Verifiable AI Governance Series Maryland HB 820 California SB 1120 Colorado SB 26-189 CMS-0057-F EU AI Act Article 12
Regulatory Analysis Series
v2.0 The Rewrite  ·  Passed 9 May 2026  ·  Effective 1 Jan 2027
Regulatory Analysis · Colorado

Colorado SB 26-189: from high-risk AI to ADMT — what the rewrite means for healthcare deployers

After two years and three legislative sessions, Colorado has rewritten its first-in-the-nation AI law. SB 26-189 narrows the obligation profile but broadens the reach — eliminating federally-regulated-entity exemptions and sweeping more healthcare entities into scope. The headline reading is that compliance got easier. The healthcare-deployer reading is more complicated.

PublishedMay 12, 2026
Reading time~21 minutes
AudienceGeneral counsel, compliance, AI governance leadership at healthcare deployers, Medicare Advantage plans, Medicaid managed care plans, and UM vendors operating in Colorado
9 May 2026Passed Colorado General Assembly (House 57–6); awaits Governor Polis
1 Jan 2027Effective date; AG mandatory rulemaking deadline
7 domainsCovered: employment, education, financial services, insurance, healthcare, residential real estate, essential government services
No PRANo private right of action; AG-only enforcement via Colorado Consumer Protection Act

01What just changed, and why it matters

On May 9, 2026 — three days before this paper's publication — the Colorado General Assembly passed SB 26-189, a complete rewrite of the state's 2024 AI law. The new statute does not amend the old one; it repeals and replaces it. For most businesses reading the headline summary, the rewrite looks like a softening. For healthcare deployers reading the fine print, it is a meaningful expansion of the regulated perimeter.

The law that SB 26-189 replaces — SB 24-205, the Colorado AI Act or "CAIA" — was signed by Governor Jared Polis in May 2024 as the first-in-the-nation comprehensive state AI statute. Governor Polis signed it with reservations, asking the legislature to revisit the law. The legislature could not reach agreement during 2025; a 2025 special session pushed the effective date from February 2026 to June 2026.

By spring 2026, the path for the original law had collapsed. xAI filed a federal constitutional challenge on April 9, 2026. The US Department of Justice intervened April 24. The Colorado District Court stayed enforcement of SB 24-205 on April 27. With the June effective date approaching and the law under federal court stay, the legislature acted. SB 26-189 was introduced May 1, passed both chambers within nine days, and arrived at the Governor's desk on May 9 with broad bipartisan support (House vote: 57–6). The bill takes effect January 1, 2027.

The shift in one sentence

SB 26-189 swaps Colorado's prescriptive risk-management-program regime for a narrower disclosure-and-transparency regime — but it eliminates the federally-regulated-entity exemptions that previously narrowed the original law's healthcare reach, with the net effect that more healthcare entities now sit inside the perimeter, doing less prescribed work but bearing direct AG enforcement exposure under the Colorado Consumer Protection Act.

02"Automated decision-making technology," defined

The most consequential structural change is the redefinition of the regulated subject matter. The original law regulated "high-risk artificial intelligence systems." The new law regulates automated decision-making technology (ADMT), defined functionally rather than technically.

"Automated decision-making technology" means technology that processes personal data and uses computation to generate output, including predictions, recommendations, classifications, rankings, scores, or other information that is used to make, guide, or assist a decision concerning an individual. Colorado SB 26-189, as engrossed (paraphrased)

The definition is technology-agnostic. It captures any system — machine-learning, rule-based, statistical, hybrid — that processes personal data and produces output used in a decision about a natural person. The framework's centre of gravity has shifted from "what is the underlying technology" to "what is the technology being used for, and against whom." This is the same conceptual move California's privacy regulator made in its 2024–2025 ADMT regulations under the CCPA, and it represents the emerging consensus in state-level AI regulation.

2.1 What the definition covers

An ADMT must satisfy three elements: it must process personal data; it must use computation to generate output; and the output must be used to make, guide, or assist a decision about an individual. In healthcare deployment this captures prior-authorization recommendation engines, clinical-decision-support systems used in coverage decisions, risk-stratification scoring models, eligibility-determination logic, and the prioritisation logic embedded in utilization-management workflows.

