Posted on
Sep 3, 2026
CMS HTI-1 Transparency: The 'Algorithm Disclosure' Rule Explained
TL;DR — HTI-1 Algorithm Disclosure for Clinical AI Scribes
The mandate in plain terms: The 2026 HTI-1 "Transparency Rule" requires AI tools to expose Human-Readable Logic—inputs, rules, thresholds, and outputs—not just a narrative summary.
The gap most vendors miss: Competing scribes can only produce a rendered note. When a Medicare MAC opens a TPE probe, a narrative cannot prove why an AI justified a code.
The Scribing.io answer here: Our Clinical Logic Bridge writes each AI finding as a FHIR R4
GuidanceResponse(linked to a versionedLibrarywith CQL summary +dataRequirements), pairs it withAuditEvent/Provenancereferencing the source audio via millisecond offsets, and finalizes aDocumentReferencetyped LOINC 34109-9 (Note).The outcome you can measure: Auditors replay the exact audio segment where longitudinal complexity was identified, match it to the triggering logic, and close probes without downcoding or recoupment.
What HTI-1 Algorithm Disclosure Demands
Surviving a MAC TPE Probe on G2211
The Clinical Logic Bridge Architecture
Implementation for Operations Directors
Pricing and Next Steps
What CMS HTI-1 "Algorithm Disclosure" Actually Demands of Clinical AI Scribes
CLINICAL UPDATE 2026: Revised for new CMS CPT G2211 standards, SB 1120 compliance, and FHIR interoperability.
For a Clinical Operations Director, the phrase "algorithm disclosure" carries a specific regulatory weight in 2026. The HTI-1 Transparency Rule reframes AI-assisted documentation from a black-box convenience into an auditable decision system. Its central expectation is Human-Readable Logic: any AI tool touching the record must expose the inputs it consumed, the rules and thresholds it applied, and the outputs it produced. Scribing.io was built around this exact obligation.
It is instructive to examine how CMS treats "algorithm" disclosure in an adjacent, more mature transparency regime. In the Hospital Price Transparency rules, CMS draws a hard line: a charge that can be resolved to a single dollar amount must be encoded as a dollar amount, while a charge that "depends on future, variable, or unknown factors" must be encoded as an explicit algorithm. Medical AI Scribing tools like Scribing.io now inherit that same philosophy.
The underlying CMS lesson is not about pricing at all. It is a consistent institutional stance: when an output is the product of variable inputs, the formula itself must be disclosed, not just the result. That is the intellectual bridge from price transparency to Ambient Clinical Intelligence.
Clinical AI scribes now face the same philosophy applied to coding and documentation. A rendered Assessment & Plan is the "single dollar amount"—the resolved output. The logic that justified a complexity-driven code like G2211 is the "algorithm," and under HTI-1 it must be exposed on demand.
This aligns directly with the data-portability expectations in the Scribing.io Onc Hti 2 Compliance Clinical Ai Scribes Reference and the machine-readable export requirements documented in the Scribing.io Onc Hti 2 Compliance Ai Scribes Data Portability Rule Reference.
CMS's Consistent "Disclose the Formula" Philosophy Across Regimes | ||
Regime | Resolved Output (the "result") | Disclosable Logic (the "algorithm") |
|---|---|---|
Hospital Price Transparency | Payer-specific negotiated charge: dollar amount | Encoded algorithm string when charge depends on variable factors |
HTI-1 Clinical AI Scribes | Rendered note / assigned code (LOINC 34109-9) | Human-Readable Logic: inputs, rules, thresholds, outputs |
Scribing.io Clinical Logic: Surviving a MAC TPE Probe on G2211
Consider the scenario that keeps Clinical Operations Directors awake: a Medicare MAC initiates a Targeted Probe and Educate audit after detecting elevated G2211 usage alongside 99214 visits in a multi-specialty clinic. G2211 is the visit-complexity add-on for longitudinal, continuous care of a serious or complex condition. Because it is high-frequency and easy to over-apply, it is a natural MAC flag.
The plan then does what HTI-1 now empowers it to do: it requests the vendor's HTI-1 algorithm disclosure for the AI scribe that structured the Assessment & Plan. This request is where the market bifurcates between tools that recorded and tools that reasoned.
A competing tool responds with a narrative note—prose describing the visit. Prose is not disclosure. It cannot demonstrate which inputs triggered the G2211 justification, which threshold was crossed, or where in the encounter the complexity was established. In a TPE, an unprovable justification is a downcoding and recoupment event.
