Colorado AI Act + CCPA ADMT, line by line.

The Colorado duty of reasonable care.

One sentence carries the regime. Reasonable care is the operative duty. It runs to any known or reasonably foreseeable risk, not just realised harm. Algorithmic discrimination is a defined term and not a colloquialism. The duty attaches to the developer in § 6-1-1702 and to the deployer in § 6-1-1703 in parallel construction.

The enforcement architecture telescopes. § 6-1-1706 routes a breach of the Part through the Colorado Consumer Protection Act. § 6-1-112 of that Act caps civil penalties at USD 20,000 per violation. Three steps from one duty in § 6-1-1702 to a per-consumer civil penalty in front of the Colorado Attorney General.

The defined term consequential decision sits at § 6-1-1701 and scopes the entire Part. A consequential decision is a decision that has a material legal or similarly significant effect on the provision or denial to any consumer of, or the cost or terms of, eight enumerated categories. Anything outside those eight is outside the Act. Anything inside, where the AI system is the maker or substantial factor, is high-risk.

What counts as a high-risk AI system.

Colorado defines the term twice over. The first move is the noun. A high-risk artificial intelligence system means any artificial intelligence system that, when deployed, makes, or is a substantial factor in making, a consequential decision. The second move scopes consequential decision to a closed list.

The closed list, drawn from § 6-1-1701, covers a decision that has a material legal or similarly significant effect on the provision or denial to any consumer of, or the cost or terms of, eight categories. Education enrolment or an education opportunity. Employment or an employment opportunity. A financial or lending service. An essential government service. Healthcare services. Housing. Insurance. A legal service. The list is exhaustive in the sense that the duty does not attach outside it.

Compare to the EU AI Act Annex III, which enumerates eight broadly parallel use cases, but Colorado's list and Annex III are not co-extensive. Colorado includes legal services, which is not an Annex III line. Annex III includes critical infrastructure, biometric identification, law-enforcement, migration and border control, and administration of justice — all outside the Colorado perimeter. Substantially similar but not identical. A system high-risk under one regime is not automatically high-risk under the other.

The Colorado statute carves out specific technologies that, on their own, are not high-risk. § 6-1-1701 names anti-fraud technology that does not use facial recognition, anti-malware, anti-virus, AI-enabled video games, calculators, cybersecurity, databases, data storage, firewalls, internet domain registration, internet website loading, networking, spam filters, spell-checking, spreadsheets, web caching, and web hosting. The carve-out is for component technologies. The minute the spell-checker is wrapped into a recruitment screen, the wrapper is the high-risk system, not the spell-checker.

Developer obligations under § 6-1-1702.

The developer is the entity that does business in Colorado and that develops or intentionally and substantially modifies an artificial intelligence system. The duty is reasonable care to protect consumers from known or reasonably foreseeable risks of algorithmic discrimination.

The duty is operationalised through three concrete obligations. Documentation handed to the deployer. A statement of the high-risk system's purpose, intended benefits, intended uses, reasonably foreseeable harms, training data summary, performance evaluation summary, mitigations applied, and the data governance measures in place. Public summary on the developer's website describing the kinds of high-risk systems the developer has placed and how the developer manages known and foreseeable risks. Notice to the Attorney General within ninety days of discovery, or receipt of a credible report, that the system has caused or is reasonably likely to have caused algorithmic discrimination.

The ninety-day clock is not the EU AI Act's serious incident clock. EU Article 73 runs to fifteen days for ordinary serious incidents and as short as two days for incidents involving widespread infringement or critical infrastructure. Colorado's clock is generous by comparison. It is also a one-way ratchet — once the developer knows, the clock starts, and the only stop is filing.

Deployer obligations under § 6-1-1703.

The deployer is the entity that does business in Colorado and that deploys a high-risk artificial intelligence system. The duty is the same — reasonable care against known or reasonably foreseeable algorithmic discrimination — but the operational obligations are different.

§ 6-1-1703(2) prescribes four artefacts. A risk management policy and program. An annual impact assessment. Consumer-facing notice that a high-risk AI system is being used to make, or is a substantial factor in making, a consequential decision concerning the consumer. An adverse-action explanation with the consumer's right to correct and the right to appeal via human review where technically feasible.

§ 6-1-1703(6) carries a small-deployer exemption. A deployer with fewer than fifty full-time-equivalent employees is excused from the risk management policy, the impact assessment, and the public-statement obligations only if the deployer does not use its own data to train the high-risk system, uses the system solely for the purposes the developer disclosed, and makes the developer's impact assessment available to consumers. The duty of reasonable care, the consumer notice, the adverse-action explanation, and the AG notice all remain in force regardless of headcount.

Risk management policy and the NIST AI RMF anchor.

NIST AI RMF is named in the statute. ISO/IEC 42001 is named in the statute. The third leg — another nationally or internationally recognized risk management framework — is left for the Attorney General's rules under § 6-1-1707 to enumerate. That is where additional frameworks like the OECD AI Principles, the Singapore Model AI Governance Framework, or the EU AI Act conformity-assessment process may be lifted in.

