Clinical Update — June 2026: This Operations Playbook has been revised to reflect the AMA's May 2026 CPT Appendix S taxonomy update for AI-assisted clinical services, CMS CY2026 Physician Fee Schedule final rule adjustments to E/M documentation thresholds, and updated FHIR R4 Provenance binding specifications. All MDM scoring references align with the current AMA E/M office visit guidelines effective January 1, 2023, as clarified through the 2026 CPT errata cycle. ICD-10-CM code references reflect FY2026 addenda.

AI Medical Coder: Automating Level 4 & Level 5 E/M Capture With Real-Time MDM Auditing

TL;DR — Why Primary Care Medical Directors Should Read This

Roughly 30% of office visits are under-coded because physicians avoid documenting the "Complexity of Data Reviewed" element in Medical Decision Making. AMA 2023+ E/M rules demand granular, auditable proof—counting unique tests, identifying external notes by distinct source, and explicitly documenting independent interpretation and specialist discussions. Scribing.io's real-time MDM Ledger automates this granularity: de-duplicating labs, normalizing external sources by NPI, diarizing phone consults, and binding every data point to FHIR Provenance for a CMS-compliant audit packet. The result is defensible Level 4 (99214) and Level 5 (99215) capture without upcoding risk, closing the revenue gap that under-coding creates across your practice.

• The Under-Coding Crisis: Why 30% of Visits Are Billed Below Clinical Work

• What the AMA Taxonomy Misses: The Granular MDM Documentation Gap

• Scribing.io Clinical Logic: High-Stakes Outpatient MDM in Real Time

• The MDM Ledger Architecture: Real-Time Data Deduplication and Source Normalization

• Technical Reference: ICD-10 Documentation Standards for Complex Multi-Condition Visits

• Level 4 vs. Level 5 E/M: A Granular MDM Scoring Comparison

• Audit-Ready by Design: FHIR Provenance and the CMS 6-Year Lookback

• Implementation Roadmap for Primary Care Medical Directors

The Under-Coding Crisis: Why 30% of Visits Are Billed Below Clinical Work

Every Primary Care Medical Director has seen it: a physician finishes a 25-minute encounter managing a patient with three chronic conditions, reviews an outside ED note, interprets an ECG, calls a specialist—then bills a 99214 because they cannot articulate why the data complexity justifies a higher level. The note reads "outside notes reviewed, labs reviewed" and the physician moves on, leaving defensible revenue on the table.

This is not a knowledge problem. It is a documentation friction problem. And it is precisely the problem Scribing.io was built to eliminate.

Current clinical benchmarks—consistent with findings reported in JAMA Internal Medicine research on E/M coding accuracy—indicate that approximately 30% of outpatient visits are under-coded because clinicians are reluctant to detail the "Complexity of Data Reviewed" element within Medical Decision Making (MDM). The reluctance is rational: under the AMA's 2023+ E/M framework, "data reviewed" is not a free-text narrative flourish. It requires specific, countable, source-attributed evidence that can survive a payer audit. When the documentation burden of proving what was reviewed exceeds the cognitive bandwidth available between patients, physicians default to the safer, lower-level code.

The downstream financial impact is substantial. For a 10-physician Family Medicine practice averaging 25 patient encounters per physician per day, a systematic 30% under-coding rate on visits that clinically qualify for 99215 translates to a revenue shortfall exceeding $300,000 annually—without any change in clinical work performed. Across a health system with 50 primary care physicians, that number approaches seven figures. These are not hypothetical dollars gained by upcoding. They are real dollars lost by under-documenting work that was already performed.

The pattern extends beyond primary care. We see identical documentation avoidance in Psychiatry, where complex medication management visits with external collateral review routinely bill at 99213 or 99214 despite meeting 99215 data thresholds. The mechanism is the same everywhere: the physician does the work, the note fails to prove it, and the code drops.

