Which AI tools are recommended for ICD-10-CM coding? | CLAIRE

Table of Contents

1. Why AI is changing how coders assign ICD-10-CM codes
2. What AI coding tools actually deliver: accuracy, speed, and revenue
3. How AI reads a chart: NLP, machine learning, and large language models
   1. Natural language processing extracts the clinical facts
   2. Machine learning ranks candidate codes
   3. Large language models read the full note in context
4. Recommended AI tools for ICD-10-CM coding at a glance
5. How the leading AI coding solutions compare
   1. Corti
   2. MediCodio
   3. Nym
   4. AAPC Codify
   5. 3M 360 Encompass
   6. MAIA
   7. RapidClaims
   8. Twofold
   9. CombineHealth
6. What to look for when evaluating an AI coding tool
7. A practical path from shortlist to implementation
8. Compliance, ethics, and the role of human oversight
9. Where AI medical coding goes next
10. FAQs about AI tools for ICD-10-CM coding
    1. How accurate is AI for ICD-10-CM coding?
    2. Does AI replace human medical coders?
    3. How is AI coding different from EHR suggestions and CAC tools?
    4. Do AI coding tools keep up with annual ICD-10-CM updates?
    5. What should I look for when choosing an AI ICD-10-CM coding tool?
11. Choose a tool that shows its reasoning

---

Which AI tools are recommended for ICD-10-CM coding?

Why AI is changing how coders assign ICD-10-CM codes

If you are asking which AI tools are recommended for ICD-10-CM coding, start with independent evidence, not vendor marketing. A 2025 systematic review of 35 studies found that large language models matched or exceeded professional coders on common diagnosis codes but dropped sharply on rare ones. Most vendor accuracy figures are self-reported, not independently audited. This guide pairs the leading tools with the academic data you need to separate validated performance from marketing claims.

ICD-10-CM contains roughly 95,000+ codes, and CMS and NCHS update the code set every October 1 on a fiscal-year cycle. Manual coding struggles to keep pace with documentation volume. Traditional computer-assisted coding (CAC) tools use a spectrum of NLP methods, from basic keyword and dictionary matching to pattern matching, statistical models, and symbolic rules. While the simplest tier surfaces keyword-matched codes, more advanced CAC engines parse phrases and apply coding rules, though they still struggle with negation, context, and full-note comprehension. Modern AI tools use natural language processing to extract conditions, laterality, episode of care, and complications from progress notes and discharge summaries, then suggest codes with confidence scores.

The coder shortage and burnout crisis drives adoption. Coders are being asked to do more with less, and AI-assisted workflows reduce the bottleneck of peer consultations and manual lookups. You can compare the two approaches in our guide to AI versus traditional computer-assisted coding. What follows is a vetted list of recommended tools, the technology behind them, evaluation criteria, and compliance considerations.

What AI coding tools actually deliver: accuracy, speed, and revenue

A 2025 systematic review of 35 LLM coding studies, registered with PROSPERO and reported per PRISMA guidelines, found a median micro-F1 of 0.79 for frequent ICD codes. Micro-F1 weights performance by code frequency, so common diagnoses dominate the score. The review screened 590 records. It concluded that LLM-based systems "reliably automate common ICD codes and frequently match or surpass professional coders." The same review found a median macro-F1 of just 0.42 for infrequent codes, which weighs all codes equally. That gap means real-world accuracy depends heavily on your case mix.

Vendor-reported outcomes add a practical dimension, but attributions matter. The 72% reduction in after-hours documentation belongs to Suki AI (AAFP Innovation Lab Phase 2 report), not DeepScribe. DeepScribe reports a reduction of up to 75%, measured at Covenant HealthCare (Aug-Nov 2023). Medmio CodeSight reports a 10.9-second median processing time per chart; MediCodio's published figure is approximately 1.5 minutes (90 seconds). The 15% increase in charge capture attributed to XpertDox comes from a client testimonial by Arihant Jain, President of Nao Medical, in a XpertDox press release dated Jan. 2, 2025, not a general product benchmark. Another XpertDox client, Urgent Care Pros, reports a 20% increase. These figures are self-published, so treat them as claims until you test them on your own charts.