2.2 What the definition excludes

SB 26-189 explicitly excludes: calculators, databases, firewalls, spell-checkers, and certain spreadsheets; tools used solely to summarise, organise, translate, or draft content for human review; advertising, content moderation, cybersecurity, fraud prevention, and AML/CTF compliance systems; consumer-facing LLMs not marketed for consequential decision-making (provided an acceptable use policy accompanies them); and routine clerical tasks. The LLM exclusion is the most interesting policy choice: the moment a general-purpose LLM is integrated into a workflow using its output to make or assist a coverage determination, the exclusion falls away.

03Covered domains and the healthcare hook

SB 26-189 regulates ADMT that materially influences a consequential decision about a consumer in one of seven specified covered domains.

DOMAIN 1
Employment
Hiring, firing, promotion, compensation, performance management, task allocation.
DOMAIN 2
Education
Admissions, evaluation, discipline, and access to educational programs and resources.
DOMAIN 3
Financial services
Credit, lending, insurance pricing, account access. Elimination of federally-regulated-entity exemptions has implications for banks and credit unions previously assumed exempt.
DOMAIN 4
Insurance
Underwriting, pricing, and claims-handling decisions across life, health, P&C, and auto.
DOMAIN 5 — KEY FOR THIS PAPER
Health care
ADMT used to materially influence access to, eligibility for, or compensation in healthcare. Captures payer prior-authorization AI, eligibility-determination logic, clinical-decision-support systems used in coverage decisions, and risk-stratification tools that drive care-management routing.
DOMAIN 6
Residential real estate
Tenant screening, mortgage underwriting, property valuation, eviction risk-scoring.
DOMAIN 7
Essential government services
Public benefits eligibility, government-administered programs, civic services access.

3.1 What "materially influences a consequential decision" means

The mature pre-rulemaking reading is that ADMT that makes the decision autonomously is materially influential; ADMT whose output is typically followed by the human reviewer is materially influential; ADMT used to route, prioritise, or pre-screen in a way that constrains the human reviewer's options is likely materially influential; ADMT used purely advisorily, where the human reviewer routinely overrides or independently evaluates the recommendation, is likely not materially influential. AG rulemaking will define this further.

3.2 The healthcare-deployer reading

For a payer, hospital system, or UM vendor operating in Colorado, three distinct categories of exposure follow. First, prior-authorization AI driving coverage determinations is in-scope. Second, clinical-decision-support systems feeding into coverage or care-management decisions — bed-utilisation models, length-of-stay predictors, discharge-readiness scoring — are likely in-scope when their outputs are operationally treated as decisional rather than advisory. Third, risk-stratification and care-management routing tools are in-scope to the extent they shape access to care.

04Developer and deployer duties

SB 26-189 distributes obligations between developers (entities that create or substantially modify covered ADMT) and deployers (entities that use covered ADMT to make consequential decisions). A single organisation can be both.

4.1 Developer duties

Developers must provide deployers with technical documentation covering: intended uses and unsuitable uses; categories of training data; known limitations and inappropriate uses; instructions for appropriate use including human review. Developers must also notify deployers of material updates and retain records for at least three years.

4.2 Deployer duties

ObligationWhat it means in practice for healthcare deployers
Consumer disclosuresClear notice to the consumer about the use of ADMT in a consequential decision — targeted, not buried in privacy-policy boilerplate.
Post-adverse-outcome explanationsWhen ADMT produces an adverse outcome, a substantive explanation of the decision allowing the consumer to understand its basis.
Correction rightsThe consumer may correct inaccurate personal data that contributed to the adverse outcome, and the deployer must reconsider the decision in light of corrected data.
Meaningful human reviewThe consumer has a right to request human review of an adverse decision — review that is substantive, not a procedural box-check by a reviewer routinely affirming the ADMT output.
The "meaningful" question

"Meaningful human review" maps directly onto the same question California SB 1120 addresses (licensed clinician competent to evaluate the specific clinical issues) and Maryland HB 820 addresses through its individual-clinical-picture requirement. A healthcare deployer that has built its workflow to satisfy California's standard is already most of the way to satisfying Colorado's. A deployer whose workflow features human reviewers rubber-stamping ADMT recommendations is exposed under all three regimes simultaneously.