The solution is not "code higher." The solution is to make the documentation granularity that justifies the correct code automatic, auditable, and embedded in the clinical workflow so physicians never have to choose between documentation fidelity and patient care velocity. See our real-time MDM Ledger that auto-binds Category 1/2/3 data to FHIR Provenance and one-click exports an auditor-ready packet into Epic/Cerner.

What the AMA Taxonomy Misses: The Granular MDM Documentation Gap

The AMA's CPT Appendix S taxonomy—updated May 2026—provides a useful classification framework for AI in medical services, distinguishing between assistive, augmentative, and autonomous software outputs. It addresses a fundamentally different problem than the one that drives under-coding in primary care.

Appendix S classifies what AI does. It does not address how AI helps physicians prove what they did.

This is the critical gap. The AMA taxonomy asks: "Is this software assistive, augmentative, or autonomous?" The Primary Care Medical Director asks: "Can this software ensure my physicians capture the correct E/M level for the clinical work they already perform—and survive an audit?"

Most published guidance—including the AMA's own E/M documentation resources and CMS's Evaluation and Management resources—treats "data reviewed" as a monolithic concept. A typical instruction reads: "Document the data you reviewed and ordered." But the 2023+ MDM table demands far more specificity.

The Three Granularity Requirements Most Resources Ignore

1. Unique Test Counting, Not Panel Subcomponent Counting

When a physician orders a Comprehensive Metabolic Panel (CMP), that is one test for MDM purposes—not 14 individual results. However, if a separately ordered INR is also reviewed, that is a second unique test. The distinction matters because the threshold between Moderate and High data complexity can hinge on the count of unique tests reviewed or ordered. Most EHR templates fail to enforce this distinction, leading to either under-counting (treating CMP + INR as "labs reviewed") or over-counting (listing every CMP subcomponent as a separate data point, which is indefensible under audit).

2. External Source Identification by Unique Origin

Reviewing "outside records" is not a data point. Reviewing a discharge summary from St. Mary's Hospital Emergency Department (Source 1) and an office note from Dr. Patel's cardiology practice (Source 2) is two external sources. The AMA 2023+ framework requires each external source to be identified distinctly. A PDF scan sitting in an EHR media tab does not automatically constitute "reviewed" data unless the note explicitly references its content and origin. This requirement is detailed in the AMA's E/M descriptors and guidelines.

3. Independent Interpretation and External Physician Discussion

Two of the highest-value MDM data activities—independent interpretation of a test (e.g., a physician personally reading an ECG rather than relying on the machine read) and discussion with an external physician about management—require explicit documentation of the interpretation rendered or the physician contacted, the content discussed, and the time. A note that reads "discussed with cardiology" is insufficient. A note that reads "Discussed warfarin management with Dr. Sarah Chen, cardiology (NPI: 1234567890), at 2:47 PM for 2 minutes; agreed to hold warfarin pending repeat INR Thursday" qualifies for Category 3 credit.

Why Post-Encounter AI Coding Falls Short

Most AI medical coding solutions operate as post-encounter auditors: the physician writes the note, and the software suggests a code afterward. This approach is better than nothing, but it cannot solve the granularity problem because the granular data was never captured in the note in the first place. You cannot extract a Category 3 phone consult detail from a note that says "discussed with cardiology." You cannot count unique external sources from "outside notes reviewed." The data gap exists at the point of documentation, not at the point of code selection.

Scribing.io operates differently—as a real-time MDM Ledger embedded in the encounter workflow.

Scribing.io Clinical Logic: High-Stakes Outpatient MDM for a 67-Year-Old With Type 2 Diabetes and New Atrial Fibrillation

Consider the encounter that epitomizes the under-coding problem:

Patient: A 67-year-old male with type 2 diabetes mellitus (E11.65) and newly diagnosed atrial fibrillation (I48.91) on warfarin presents to an internist's office with palpitations. He was seen in the ED three days ago and had a cardiology consult from an outside practice. The internist reviews the ED records, reviews the cardiology consult, personally interprets an in-office ECG, calls the cardiologist to discuss anticoagulation management, reviews the patient's CMP and a separately ordered INR, and adjusts the warfarin dose with a monitoring plan.

