What problem does it solve?
Every patient encounter has to be translated into standard codes before a provider can bill for it. Certified coders read the notes and choose among tens of thousands of diagnosis and procedure codes, modifiers and add ons, following payer rules that change every year. Coders are scarce, backlogs delay cash, and inconsistent coding causes claim denials, rework and lost revenue.
Coding also carries compliance risk in both directions. Undercoding loses revenue that the documentation supports; overcoding, especially of diagnoses that raise risk adjustment payments, leads to audits, repayments and fraud allegations. Computer assisted coding has suggested codes for years; newer systems code the simpler encounters on their own, which raises the question of who checks their work.
How does it work?
- Receive the signed documentation. When an encounter is closed, the notes, orders and results are sent to the coding engine.
- Assign codes. The model proposes diagnosis and procedure codes, modifiers and sequencing, with the text that supports each code and a confidence level.
- Apply rules. Payer edits, coding guidelines and the organization's own policies are checked, and documentation gaps are flagged for the clinician.
- Route by confidence. High confidence encounters of approved types are released to billing automatically; the rest go to a coder's worklist with the suggested codes.
- Audit and learn. Coders and auditors sample automated encounters, and denials and audit findings feed back into rules and models.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | 98.3% | 1 | 1 organization |
| Automation rate | Too few to pool | 95.5% | 1 | 1 organization |
| Error reduction | Too few to pool | 59% | 1 | 1 organization |
| Productivity gain | Too few to pool | 70% | 1 | 1 organization |
Value drivers: Lower cost to serve, Speed and cycle time, Compliance quality, Revenue growth, Employee productivity.
Indicative value
A physician group with 1 million coded encounters a year
USD 583,333 to USD 4.3 million
Coder effort released per year
How this is calculated
Formula: encounters * automationShare * minutesPerEncounter / 60 * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Encounters coded per year encounters, encounters per year | 1,000,000 | 1,000,000 | The reference physician group. |
| Share of encounters coded without a coder automationShare, fraction of encounters | 0.5 | 0.85 | Conservative against the 95.5% encounter level automation that Your Health reports on this page. The lower range is an editorial caution, not a sourced figure: one self reported result is thin evidence, and it does not say how "automated" is defined or measured. |
| Coder minutes per encounter today minutesPerEncounter, minutes per encounter | 2 | 5 | Editorial assumption for professional fee coding. Replace with your own productivity data. |
| Fully loaded cost per coder hour costPerHour, USD per hour | 35 | 60 | Editorial assumption covering internal and outsourced coders. Replace with your own. |
What it leaves out: Coder effort only. It leaves out the effect on denials, days to bill and cash, the revenue effect of more complete coding (which must be supported by documentation), audit and compliance costs, and the licence cost, often charged per encounter.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
US Department of Veterans Affairs, Veterans Health Administration
United States · Healthcare · 2025
The Veterans Health Administration reports in the 2025 federal AI use case inventory that it uses the Solventum (formerly 3M) 360 Encompass computer assisted coding system to suggest ICD-10-CM, CPT and HCPCS codes from the clinical documentation of each encounter. Medical coders review, validate and select the codes, so the AI speeds up the coder rather than coding on its own. The inventory lists the use as deployed, classifies it as high impact and publishes no outcome figures.
No outcome disclosed.
Your Health
United States · Healthcare · 2025
Your Health, a senior focused primary and specialty care group in South Carolina and Georgia with about one million patient visits a year, went live with Fathom's autonomous coding in October 2025 across every service line and place of service, integrated with its athenahealth record system. In its own newsroom Your Health reports a 95.5% encounter level automation rate and coding accuracy up from 96.3% to 98.3% since implementation; Fathom's release adds more complete diagnosis capture that raised average risk adjustment scores. Neither source says how accuracy was measured or how the remaining encounters are handled.
- Automation rate: 95.5%, encounter level, since implementation in October 2025
"Since implementation in October 2025, we have achieved a 95.5% encounter-level automation rate and increased coding accuracy from 96.3% to 98.3%, reflecting meaningful gains in efficiency, precision, and scalability across the organization."
Claimed by: organization - Accuracy: 98.3%, since implementation in October 2025
"Since implementation in October 2025, we have achieved a 95.5% encounter-level automation rate and increased coding accuracy from 96.3% to 98.3%, reflecting meaningful gains in efficiency, precision, and scalability across the organization."
Claimed by: organization
Mass General Brigham
United States · Healthcare · 2023
CodaMetrix, a company spun out of Mass General Brigham, uses machine learning and natural language processing on the clinical record to translate clinical notes into procedure and diagnosis codes automatically and reduce the workload of human coders. In CodaMetrix's February 2023 funding release, Mass General Brigham's vice president of physician revenue cycle services cited "a 70% reduction in manual labor" and a 59% reduction in denials due to coding as "our outcomes", without saying which work, period or sites the figures cover. Mass General Brigham physician organizations also invested in the company.
- Productivity gain: 70%, manual labor, scope and period not stated
"Our outcomes — a 70% reduction in manual labor — 59% reduction in denials due to coding, and a significant increase in cost savings — is the proof." said Michael Mercurio, Vice President of Physician Revenue Cycle Services at Mass General Brigham."
