Policy — Artificial intelligence

Colorado SB24-205 obligates both developers and deployers of high-risk AI systems to operate a written risk management program (RMP) that detects, mitigates, and monitors algorithmic discrimination across the lifecycle. Enforcement begins on 1 February 2026, leaving five months to stand up policies aligned with NIST AI RMF or ISO/IEC 42001, embed testing in MLOps, and link consumer notices and appeals to documented safeguards. This brief translating these requirements into control maps, playbooks, and evidence binders that connect with the AI pillar hub, the Colorado compliance guide, and related briefs on high-risk readiness and impact assessments.

What §6-1-1704 demands

Colorado’s law requires the RMP to be documented, risk-based, and iterative. Key elements include:

• Risk identification: Document intended purpose, affected individuals, and potential algorithmic discrimination vectors (for example, protected classes across employment, housing, lending, education, healthcare, insurance).
• Controls and mitigations: Specify technical and procedural safeguards—feature constraints, monitoring thresholds, human-in-the-loop approvals, and override authority—proportional to risk.
• Evaluation and testing: Run pre-deployment and periodic testing for bias, robustness, and drift; record metrics, datasets, and acceptance criteria.
• Governance and accountability: Assign owners, escalation paths, and documentation duties; keep a public statement describing high-risk uses and safeguards.
• Continuous monitoring: Track incidents and material changes; retrain or rollback when triggers fire; retain artifacts for regulatory requests.

Program blueprint and ownership

Structuring the RMP around the NIST AI RMF functions—Map, Measure, Manage, and Govern—paired with clear ownership.

Diagram: How the RMP fits lifecycle

        Map (scope + purpose) → Measure (test & metrics) → Manage (mitigate & deploy)
         ↑ ↓
         └──────────── Govern (policies, roles, oversight, transparency) ───────────┘
         

Embedding into MLOps

To avoid shelfware, the RMP is wired into CI/CD and incident processes:

• Pre-commit checks: Lint model cards for required fields (purpose, datasets, protected-class coverage, limitations); block merges missing risk statements.
• Automated tests: Run fairness and robustness suites during CI; store outputs with model version and dataset hash for reproducibility.
• Release gates: Require legal sign-off after consumer notice copy and appeal SLAs are validated; tag deployments with effective policy version.
• Monitoring hooks: Stream drift and performance metrics to dashboards; alert when protected-class gaps exceed thresholds; trigger retraining or rollback.
• Change control: Log all parameter updates and data refreshes; refresh impact assessments after material changes; notify customers if changes alter risk profile.

Risk scenarios and mitigations

Colorado focuses on consequential decisions. Using scenario playbooks to test and mitigate.

• Employment screening: Bias testing on gender, race, and age with stratified sampling; mask correlating features; require recruiter override before rejection.
• Credit underwriting: Examine disparate impact on protected classes; cap feature influence for proxy attributes; document adverse action rationales in consumer notices.
• Housing eligibility: Validate geographic and demographic balance; prohibit using past eviction data as sole factor; route borderline cases to human review.
• Insurance pricing: Stress-test for disability and age proxies; apply fairness-constrained improvement; explain key factors in notice templates.

Metrics and thresholds

Each high-risk system gets quantitative targets tied to escalation.

Transparency and documentation

Colorado requires a public statement summarizing high-risk uses and safeguards, plus the ability to furnish detailed documentation to the Attorney General. Building a layered package:

1. Public layer: Web page summarizing high-risk systems, purposes, core safeguards, and appeal routes.
2. Customer layer: Consumer notice copy, appeal SLAs, and channel-specific explanations of key factors.
3. Regulatory layer: Full model cards, assessment reports, testing evidence, incident history, and developer attestations.

Integration with developers

Because the statute allocates duties to developers (§6-1-1705) and deployers (§6-1-1706), the RMP includes vendor coordination:

• Require developers to provide intended use, data sources, limitations, evaluation metrics, and known risks.
• Negotiate contractual commitments for timely notice of material changes and incident cooperation.
• Share The mitigation playbooks so developer updates align with deployer obligations.

Incident response and AG notification

An RMP without incident muscle will fail audits. this brief codifies a 72-hour internal investigation and 90-day AG notification plan with roles and templates. Evidence includes root-cause analyzes, consumer communication scripts, mitigation steps, and verification of restored performance.

Training and attestation

SB24-205 expects organizational buy-in. Running quarterly training for engineers, product managers, customer-support leads, and legal reviewers, followed by attestations that policies were read and applied. Completion metrics feed the board dashboard and readiness attestation.

Documentation

• Colorado SB24-205 — Artificial Intelligence Act
• Colorado Attorney General AI Act Fact Sheet
• NIST AI Risk Management Framework
• ISO/IEC 42001:2023 Artificial Intelligence Management System

Aligning Colorado risk management programs with actionable tests, notices, and incident response so compliance is auditable and repeatable.