• Accuracy: LLMs match or exceed professional coders on frequent codes (median micro-F1 0.79) but drop to 0.42 macro-F1 on rare codes, so benefits depend on case mix
• Speed: vendor-reported chart processing times range from 10.9 seconds (Medmio CodeSight) to same-shift turnaround, reducing backlog
• Revenue: NCCI edit checking before claim submission reduces denials; a XpertDox client reports 15% higher charge capture
• Cost: AI-assisted review lets experienced coders focus on complex charts while routine codes are pre-validated, reducing overtime and contract labor

How AI reads a chart: NLP, machine learning, and large language models

Three technologies power modern AI coding tools. Natural language processing (NLP) extracts conditions, laterality, episode of care, and complications from progress notes, discharge summaries, and problem lists. Machine learning (ML) classification models, trained on labeled coded charts, rank candidate codes by likelihood. Large language models (LLMs) read the full clinical note in context. They reason over the documentation and produce a code with a justification grounded in the clinical text.

A MIMIC-IV validation study analyzed 865,079 ICD-10-CM code assignments across 64,764 hospital admissions and found that a GPT-4o-based system achieved 93.6% accuracy, with 95.4% sensitivity and 85.2% specificity. Principal diagnoses showed a PPV of 93.9%, while secondary diagnoses dropped to 83.8%. The distinction between autonomous coding (AI assigns and posts codes above a confidence threshold) and AI-assisted coding (AI suggests, coder decides) matters for compliance. Most organizations run a hybrid model with human-in-the-loop controls.

Natural language processing extracts the clinical facts

NLP identifies the conditions, procedures, and modifiers documented in a clinical note. It parses free-text fields, problem lists, and discharge summaries. The goal is to find the facts a coder needs: the diagnosis, its laterality, whether it is acute or chronic, and whether it was present on admission.

Machine learning ranks candidate codes

ML models trained on labeled coded charts rank candidate codes by likelihood. Medmio's ICD-10 model, for example, was trained on 72,000 labels. But ML models trained on prior-year codes degrade when CMS releases annual updates unless the vendor retrains or embeds coding logic in a rules engine.

Large language models read the full note in context

LLMs go beyond keyword matching. They read the entire clinical note and reason over the documentation, producing a code with a justification grounded in the text. This handles ambiguity and context that keyword-based systems cannot.

Recommended AI tools for ICD-10-CM coding at a glance

The ten platforms below span four categories: autonomous coding engines, AI-assisted suggestion tools, searchable code references, and enterprise CAC-plus-CDI suites. Use the list for a quick answer and the table for side-by-side comparison. The subsections that follow cover the distinctions that matter for each tool.

• Corti: API-native coding engine returning ICD-10-CM, ICD-10-PCS, and CPT predictions with supporting evidence and audit trail
• MediCodio: AI coding assistant proposing ICD-10, CPT, and HCPCS codes for coder review
• Nym: autonomous coding for emergency department and radiology charts with full audit trail
• AAPC Codify: searchable code reference with crosswalks and NCCI edits for lookup and compliance checking
• 3M 360 Encompass: enterprise CAC platform combining computer-assisted coding with CDI workflows
• MAIA: reads clinical notes and returns ranked ICD-10, CPT, and HCPCS suggestions with confidence scores and CMS guideline validation
• RapidClaims: newer entrant focused on autonomous coding (limited public documentation at time of writing)
• Twofold: newer entrant focused on RCM automation (limited public documentation at time of writing)
• CombineHealth: newer entrant focused on denial prevention (limited public documentation at time of writing)
• An AI assistant built by coders for coders, interpreting clinical documentation and suggesting ICD-10-CM, CPT, and HCPCS codes with explanations of clinical reasoning

How the leading AI coding solutions compare

Vendor accuracy figures in this table are self-reported. Where public documentation was thin, we flag that rather than guess. For example, Medmio's 98.1% accuracy figure was measured on 945 billed-and-paid charts, but that benchmark is vendor-published, not independently audited.

Corti

Corti's API-native coding engine returns ICD-10-CM, ICD-10-PCS, and CPT predictions with supporting evidence and a complete audit trail. Corti reports 74% F1 accuracy on automated coding benchmarks versus 48% for raw OpenAI models. The coding logic is built into the engine to avoid the October retraining gap.