What typically happens: The internist bills 99214 (Level 4, Moderate MDM). The note reads: "Reviewed outside notes and ECG. Labs reviewed. Discussed with cardiology. Warfarin dose adjusted." The physician knows this was a complex visit but documents conservatively because "outside notes and ECG reviewed" feels risky to defend at a higher level. The practice loses the 99215–99214 differential—roughly $60–$80 per encounter depending on payer mix—multiplied across every similar visit, every day.

What happens with Scribing.io's live MDM Ledger:

Step-by-Step MDM Ledger Workflow

The MDM Level Determination

With these data elements captured and attributed, the MDM Ledger scores the encounter across all three MDM axes:

• Number and Complexity of Problems Addressed: Two chronic conditions — type 2 diabetes with hyperglycemia (E11.65) requiring ongoing medication management, and new atrial fibrillation (I48.91) with initiation of anticoagulation. A new problem requiring a prescription drug qualifies as Moderate to High. Combined with the chronic illness requiring drug management = High.

• Amount and/or Complexity of Data Reviewed: Two unique external sources (Category 1 × 2) + independent interpretation of ECG (Category 2) + external physician discussion with documented name/NPI/time/content (Category 3) + two unique lab tests reviewed (Category 1). Per the AMA MDM data table, achieving Category 3 credit or the combination of Category 1 + Category 2 qualifies as Extensive data complexity. This encounter achieves both.

• Risk of Complications and/or Morbidity or Mortality: Warfarin is a prescription drug requiring intensive monitoring for toxicity (narrow therapeutic index, bleeding risk, INR variability). This alone qualifies as High risk per the CMS/AMA risk table.

Result: Two of three MDM elements meet High complexity. Under the 2023+ E/M rules, where MDM level is determined by the two highest of three elements, this encounter qualifies as 99215 — High MDM.

Without Scribing.io, the same clinical work generates a note that supports only 99214 because the documentation lacks the granularity to prove High data complexity. The physician did the work. The note did not capture it. The code drops.

With Scribing.io, the live MDM Ledger produces a clean 99215 with an auditor-ready packet attached directly into Epic or Cerner—denial risk reduced, revenue per visit aligned to clinical work performed.

The MDM Ledger Architecture: Real-Time Data Deduplication and Source Normalization

The MDM Ledger is not a charting template or a coding suggestion pop-up. It is a parallel data structure that runs alongside the encounter, ingesting structured and unstructured clinical data from multiple channels and normalizing it against AMA E/M counting rules in real time.

Four Core Technical Capabilities

1. Lab De-Duplication Engine

The engine maps incoming lab results to their parent order using LOINC codes and the ordering facility's compendium. A CMP (LOINC panel 24323-8) is recognized as a single orderable test regardless of how many individual analytes the EHR displays. A separately ordered PT/INR (LOINC 6301-6 / 34714-6) is counted as an additional unique test. This prevents both under-counting ("labs reviewed") and over-counting (listing 14 CMP components individually), either of which creates audit vulnerability.

2. External Source Normalization

External documents arrive in EHRs through multiple pathways: Direct messaging, fax-to-EHR conversion, patient portal uploads, and HL7 v2 MDM messages. Many land in the Media tab as unstructured PDFs—not exposed through standard FHIR R4 DocumentReference APIs. Scribing.io addresses this by:

• Ingesting HL7 v2 MDM/TXA segments and XDS.b metadata where available

• Extracting sending facility OID, provider NPI, and document type from message headers

• Creating normalized FHIR DocumentReference resources with unique author and custodian references

• Presenting each source to the physician as a discrete, attributable data item in the MDM Ledger rather than a generic "outside records" blob

3. Phone Consult Diarization

Ambient audio capture—already active for encounter documentation—extends to phone consultations when the physician initiates a specialist call during the visit. The diarization engine separates speakers, identifies the external physician by matching voice segments against known call context (outbound number, scheduled callback), and extracts:

• Specialist name and NPI (cross-referenced against CMS NPPES)

• Call start time and duration

• Discussion content summarized into a structured Category 3 data element

The physician reviews and confirms the diarized summary before it is committed to the note. This is not autonomous documentation—it is augmentative capture that requires physician attestation, consistent with AMA Appendix S augmentative AI classification.

