Statutory Compliance Architecture for Colorado SB 24-205 Standards
Documenting multi-specialty clinical encounters under Colorado SB 24-205, codified at C.R.S. §§ 6-1-1701 through 6-1-1707, requires that ambient documentation be evaluated as a potential high-risk artificial intelligence system. When a generated note materially influences a consequential decision, such as coding, reimbursement, eligibility, prior authorization, or risk stratification, the deployer assumes a duty of reasonable care to protect patients from algorithmic discrimination. The Attorney General enforces violations as deceptive trade practices under the Colorado Consumer Protection Act.
Deployer obligations are layered and cumulative. Section 6-1-1703(2) requires a documented risk management policy and program that is reviewed throughout the system's lifecycle. Section 6-1-1703(3) requires an impact assessment before deployment, annually thereafter, and within ninety days of any intentional and substantial modification. Section 6-1-1703(4) requires several protections for the consumer: notice, an explanation of adverse decisions, an opportunity to correct inaccurate personal data, and an appeal via human review where technically feasible.
Operative dates have shifted repeatedly. The 2025 extraordinary session moved the effective date to June 30, 2026. Subsequent proposals have contemplated small-business exemptions and revised enforcement timing. Compliance officers should verify the current statutory text before relying on any date. An architecture built to the strictest reading remains defensible under any amendment. Section 6-1-1706 also provides an affirmative defense for deployers aligned with the NIST AI Risk Management Framework or ISO/IEC 42001.
Human-Directed Review Across Specialty Diagnostic Frameworks
Merry AI enforces two-step insertion. Every draft is generated in a side panel and bound to the patient, encounter, note template, and specialty context read from the EHR DOM or URL. Insertion into textarea, contenteditable, or iframe-hosted editors is blocked if the context changes between generation and confirmation. This prevents wrong-chart injection in multi-tab workflows. The clinician's explicit confirmation becomes the human-review artifact on which § 6-1-1703(4) appeals depend.
Specialty criteria are presented as hypotheses, never as assertions. The AI output identifies candidate findings, and the clinician confirms, edits, or rejects each element before the note is committed. This distinction carries statutory weight. An impact assessment must describe known limitations, and fabricated criteria are precisely the failure mode that produces uneven documentation quality across patient populations.
Behavioral Health and Cardiology Criteria Integrity
In behavioral health, DSM-5-TR diagnoses require explicit criteria clusters, onset, duration, severity, and functional impairment, alongside a complete Mental Status Examination. Merry AI does not auto-populate diagnoses from conversation alone. Any inferred content on suicidality, violence, or substance use is labeled AI-suggested and held for mandatory clinician review.
In cardiology, NYHA class and LVEF are never inferred. LVEF is captured as a numeric value with its modality source: echocardiography, cardiac MRI, or nuclear imaging. When no value is stated, the note reads "LVEF not documented." ACC/AHA heart failure stage and CHA₂DS₂-VASc scores follow the same confirmation rule.
Orthopedics and Pediatrics Measurement Fidelity
Orthopedic notes demand measured values. Templates prompt for range-of-motion degrees, manual muscle testing grades from 0 to 5, and named maneuvers such as Lachman, McMurray, and straight leg raise with explicit results. Historical ROM values cannot carry forward into the current encounter without clinician confirmation.
Pediatric well-child visits follow Bright Futures age bands. Growth percentiles, developmental milestones, and screening results for vision, hearing, depression, and autism are surfaced for confirmation. Omitted elements are written explicitly, for example "Hearing screening not performed today," so that completion is never implied.
Audit Logging, Impact Assessments, and CMS Coding Defensibility
Every AI interaction is logged as an append-only, time-stamped event keyed to patient, encounter, note ID, specialty, user, and model version. Logged events include audio capture, transcript, draft generation, each clinician edit, and each DOM write. Inserted content carries machine-readable metadata and a visible note-header marker indicating AI assistance. This design is consistent with the disclosure principles of § 6-1-1704 and supports downstream detection of synthetic content.
These logs feed the annual impact assessment directly. Merry AI stratifies documentation completeness and complexity capture by specialty and demographic cohort. Deployers can therefore detect systematic under-documentation or overstated severity for any protected group before it influences a consequential decision.
G2211 and NCCI Modifier Distinctness
HCPCS G2211 rewards documented longitudinal complexity. CMS denies G2211 when the base office/outpatient E/M service is reported with modifier 25. The exception applies when the same-day service is an Annual Wellness Visit, vaccine administration, or a Medicare Part B preventive service. Beginning January 1, 2026, that exception extends to home or residence E/M codes. The note must articulate continuity, multimorbidity, and care coordination rather than restate the visit.
NCCI 2026 PTP edits require documented distinctness. Modifier 25 requires a significant, separately identifiable E/M service. Modifiers 59, XE, XS, XP, and XU require distinct sessions, sites, lesions, or practitioners. One example is CPT 45385 reported with 45380 for a separate lesion. Merry AI tags the supporting narrative segments so that coders never infer distinctness the clinician did not document. Incorrect coding accounted for 49.1% of 2024 E/M improper payments, so this traceability carries real financial weight.
Statutory Comparison and Implementation Sequence
| Compliance Dimension | Manual Charting | Standard Generic AI Scribes | Merry AI Compliance Architecture |
|---|---|---|---|
| Human review (§ 6-1-1703(4)) | Inherent but unlogged | Optional, often post-insertion | Mandatory two-step confirmation, logged |
| Impact assessment evidence (§ 6-1-1703(3)) | Not applicable | Model accuracy metrics only | Specialty and cohort discrimination analysis |
| AI disclosure (§ 6-1-1704) | Not applicable | Inconsistent or absent | Machine-readable metadata plus visible header |
| Audit trail granularity | EHR timestamp only | Transcript and final note | Append-only event log per DOM write |
| Specialty criteria integrity | Clinician-dependent | Generic language model | DSM-5-TR, NYHA/LVEF, ROM, Bright Futures prompts |
| Modifier distinctness | Coder inference | Unflagged narrative | Tagged segments for 25, 59, XE, XS |
Implementation proceeds in four stages. First, adopt a risk management policy mapped to the NIST AI RMF. Second, complete baseline impact assessments for each specialty. Third, enable context-bound insertion and event logging. Fourth, set the cadence for annual review. Peer-reviewed evidence on ambient documentation accuracy is indexed at https://www.ncbi.nlm.nih.gov/pmc/. To map your EHR stack to these controls, Schedule a 15-Minute Workflow Audit.