The Clinical Logic Bridge Export
Scribing.io exports a FHIR R4 bundle instead of a paragraph. Each element is machine-verifiable and traces back to the source audio. The two chronic conditions at the center of the complexity claim— E11.9 (ICD-10-CM) and I10 (ICD-10-CM) —are captured as discrete findings, not adjectives.
Scribing.io Clinical Logic Bridge — FHIR Bundle Components for TPE Defense | ||
FHIR Resource | Role in Audit Defense | What the Auditor Sees |
|---|---|---|
| Represents each AI-derived clinical finding | The specific finding plus its |
| Discloses the actual rule set and thresholds applied | Versioned logic mapping documented complexity to G2211 eligibility |
| Establishes who/what generated the finding | Attribution to the AI agent and the source |
| Links the finding to the exact audio timecode | The precise timestamp range where complexity was identified |
| Types and anchors the finalized rendered note | The human-facing Note, correctly coded and logic-linked |
How the Probe Closes
The auditors replay the exact audio segments flagged by the AuditEvent offsets—the moments where the clinician discussed diabetes and hypertension medication titration and coordinated ongoing care. They then match those segments to the GuidanceResponse and its Library-linked logic that triggered the G2211 justification.
Because the audio, the finding, the rule, the threshold, and the finalized note form an unbroken chain, the justification is provable rather than asserted. The probe closes without downcoding and without recoupment.
Before you model the exposure of a failed versus defended probe across your specialties, run the numbers with the AI Medical Scribe ROI Calculator.
The Clinical Logic Bridge: Closing the Gap Others Leave Open
The transparency gap in the current market is not that vendors refuse to disclose—it is that they disclose the wrong artifact. They surface the rendered note as if it were the logic. Under HTI-1's Human-Readable Logic standard, that hides the very formula CMS wants exposed.
Scribing.io's architectural insight is to treat every AI-derived finding as a first-class, disclosable decision object. Each finding becomes a FHIR R4 GuidanceResponse that references a versioned Library carrying the CQL summary and dataRequirements, paired with an AuditEvent/Provenance that links back to the original audio via millisecond offsets.
The finalized DocumentReference is typed with LOINC 34109-9 (Note). This is the Clinical Logic Bridge, and it is what most tools miss because they never stored the reasoning in the first place.
The Four Disclosure Dimensions HTI-1 Requires — And Where Competitors Fall Short | ||
Disclosure Dimension | Clinical Logic Bridge Artifact | Typical Vendor Output |
|---|---|---|
Inputs |
| Not itemized; buried in prose |
Rules | CQL summary in the | Undisclosed proprietary black box |
Thresholds | Evaluation logic referenced by | Implicit and unprovable |
Outputs |
| Narrative note only |
Traceability to source |
| None; no link back to encounter audio |
Implementation Checklist for the Clinical Operations Director
Operational readiness for HTI-1 disclosure is not a compliance afterthought—it is a documentation-integrity discipline. The following sequence prepares a multi-specialty clinic before, not during, a probe.
Inventory your G2211 exposure first, flagging every specialty where the add-on pairs frequently with 99214 or 99215 established-patient visits.
Confirm versioned Library retention, ensuring the CQL logic active on each encounter date is retrievable, not overwritten by later model updates.
Validate audio offset integrity, testing that
AuditEventmillisecond ranges resolve to the correctMediasegments on replay.Standardize note typing on LOINC 34109-9, so every finalized
DocumentReferenceis uniformly discoverable during export.Rehearse a mock TPE export, generating a full FHIR bundle and having a coder verify the finding-to-audio chain end to end.
The distinction that matters most for your specialty mix is documented in the Scribing.io specialty coverage, where complexity-coding patterns differ sharply between primary care, cardiology, and endocrinology.
State-level obligations layer on top, including California's SB 1120 requirements governing AI in utilization and clinical decisions—reviewed against scribe workflows in the Scribing.io AI scribe law library.
Pricing, Integration, and Next Steps
The cost of a Clinical-Grade Scribing platform should be measured against the recoupment it prevents, not the per-provider line item alone. A single successfully defended G2211 probe frequently offsets an annual deployment.
Integration into existing EHR workflows follows the FHIR-native path detailed in the Scribing.io integration directory, so the disclosure bundle exports without a bolt-on middleware layer.
Review deployment tiers and modeling against your provider count on the Scribing.io Pricing & Plans page, then pair it with the earlier ROI analysis to build your internal business case.
The core principle to carry forward is simple: under HTI-1, a note that cannot explain itself is a liability. Ambient Clinical Intelligence that exposes its logic is the defensible standard.