05The eliminated exemptions — and who gets swept in

SB 24-205 contained conditional exemptions for federally-regulated entities — meaning Medicare Advantage organisations, Medicaid managed care plans, and similar federal-payer entities could plausibly argue exemption. SB 26-189 eliminates these exemptions. The Colorado AI Policy Work Group recommended this on the rationale that those entities are among the most consequential users of ADMT in the covered domains.

5.1 Who is now in scope who arguably was not before

As of mid-2026, no federal rule of the kind that would qualify for the remaining narrow exemption exists in the healthcare-AI context — CMS-0057-F is an infrastructure rule, not a substantive AI regulation — so the practical effect of the exemption is limited.

06Enforcement, the cure period, and the xAI litigation

6.1 AG-only enforcement via the Colorado Consumer Protection Act

The Colorado Attorney General is the sole public enforcer. Violations are channeled through the Colorado Consumer Protection Act, with a violation of SB 26-189 deemed a deceptive trade practice. Available remedies include civil penalties, restitution, disgorgement, and injunctive relief. There is no private right of action.

6.2 The 60-day notice-and-cure period

Before initiating enforcement, the AG must provide 60 days' notice and an opportunity to cure the alleged violation, where cure is deemed possible. The notice-and-cure provision is scheduled to expire January 1, 2030. After that date, the AG may proceed directly to enforcement. Healthcare deployers should plan their compliance posture around the cure-period availability through 2029, and its unavailability from 2030 forward.

6.3 The xAI v. Weiser litigation

xAI LLC v. Weiser, Civil Action No. 1:26-cv-01515-DDD-CYC (D. Colo.), was filed April 9, 2026, raising First Amendment, Dormant Commerce Clause, Due Process, and Equal Protection challenges to SB 24-205. DOJ intervened April 24. The court stayed SB 24-205 on April 27. SB 26-189 is structured as a repeal and replacement — whether the stay reaches the new statute is an open question the courts will need to resolve.

The litigation track to monitor

The first procedural question is whether xAI and DOJ amend their complaint to add SB 26-189 as a target, or whether the parties stipulate to litigation under the new statute. The procedural disposition will likely come within the next 60–90 days. Healthcare deployers should plan to the law as enacted (effective January 1, 2027) and adjust if and when the court intervenes — not delay compliance work in anticipation of judicial relief that may not arrive.

07Timeline and rulemaking

May 2024
SB 24-205 signed
Governor Polis signs the original Colorado AI Act with reservations. Original effective date: February 2026.
2025 session + special
Effective date pushed to June 2026
Legislature could not reach agreement. Special session extended the effective date four months.
17 March 2026
AI Policy Work Group proposal released
Unanimous Work Group proposal becomes the basis for SB 26-189.
9 April 2026
xAI files federal lawsuit
Constitutional challenge to SB 24-205 on First Amendment, Dormant Commerce Clause, Due Process, and Equal Protection grounds.
27 April 2026
Federal court stays SB 24-205 enforcement
US District Court for the District of Colorado stays enforcement of the original Colorado AI Act.
9 May 2026 · Three days ago
SB 26-189 passes both chambers
House passes 57–6. Proceeds to Governor Polis for signature; Governor's office has stated intent to sign.
By Q3 2026
AG mandatory rulemaking commences
Mandatory rulemaking begins on materiality threshold, disclosure format, explanation standard, meaningful-review definition, and remaining exemption scope.
1 January 2027
SB 26-189 effective date
Substantive provisions take effect. AG rulemaking must be complete. ADMT inventories, consumer-disclosure mechanisms, explanation pathways, correction processes, and meaningful-human-review capability must be operational.
January 2028
AG annual reporting begins
First public visibility into Colorado's enforcement pattern.
1 January 2030
Cure period expires
60-day notice-and-cure provision sunsets. AG may proceed directly to enforcement.