Claimed by: organization - Error reduction: 59%, claim denials due to coding
"Our outcomes — a 70% reduction in manual labor — 59% reduction in denials due to coding, and a significant increase in cost savings — is the proof." said Michael Mercurio, Vice President of Physician Revenue Cycle Services at Mass General Brigham."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Historical encounters with documentation and final codes for calibration and testing
- Current code sets, payer rules and the organization's coding policies
- Denial and audit history by code and specialty
Systems to integrate
- Electronic health record for signed documentation
- Practice management or billing system for charges and claims
- Coder worklist and query tools for clinician documentation questions
- Denial management and audit systems
Complexity: Medium
Mature vendors exist and integrate with common record systems. The work is in choosing which encounter types may be released without a coder, calibrating to payer mix and local guidelines, and building the audit loop that compliance teams need.
- 1
Start with suggestions, then release by encounter type
Run the engine as a suggestion tool first, measure agreement with coders per encounter type, and allow automatic release only where agreement stays high.
- 2
Set confidence thresholds with compliance
Agree with compliance which encounter types, specialties and codes may be released automatically and which, such as risk adjustment diagnoses, always need a coder.
- 3
Close the documentation loop
Route documentation gaps back to clinicians as queries instead of coding around them, so codes stay supported by the record.
- 4
Audit automated encounters continuously
Sample automated encounters every week against a coder's review, track accuracy by code family and feed errors back.
- 5
Watch denials and payer feedback
Compare denial rates and reasons before and after, per payer, and investigate any rise in high value codes.
Guardrails
- Only encounter types with proven accuracy are released without a coder
- Every automated code is supported by text in the signed documentation, stored with the claim
- Risk adjustment diagnoses and high value procedures reviewed by a certified coder
- Documentation gaps sent to the clinician as a query, never filled in by the AI
- Weekly audit sample of automated encounters with results reported to compliance
KPIs to instrument
- Share of encounters released without a coder, by specialty
- Coding accuracy on a weekly audit sample
- Denials due to coding, per payer
- Days from encounter to claim
- Shift in the distribution of evaluation and management levels and risk adjustment scores
Human in the loop
Certified coders handle every encounter below the confidence threshold and all excluded code families, and they audit a sample of automated encounters. Compliance owns the release rules, and clinicians answer documentation queries.
Common failure modes
- Upcoding at scale
- The model systematically selects higher levels or adds unsupported diagnoses, which creates overpayment and fraud exposure. Monitor code distributions and audit high value codes.
- Accuracy measured on the wrong sample
- Vendor and customer accuracy figures may cover only the encounters that were automated, or a sample chosen by the vendor; the Your Health figures on this page do not say how accuracy was measured. Audit a random sample of automated encounters yourself.
- Stale rules
- Annual code set and payer rule changes are not reflected in time. Assign owners for updates and test before each effective date.
- Coders lose the skill to audit
- If coders only handle exceptions, the organization may lose the expertise to check the machine. Keep audit and training time in coder roles.
What are the risks and rules?
EU AI Act
Minimal risk
Assigning billing and statistical codes from clinical documentation is not listed in Annex III and does not decide on a person's access to care, so no specific AI Act obligations apply beyond AI literacy. Health data processing falls under GDPR Article 9, and in the United States under HIPAA and the payment integrity rules of public payers. Minimal under the AI Act does not mean low stakes: the Veterans Health Administration classifies its computer assisted coding deployment on this page as high impact in the 2025 US federal AI use case inventory, even though coders select every code.
Rules that apply
Guidance
- General Compliance Program Guidance (Office of Inspector General, US Department of Health and Human Services, North America). Voluntary OIG guidance on the federal fraud and abuse laws and the seven elements of a compliance program, the frame against which US providers audit and monitor billing and coding, whether done by people or software.
- Medicare Advantage Risk Adjustment Data Validation Program (Centers for Medicare & Medicaid Services, North America). Explains how CMS audits whether diagnoses that Medicare Advantage organizations submit for risk adjustment are supported by medical records. Providers are not the audited party, but the plans they code for are, so unsupported diagnoses from automated coding flow into this exposure.
- ICD-10 (Centers for Medicare & Medicaid Services, North America). CMS page for the ICD-10 code sets. It publishes the annual ICD-10-PCS procedure code files and relays the ICD-10-CM diagnosis code updates that CDC develops and announces; an automated coding engine must track both.
Controls to put in place
- Written release rules per encounter type and code family, approved by compliance
- Evidence link from every code to the supporting documentation, kept with the claim
- Weekly audit of automated encounters and trend monitoring of code distributions
- Change control for model, rule and code set updates
- Access controls and logging for clinical documentation used by the engine
Frequently asked questions
- What share of encounters can AI code without a coder?
- Published figures are few and define automation differently. Your Health reports that it codes 95.5% of encounters automatically with Fathom across all service lines, with accuracy rising from 96.3% to 98.3%. A Mass General Brigham executive cited "a 70% reduction in manual labor" with CodaMetrix, without stating scope or period. Our editorial advice is to treat these self reported figures as upper bounds and measure on your own audit sample.
- Does automated coding raise compliance risk?
- It can, because errors repeat at scale. The main exposure is unsupported diagnoses or higher service levels that increase payments, which public payers audit. Keep evidence for every code, audit automated encounters and route risk adjustment diagnoses to coders.
- Is computer assisted coding the same as autonomous coding?
- No. Computer assisted coding suggests codes that a coder confirms, as in the Veterans Health Administration's deployment; autonomous coding releases confident encounters to billing without a coder and sends the rest to a worklist.
How to cite this page
Blits.ai AI Use Case Library, "AI medical coding for clinical encounters", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/medical-coding-automation. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published