cdt.org
- Schellman — what you need to know: schellman.com
- Pacific.ai — developer/deployer compliance guide: pacific.ai
- TrustArc — SB 24-205 compliance guide: trustarc.com
- CO-AIMS — annotated full text: co-aims.com
Data Sources
- Future of Privacy Forum — Colorado AI Act policy brief (PDF): FPF policy brief
- NAAG — deep dive on Colorado’s AI Act: naag.org
Related Reading
US AI Regulation Series:
- NIST AI Risk Management Framework — The framework Colorado names as an affirmative defense
- AI Liability in the US — How developer reasonable-care duties land in actual litigation
- Texas TRAIGA Compliance Guide — The other major 2026 state AI law
- White House AI Framework 2026 — The federal preemption push targeting state laws like Colorado’s
- SEC and AI: What Financial Firms Need to Know — sector enforcement playbook, six AI washing cases
- EU vs US AI Regulation: The Definitive Comparison — cross-jurisdiction picture, NIST RMF as bridge, Digital Omnibus status
- Colorado High-Risk AI Classifier (interactive tool) — test your AI system against SB 24-205 in six questions
- Illinois AI Employment Law 2026: AIVICA + HB 3773 — disparate-impact standard with private right of action via IHRA
- NYC Local Law 144: AI Bias Audit Guide — first US mandatory bias audit; DCWP enforcement now active
- FDA AI Medical Devices: PCCP + EU AI Act Comparison — 1,451+ AI devices authorized; PCCP framework for adaptive AI
EU Comparison:
- EU AI Act Annex III — The high-risk classification system Colorado borrowed and adapted
This article provides general information about AI regulation and does not constitute legal advice. Laws and policies change frequently. Consult qualified legal counsel for compliance decisions specific to your organization. Reg Intel is not a law firm and does not provide legal services.
Last verified: April 27, 2026. The KILO draft / Proposed ADMT Framework status is changing fast — re-verify legislative status before relying on the dual-deadline analysis.