MediCodio

MediCodio proposes ICD-10, CPT, and HCPCS codes for coder review. It is a strong fit for teams wanting suggestion-and-review workflows rather than autonomous posting. MediCodio's published processing time is approximately 1.5 minutes per chart.

Nym

Nym handles autonomous coding for emergency department and radiology charts, assigning ICD-10-CM and CPT codes with full audit trails and no human intervention on high-confidence charts. Complex cases route to credentialed coders.

AAPC Codify

AAPC Codify is a searchable code reference with crosswalks and NCCI edits. It is best for coders who want a lookup tool with compliance checking rather than note-reading AI.

3M 360 Encompass

3M 360 Encompass (now Solventum) is an enterprise CAC platform combining computer-assisted coding with CDI workflows. It offers deep EHR integration for large health systems.

MAIA

MAIA reads clinical notes and returns ranked ICD-10, CPT, and HCPCS suggestions with confidence scores. A deterministic CMS guideline validator catches Excludes1 conflicts, missing companion codes, and non-billable codes before the coder sees the chart.

RapidClaims

RapidClaims is a newer entrant focused on autonomous coding. Public documentation was limited at time of writing, so verify current capabilities directly with the vendor.

Twofold

Twofold focuses on RCM automation. Public documentation was limited at time of writing, so contact the vendor directly for current feature lists.

CombineHealth

CombineHealth focuses on denial prevention. Public documentation was limited at time of writing, so contact the vendor directly for current capabilities.

Coders get instant, expert-level clarification that removes the delay of peer consultations and manual lookups. The integrated searchable reference covers ICD-10-CM, CPT, HCPCS, and ICD-10-PCS in one place, so coders verify codes without switching between tools. Every suggestion includes the clinical reasoning behind the coding pathway, so coders can confirm or override with confidence rather than guess.

What to look for when evaluating an AI coding tool

The academic evidence shows a large gap between frequent and rare code performance. A single headline accuracy number does not tell you how a tool will perform on your charts. Our 2026 buyer's guide for AI coding assistants covers these criteria in more detail. Here are the seven features to prioritize.

• NLP and LLM accuracy on your case mix: ask vendors for performance data broken down by specialty and code frequency, not a single headline number
• EHR integration: confirm pre-built connections to your system (Epic, Cerner, Athenahealth, eClinicalWorks) and support for HL7v2 and FHIR interfaces
• Audit trails and explainability: every suggested code should show the supporting documentation snippet and reasoning for compliance reviews
• Support for multiple code sets: ICD-10-CM alone is rarely enough; look for CPT, HCPCS, and ICD-10-PCS plus fiscal-year switching for retrospective audits
• Compliance checks: NCCI edits, MUE limits, Excludes1 and Excludes2 conventions, and HCC, CC/MCC, HAC, and POA indicators where risk adjustment applies
• Security posture: HIPAA compliance with a signed BAA, SOC 2 Type II or HITRUST certification, encryption in transit and at rest, and a commitment not to train models on your data
• Human-in-the-loop controls: configurable confidence thresholds and review queues so uncertain charts always route to a credentialed coder

A practical path from shortlist to implementation

1. Define scope: pick one service line or chart type (ED notes, radiology, primary care) for a pilot rather than a big-bang rollout
2. Run a parallel test: code the same charts manually and with the AI tool, then compare. Some vendors, for example Medonix, offer a 30-day parallel-run guarantee with targets written into the contract
3. Set confidence thresholds: many practices auto-post only above 95% confidence and route everything else to human review
4. Train the team: coders need to understand what the tool does and does not do; adoption fails when AI output arrives without context or explanation
5. Monitor and audit: track denial rates, coding productivity, and audit findings monthly; expect implementation timelines of roughly 2 to 6 weeks from contract to production for integration-heavy tools

Three challenges deserve attention during implementation. October code updates require vendor readiness; confirm the tool handles fiscal-year changes without retraining. Physician documentation quality affects AI accuracy, so pair the tool with CDI query workflows. Staff resistance fades when coders see the AI explaining its reasoning rather than issuing opaque suggestions.

Compliance, ethics, and the role of human oversight

Any AI tool touching protected health information must operate under a Business Associate Agreement (BAA). Verify SOC 2 Type II or HITRUST certification and data-encryption practices before sharing a single chart. You can read more about regulatory compliance for AI medical coding tools on our blog. The coder of record remains responsible for final code assignment, regardless of what the AI suggests.