4. FHIR Provenance Binding

Every data item in the MDM Ledger—each external source, each lab result, each independent interpretation, each phone consult—is linked to a FHIR Provenance resource that records:

• agent: The physician who reviewed or interpreted the data

• entity: The source document or test result (by FHIR resource ID)

• recorded: Timestamp of review/interpretation

• activity: The type of MDM activity (review, independent interpretation, external discussion)

This creates a tamper-evident, exportable audit chain that persists independently of the narrative note—critical for compliance with CMS's 6-year overpayment lookback window under the 60-day repayment rule.

Technical Reference: ICD-10 Documentation Standards for Complex Multi-Condition Visits

Accurate E/M leveling depends on accurate problem identification, and accurate problem identification depends on ICD-10-CM code specificity. Vague or unspecified codes weaken MDM justification in two ways: they reduce the documented complexity of problems addressed, and they trigger payer edits that slow or deny reimbursement.

For the clinical scenario described above, the relevant codes are:

E11.65 - Type 2 diabetes mellitus with hyperglycemia; I48.91 - Unspecified atrial fibrillation

How Scribing.io Ensures Maximum Code Specificity

E11.65 — Type 2 diabetes mellitus with hyperglycemia: Scribing.io's real-time AI Coder cross-references the problem list, active medications (metformin, insulin), and lab values (glucose, HbA1c) against the ICD-10-CM Official Guidelines for Coding and Reporting. If the physician documents "diabetes" without specifying type, hyperglycemic status, or complications, the system prompts for clarification before note finalization. This prevents the common fallback to E11.9 (Type 2 diabetes mellitus without complications), which understates disease complexity and weakens the "Number and Complexity of Problems" MDM element.

I48.91 — Unspecified atrial fibrillation: While I48.91 is the correct code when AF type (paroxysmal, persistent, permanent) has not yet been determined—as in a new diagnosis—Scribing.io flags this as a specificity opportunity. If the ED note or cardiology consult characterizes the AF as paroxysmal (I48.0), persistent (I48.1), or chronic/permanent (I48.2/I48.21), the system suggests upgrading specificity based on available external documentation. This prevents both under-specification (which invites payer queries) and over-specification (assigning a type not supported by the record).

Multi-Condition Documentation Integrity

For encounters involving multiple chronic conditions, Scribing.io enforces ICD-10-CM Guideline I.B.8 (code to highest specificity) and Guideline IV.J (sequencing for coexistence): the code most relevant to the reason for the visit is listed first, with additional codes reflecting all conditions actively managed. The system also cross-references the CMS ICD-10-CM code edits for mutually exclusive code pairs, manifestation/etiology sequencing requirements, and age/sex conflicts—catching errors that would otherwise trigger claim denials days or weeks after the encounter.

Level 4 vs. Level 5 E/M: A Granular MDM Scoring Comparison

The financial and clinical distinction between 99214 and 99215 is significant but frequently misunderstood. The table below maps the specific MDM thresholds that differentiate these levels, with the clinical scenario data overlaid.

At scale, that $60 per-encounter delta applied to the estimated 30% of under-coded visits across a mid-size primary care group represents the single largest recoverable revenue opportunity that does not require seeing more patients, hiring more staff, or changing clinical protocols.

Audit-Ready by Design: FHIR Provenance and the CMS 6-Year Lookback

Defensible coding is not just about getting the level right at the time of billing. It is about maintaining a retrievable, tamper-evident record that proves the level was justified years after the encounter. CMS's 60-day repayment rule (42 CFR § 401.305) establishes a 6-year lookback window for overpayment identification—meaning a 99215 billed today must be defensible through 2032.