08How SB 26-189 fits into the five-regime landscape

Dimension
Colorado SB 24-205 (repealed)
Colorado SB 26-189 (effective 2027)
Regulated subject
High-risk AI systems
Automated decision-making technology (ADMT)
Triggering harm
Algorithmic discrimination
Material influence on consequential decisions
Deployer governance
Risk management programs, impact assessments, annual reviews, public summaries
Disclosures, post-adverse-outcome explanations, correction rights, meaningful human review
Federal-entity exemptions
Limited but available
Eliminated — federal-payer entities now in scope
AG rulemaking
Permissive
Mandatory; deadline January 1, 2027
Private right of action
None, but ambiguities feared
None; ambiguities closed
Cure period
Not specified
60-day notice-and-cure; sunsets January 1, 2030
Effective date
Was February 2026, then June 2026; never operative
January 1, 2027

8.1 What sits at the intersection of all five regimes

The healthcare deployer reading this paper alongside SB 1120, HB 820, CMS-0057-F, and EU AI Act Article 12 will find that the architecture which satisfies all five is one architecture:

09What healthcare deployers should do in the next 90 days

9.1 Build the ADMT inventory and confirm Colorado-operations scope

Begin with a documented inventory of every AI, algorithmic, or rule-based system touching a consequential decision in the healthcare domain for a Colorado consumer — including systems operated by vendors on the entity's behalf. Critically, include systems previously considered exempt under SB 24-205's federally-regulated-entity exemption. Those exemptions are gone. Medicare Advantage plans, Medicaid managed care plans, and CHIP managed care entities operating in Colorado need to assume in-scope.

9.2 Track the AG rulemaking and engage where possible

Mandatory AG rulemaking will produce the most consequential operational guidance before January 1, 2027. The AG's office has historically run rulemaking in a notice-and-comment posture. Healthcare entities operating across multiple state regimes should plan to file comments when notice issues, particularly on the interaction between SB 26-189 and the parallel federal CMS-0057-F infrastructure.

9.3 Architect the per-consumer rights framework now, not in late 2026

The four deployer duties describe operational capabilities, not policy documents. A Colorado-resident enrollee who receives an adverse coverage determination on January 2, 2027 has an immediate right to disclosure, explanation, correction-data-submission, and human-review request. Healthcare deployers should treat the next 33 weeks as the build-and-test window for those processes.

A coordinated 90-day kickoff

Days 1–30: complete the ADMT inventory for Colorado healthcare operations; identify systems previously assumed exempt; map the gap between current capability and the four deployer duties. Days 31–60: begin building the consumer-disclosure, explanation, correction, and human-review processes; engage Colorado counsel on outstanding questions; file comments on AG rulemaking notice if issued. Days 61–90: end-to-end test the consumer-rights workflows on real consequential decisions; reconcile the resulting architecture against the parallel California, Maryland, federal, and (if applicable) EU regimes.

References & citations

  1. Colorado General Assembly. Senate Bill 26-189 — Automated Decision-Making Technology. Passed May 9, 2026; awaits Governor's signature.
  2. Colorado General Assembly. Senate Bill 24-205 — Consumer Protections for Artificial Intelligence (repealed by SB 26-189).
  3. Szewczyk et al. Colorado Rewrites Its Landmark AI Law: Unpacking SB 26-189. Ballard Spahr CyberAdviser, May 11, 2026.
  4. Reed Smith LLP. SB 26-189: Colorado Legislature Kicks Off CAIA Rewrite Race.
  5. Proskauer Rose LLP. Major Developments Put Colorado's AI Law on Ice Ahead of Implementation, May 2026.
  6. Colorado Newsline. New bill would narrow scope of Colorado's landmark 2024 AI law, May 4, 2026.
  7. xAI LLC v. Weiser, Civil Action No. 1:26-cv-01515-DDD-CYC (D. Colo., filed April 9, 2026); DOJ complaint in intervention filed April 24, 2026; stay entered April 27, 2026.

One architecture for the rewrite — and everything after

Predicate ZK is a verifiable AI governance infrastructure that produces per-determination structured evidence — satisfying Colorado SB 26-189's meaningful-human-review and explanation requirements, California SB 1120's clinician-decision standard, Maryland HB 820's individual-clinical-picture standard, CMS-0057-F's denial-reasoning requirements, and EU AI Act Article 12's record-keeping mandate. Built once, deployed across the entire regulatory landscape.

Read the architecture Contact us