AI suggestions must follow the official ICD-10-CM guidelines and NCCI edits published by CMS. The guidelines, approved by the four Cooperating Parties (AHA, AHIMA, CMS, and NCHS), describe how to apply the code set correctly. HIPAA requires covered entities to use standard code sets like ICD-10-CM for covered transactions, but the official guidelines themselves are interpretive rules, not statutory mandates. The practical action for buyers: confirm the tool follows current guideline updates, produces auditable code assignments, and keeps the coder of record in control of final decisions.

• HIPAA: require a signed BAA before any PHI enters the tool; verify encryption in transit and at rest
• CMS guidelines: confirm the tool follows current ICD-10-CM official guidelines and NCCI edits; the coder of record is responsible
• Algorithmic bias: the systematic review found accuracy declines for rare diagnoses and only 3 of 35 studies conducted external validation; monitor performance across patient populations
• Data privacy: confirm the vendor never uses your clinical data to train models without explicit agreement

Where AI medical coding goes next

The 2025 systematic review concluded that LLM-based systems "reliably automate common ICD codes and frequently match or surpass professional coders." It noted that "prospective, multicenter trials and transparent reporting of prompts and post-processing rules are required before clinical deployment." As models improve, expect the rare-code gap to narrow. The MIMIC-IV study found that PPV declines with diagnosis position: each step down produced a 0.12% decrease in PPV.

Deeper CDI integration is the next frontier. AI that generates compliant, clinically supported queries to physicians improves documentation at the source rather than fixing it downstream. The coder role is shifting from manual lookup to verification, auditing, and clinical reasoning. Coders who adopt AI-assisted workflows position themselves for the higher-value work that remains uniquely human.

FAQs about AI tools for ICD-10-CM coding

How accurate is AI for ICD-10-CM coding?

The evidence covered above applies here too. Frequent ICD codes see a median micro-F1 of 0.79, often matching or exceeding professional coders. Rare diagnoses drop to a median macro-F1 of 0.42. Vendor-reported figures run higher, but those are self-published. Ask vendors for performance data on your specialty and code mix.

Does AI replace human medical coders?

No. Every credible tool positions AI as an augmentation, not a replacement. Most systems use a human-in-the-loop model: the AI suggests codes with confidence scores, and anything below a set threshold routes to a credentialed coder. The coder of record remains responsible for final code assignment and compliance.

How is AI coding different from EHR suggestions and CAC tools?

Traditional CAC tools use a spectrum of NLP methods, from keyword matching to statistical models and symbolic rules, but they struggle with negation, context, and full-note comprehension. Modern AI tools read the full clinical note using NLP and large language models, apply payer and NCCI rules, and assign or suggest codes with a confidence score. The result is fewer unspecified codes and better specificity.

Do AI coding tools keep up with annual ICD-10-CM updates?

It varies by architecture. ICD-10-CM updates every October 1 on a fiscal-year cycle, and ML-based tools trained on prior-year codes can suffer an accuracy gap until they are retrained. Some vendors embed coding logic in a rules engine so updates apply without retraining. Ask each vendor how they handle the October update.

What should I look for when choosing an AI ICD-10-CM coding tool?

Prioritize six things. First, accuracy data specific to your case mix. Second, EHR integration via HL7 or FHIR. Third, audit trails that show the supporting documentation. Fourth, support for CPT and HCPCS alongside ICD-10-CM. Fifth, built-in compliance checks for NCCI edits and MUE limits. Sixth, HIPAA compliance with a signed BAA. Explainability matters most; a tool that shows its clinical reasoning builds coder trust and survives audits.

Choose a tool that shows its reasoning

Match the tool's approach to your team's workflow. If you want autonomous posting for high-volume specialties, look for audit trails and configurable confidence thresholds. If you want AI-assisted coding that keeps the coder in control, prioritize explainability. Verify accuracy claims with your own parallel test before trusting any headline number.

The right tool suggests ICD-10-CM, CPT, and HCPCS codes with explanations of the clinical reasoning behind each coding pathway, so coders make faster, more confident decisions without giving up judgment. See how it works and try it on your own charts.

• Book a demo to see the assistant read your clinical notes, suggest codes with full reasoning, and generate compliant CDI queries your physicians will actually respond to