Roadmap to February 2026

This brief guides clients through a structured cadence: September focuses on finalizing the policy and governance charter; October centers on embedding controls into CI/CD and running dry-run assessments; November publishes transparency statements and trains customer-facing teams; December completes incident tabletops, validates evidence retention, and captures board acknowledgement ahead of go-live.

Audit trail and retention

Evidence must survive regulatory review. Retaining signed policies, model cards, test datasets, metric outputs, consumer notice versions, appeal logs, and incident reports for at least three years, tagged to model version and deployment date. Each artifact is cross-referenced in an index so legal teams can prove which safeguards were active for any decision.

Crosswalk to other frameworks

• EU AI Act Article 9 risk management: Colorado’s §6-1-1704 dovetails with Article 9 requirements—teams can reuse hazard identification, testing, and monitoring steps with localized notice language.
• FTC unfairness guidance: Documentation of foreseeable harms and mitigations supports U.S. federal expectations for deceptive or unfair AI practices.
• ISO/IEC 42001 clauses: Policies, roles, awareness training, monitoring, and continual improvement map directly to management-system controls for certification-ready organizations.

KPIs for board oversight

Board and executive sponsors monitor a small set of KPIs each quarter: percentage of high-risk systems with current assessments; number of policy deviations with corrective actions; appeal volumes and reversal rates by product line; time-to-close for incident investigations; and training completion by role. These KPIs are paired with confidence notes explaining data lineage and any blind spots.

Lessons from pilots

Pilot setups show that early pairing of legal and data science cuts assessment rework by half, and publishing draft notices alongside model cards surfaces clarity gaps before launch. This brief catalogs these lessons and reuses them as guardrails for subsequent models to keep the RMP living and adaptive.

Policy Development and Analysis

Policy analysis should assess the implications of this development for organizational operations, compliance obligations, and strategic positioning. Impact assessments should consider both direct requirements and indirect effects through industry practices, customer expectations, and competitive dynamics.

Policy development processes should engage relevant teams to ensure full consideration of diverse perspectives and practical setup constraints. Feedback mechanisms should capture lessons learned and drive policy refinements based on operational experience.

Policy Implementation Monitoring

Policy teams should track setup progress and monitor for developments that may affect requirements or interpretation. Stakeholder engagement should ensure relevant parties understand policy implications and their responsibilities for compliance. Documentation should support audit and examination processes by demonstrating timely awareness and appropriate response to policy developments.

Regular reviews should assess ongoing compliance status and identify any gaps requiring additional attention or resource allocation.

Bias Testing and Impact Assessment

Colorado AI Act risk management requirements emphasize algorithmic discrimination prevention through systematic bias testing and impact assessment. Testing protocols should cover protected class disparate impact analysis across relevant decision contexts. Documentation of testing methodologies, findings, and remediation actions supports compliance evidence.

Impact assessment processes should engage affected stakeholder perspectives and consider cumulative effects of AI system deployment. Assessment findings inform risk mitigation strategies and ongoing monitoring approaches that maintain compliance as systems operate in production environments.

Human Oversight and Intervention Capabilities

Risk management programs must ensure meaningful human oversight of high-risk AI decision-making. Technical setups should support human review of AI outputs before consequential decisions, with override capabilities for identified errors or concerns. Staff training on oversight responsibilities and intervention procedures maintains effective human control.

Documentation of human oversight activities, intervention frequency, and outcome corrections shows operational compliance with oversight requirements. Metrics on human review patterns inform ongoing system refinement and risk management improvements.

Bias Testing and Impact Assessment

Human Oversight and Intervention Capabilities

Risk Mitigation Control Implementation

Identified risks require appropriate mitigation controls proportionate to potential harm. Technical controls, procedural safeguards, and organizational measures should address specific risk factors documented in impact assessments. Control effectiveness monitoring validates that implemented measures achieve intended risk reduction.

Defense in depth approaches layer multiple controls addressing similar risks. Redundant safeguards provide resilience against individual control failures and show strong risk management.

Incident Response and Remediation

Risk management programs should include incident response procedures for AI system failures, discrimination complaints, and unexpected adverse outcomes. Rapid response capabilities limit harm from incidents while remediation processes address root causes and prevent recurrence.

Incident documentation supports regulatory compliance demonstrations and continuous improvement. Lessons learned from incidents inform risk assessment updates and control refinements.

Ongoing Monitoring and Model Governance

Risk management extends throughout AI system lifecycles, requiring ongoing monitoring of system performance, fairness metrics, and emerging risks. Model governance frameworks establish clear responsibilities for monitoring activities and response to identified concerns.

Regular risk reassessment ensures that risk management remains current as systems evolve, data distributions shift, and deployment contexts change. Adaptive risk management supports sustained compliance as circumstances develop.