§ 6-1-1703(3) creates a rebuttable presumption. In any enforcement action by the Attorney General, the deployer is presumed to have used reasonable care if the risk management policy and program is reasonable considering one of the named frameworks and the deployer otherwise complies with the Part. The presumption is not a safe harbour. It shifts the burden onto the AG to rebut on the facts. A policy that names NIST on the cover page and is implemented nowhere does not raise the presumption.

The annual impact assessment, prescribed by § 6-1-1703(3), must include the system's purpose and intended use, an analysis of whether deployment poses any known or reasonably foreseeable risks of algorithmic discrimination, the data inputs, the outputs the deployer relied on, the metrics used to evaluate performance and mitigations, the post-deployment monitoring measures, and the user safeguards. The deployer keeps the assessment for at least three years following the final deployment of the system.

The cross-reading is straightforward. NIST AI RMF function GOVERN sits underneath the risk management policy. NIST function MAP sits underneath the impact assessment. NIST function MEASURE sits underneath the performance metrics. NIST function MANAGE sits underneath the mitigation actions and the post-deployment monitoring. The full per-function reading is in entry № 18.

The CPRA grant in § 1798.185(a)(15) and (16).

California's Automated Decision-making Technology rules do not sit in primary statute. They sit in regulations. The regulations are issued by the California Privacy Protection Agency under a delegation from the legislature in California Civil Code § 1798.185.

§ 1798.185(a)(15) directs the CPPA to issue regulations governing access and opt-out rights with respect to businesses' use of automated decision-making technology, including profiling. § 1798.185(a)(16) directs regulations on cybersecurity audits. § 1798.185(a)(14) sets up annual risk assessments where processing presents significant risk to consumers' privacy or security. The trio — ADMT regulations, cybersecurity audits, risk assessments — is the package the CPPA Board adopted on 24 July 2025 and the Office of Administrative Law approved on 22 September 2025.

The package took effect 1 January 2026. Substantive ADMT compliance phases in 1 January 2027 for businesses using ADMT to make significant decisions. Risk-assessment processing initiated before 1 January 2026 must be assessed and documented no later than 31 December 2027. The first attestation filing to the CPPA covering 2026 and 2027 risk assessments is due 1 April 2028. The compliance calendar is staggered, but the legal effective date is one date.

What ADMT actually covers.

The CPPA-finalised regulations define automated decision-making technology and bind it to a closed list of triggering uses. The most substantive trigger is a significant decision concerning a consumer.

A significant decision, in the regulations, is a decision that results in the provision or denial of financial or lending services, housing, education enrolment or opportunities, employment or independent contracting opportunities or compensation, or healthcare services. The five-category list overlaps Colorado's eight-category consequential-decision list on five lines. Lending, housing, education, employment, healthcare are common to both. Insurance, government services, and legal services are Colorado-only. Critical infrastructure, biometric identification, law enforcement, migration, and administration of justice are EU-only.

The regulations also reach extensive profiling — profiling a consumer in the context of work or educational performance, in publicly accessible places, or for behavioural advertising. The behavioural-advertising overlay attaches consumer rights even where no significant decision is being made.

The fourth trigger is using personal information to train an automated decision-making technology that the business intends to use to make a significant decision, conduct extensive profiling, or perform identification or biometric verification. This is the regulation's reach into model training. The training itself is the trigger, not the inference.

Consumer rights under the ADMT regulations.

The ADMT regulations layer three consumer rights on top of the standard CCPA right of access, right to delete, right to correct, right to know, and right to opt out of sale or share.

The first is pre-use notice. Before using ADMT for a significant decision, extensive profiling, or training-of-significant-ADMT, the business must provide a notice describing the business's use of the technology, the purpose, the consumer's rights with respect to it, and an explanation of how to exercise those rights.

The second is the opt-out right. Subject to enumerated exceptions — security, fraud detection, safety, performance of a service the consumer requested, and a narrow human-appeal exception where the business permits the consumer to appeal an adverse outcome to a qualified human reviewer — the consumer may direct the business to stop using ADMT to make the relevant significant decision.

The third is the access right. The consumer may request information about the business's use of ADMT with respect to the consumer, including a plain-language explanation of the logic of the ADMT, the role the ADMT outputs played in the decision, and the consumer's options to exercise other rights. The access right is a superset of the standard CCPA access right. The standard right reaches the personal information collected. The ADMT access right reaches the system's logic and output role.

The ADMT risk-assessment requirement.

The risk-assessment regulations apply to processing that presents a significant risk to consumers' privacy or security. The regulations enumerate triggering activities, of which significant-decision ADMT use is one. Selling or sharing personal information, processing sensitive personal information, processing personal information of consumers under sixteen, and using personal information to train ADMT for significant decisions or extensive profiling are the others.

The assessment must, at minimum, identify the categories of personal information processed, the operational details of the processing, the purposes, the benefits, the negative impacts including potential to disparate-impact discrimination, the safeguards considered and the safeguards implemented, and a determination of whether the benefits outweigh the negative impacts. The assessment is signed off by an executive responsible for the processing. It is retained for as long as the processing continues plus five years.