The Narrative Note Problem

Most E/M defense today relies on the narrative clinical note. But narrative notes are fragile audit artifacts:

• They can be amended (raising questions about contemporaneous documentation)

• They do not inherently prove when data was reviewed (only that the physician wrote "reviewed")

• They do not link to the source documents that were purportedly reviewed

• They cannot independently verify that a phone consult occurred at the stated time

How the MDM Ledger Solves This

Scribing.io's FHIR Provenance binding creates a parallel evidence chain that supplements the narrative note with structured, timestamped, source-linked metadata:

This packet is exportable as a single, structured document via one-click integration into Epic (as a SmartData Element or Note Addendum) or Cerner (as a PowerNote component). The auditor does not need to reconstruct the MDM from a narrative note. They receive a structured, timestamped, source-linked evidence file that maps directly to the AMA MDM table.

Implementation Roadmap for Primary Care Medical Directors

Deploying real-time MDM auditing is a clinical operations initiative, not an IT project. The following roadmap reflects implementation patterns from Scribing.io deployments across primary care, internal medicine, and multi-specialty practices.

Phase 1: Baseline Audit (Weeks 1–2)

• Pull a random sample of 200 encounters per physician from the prior 90 days

• Compare billed E/M level against documented MDM elements using the AMA 2023+ MDM table

• Quantify the under-coding rate: what percentage of encounters documented sufficient clinical complexity for a higher E/M level but were billed lower?

• Identify the specific documentation gap: In nearly all cases, it will be "Amount and/or Complexity of Data Reviewed" — the element that requires the granularity described above

Phase 2: MDM Ledger Configuration (Weeks 2–4)

• Configure Scribing.io's EHR integration (Epic via FHIR R4 + HL7 v2 ADT/MDM feeds; Cerner via FHIR R4 + Millennium Open APIs)

• Map practice-specific lab compendium to LOINC for accurate panel de-duplication

• Configure external source ingestion pathways (Direct messaging, fax-to-EHR, patient portal uploads)

• Set ambient audio capture parameters for phone consult diarization (microphone placement, consent workflow per state law)

Phase 3: Physician Onboarding (Weeks 3–5)

• Train physicians on the MDM Ledger interface: where it surfaces during the encounter, how to confirm/edit auto-captured data elements, how to enter the 1-line independent interpretation prompt

• Run 10 encounters per physician in "shadow mode" — MDM Ledger runs and generates the audit packet, but does not alter the submitted code. Compare Ledger-recommended code vs. physician-selected code to quantify the capture gap in real time

• Review shadow-mode results with each physician individually. The goal is not to say "you should have coded higher" but to show "here is the data you reviewed that your note did not capture — and here is what the Ledger would have documented for you"

Phase 4: Go-Live and Continuous Monitoring (Week 6+)

• Activate real-time MDM Ledger with AI Coder recommendations visible to the physician before note finalization

• Physician retains full authority over final code selection — the system recommends, it does not auto-submit

• Monitor weekly dashboards: E/M level distribution shift, denial rate for 99215 claims, audit packet generation rate, physician adoption metrics

• Conduct monthly compliance reviews: random sample of Ledger-supported 99215 claims reviewed by a certified coder to validate ongoing accuracy

Expected Outcomes (90-Day Benchmark)

These are not projections based on upcoding. They reflect the revenue recovery that occurs when documentation accuracy matches clinical work already performed. The NIH and JAMA literature consistently supports the finding that physician documentation under-represents clinical complexity, particularly in data-intensive encounters. Scribing.io does not change the clinical work. It makes the documentation tell the truth about the clinical work.

The core principle is straightforward: doctors under-code 30% of visits because they don't want to document the "Complexity of Data Reviewed." Scribing.io's AI Coder audits the note in real time to ensure the CPT code matches the clinical work performed. That is the entire value proposition—no more, no less.