The filing schedule. Risk assessments for processing initiated before 1 January 2026 that continues after that date must be conducted and documented no later than 31 December 2027. Assessments conducted in 2026 and 2027 must be filed with the CPPA via the abbreviated attestation form by 1 April 2028. Subsequent annual filings follow on 1 April of the following year. The substantive document stays internal. What goes to the CPPA is an attestation that the substantive document exists and a summary of in-scope activities.

One agent, three concurrent regimes.

Take an AI agent that scores credit applications. The applicants reside in Colorado, California, and Germany. The provider is a Delaware-domiciled subsidiary of an EU-headquartered group. The system is in production today.

EU AI Act Article 12 attaches on 2 August 2026 in the AI Act as enacted (subject to the May 2026 Omnibus provisional deferral to 2 December 2027 pending OJEU) because Annex III(5)(b) covers credit scoring. Article 12 binds the provider to automatic event logging over the lifetime of the system. Article 99(4) caps the penalty at EUR 15 million or 3 percent of global turnover.

Colorado SB24-205 § 6-1-1703 attaches on 30 June 2026 — assuming no further delay — because credit scoring is a financial or lending service under § 6-1-1701. The deployer must run the risk management policy, conduct the annual impact assessment, file the AG notice on discovery of algorithmic discrimination, and provide the consumer-facing notice and adverse-action explanation under § 6-1-1704.

California ADMT obligations attach on 1 January 2027 because the financial or lending service is a significant decision. The business must give the pre-use notice, honour the opt-out, satisfy the ADMT access right, and have completed and filed the risk assessment by the calendar deadlines.

The artefact pattern is one trace, three projections. Entry № 19 walks through the multi-jurisdiction architecture in detail. The execution log is the same. What differs is which fields the regulator-facing evidence package projects, and which clock starts when.

How Warrant maps the US sub-federal stack.

Enforcement and guidance signals, Q1–Q2 2026.

The Colorado AI Act has had a noisy first quarter of 2026. On 28 August 2025, Governor Polis signed SB 25B-004 in special session, postponing the SB24-205 effective date from 1 February 2026 to 30 June 2026. On 27 April 2026, the United States District Court for the District of Colorado granted a joint motion in xAI LLC v. Weiser to suspend case deadlines and stay enforcement. Under the order, the Colorado Attorney General is to refrain from initiating enforcement, including investigations, until fourteen days after the court rules on xAI's forthcoming preliminary-injunction motion.

On the rule-making track, the Attorney General's office has indicated it does not intend to promulgate substantive rules under § 6-1-1707 until the legislative session concludes. The Colorado General Assembly's 2026 regular session ended 13 May 2026. The practical implication is that, as of this writing, no substantive enforcement action under SB24-205 has been initiated, no rules under § 6-1-1707 have been finalised, and the duty in § 6-1-1702 / 1703 is on the calendar but stayed in front of one named federal court.

California is on the opposite trajectory. The CPPA package took effect 1 January 2026. Substantive ADMT obligations phase in 1 January 2027. As of May 2026 the agency has issued no public ADMT enforcement action — the substantive obligations are not yet operational. Cybersecurity-audit and risk-assessment substantive obligations have begun their phase-in for the largest covered businesses. The CPPA's enforcement division has signalled, in its 2025 and 2026 published materials, that the first wave of focused enforcement will track the staggered compliance dates.

The signals to read for the rest of 2026. The xAI preliminary-injunction ruling. The Colorado AG's first formal rulemaking notice under § 6-1-1707. The CPPA's first ADMT-related sweep, expected to follow the 1 January 2027 trigger. The window to engineer the artefact is the window before the first action lands. Today, in May 2026, that window is open in both states.

Questions a compliance officer asks first.

Read the source directly.

• Colorado SB24-205 · Consumer Protections for Artificial Intelligence · leg.colorado.gov
• SB24-205 signed text · 17 May 2024 · PDF
• Colorado SB 25B-004 · special-session amendment postponing effective date to 30 June 2026
• Colorado Attorney General · Anti-Discrimination in AI rulemaking page
• California Civil Code § 1798.185 · CPRA rulemaking grant
• CPPA · CCPA Updates: Cybersecurity Audits, Risk Assessments, ADMT, Insurance · finalised regulations page
• CPPA announcement · OAL approval of ADMT package · 23 September 2025
• California Civil Code § 1798.155 · CCPA civil penalties
• NIST AI Risk Management Framework · referenced in § 6-1-1703(3)
• Warrant entry № 18 · NIST AI RMF, line by line
• Warrant entry № 01 · EU AI Act Article 12, line by line
• Warrant entry № 19 · One agent, many jurisdictions

Authored by Warrant Compliance, the regulatory-analysis function at Warrant. [email protected]. Editorial commentary on regulatory text. Not legal advice. Statutory paraphrases are drawn from the official enrolled text of Colorado SB24-205 as filed with the Colorado Secretary of State on 17 May 2024 and from the Final Statement of Reasons accompanying the CPPA's CCPA Updates regulations approved by the Office of Administrative Law on 22 September 2025.