Medical coding automation can safely code a narrow band of work: routine, well-documented encounters with few codes and no modifiers, such as a standard office visit or a single lab panel. Complex visits, surgery, modifiers and uncertain diagnoses need a certified coder. Whatever software codes, a regular audit against the clinical note should check it.
Encounters medical coding automation can handle
Medical coding automation is software that assigns procedure codes (CPT or HCPCS) and diagnosis codes (ICD-10-CM) to an encounter from its documentation. It handles encounters well when the note is structured, the services are few and repeated, and the rules that choose the code are explicit. The simpler and more uniform the visit, the safer the automation.
Code sets set the stage. The AMA overview of the CPT code set describes CPT as "a listing of terms and five-digit codes that primarily describe medical services and procedures," maintained by the CPT Editorial Panel. Diagnosis codes follow the ICD-10-CM Official Guidelines, and the FY 2026 ICD-10-CM Official Guidelines for Coding and Reporting say adherence to them is required under HIPAA.
Encounter types that tend to suit automation share a profile:
- Routine office visits with a clear, documented basis for the level. The AMA page on CPT code 99213 describes that established-patient visit as requiring a low level of medical decision making, or 20 to 29 minutes of total time on the date of the encounter when time is used to select the code.
- Single-test encounters, such as one lab panel or one immunization, where the order names the service directly.
- Recurring therapy visits with the same documented services each time.
- Preventive visits that follow a standard template with no added problems.
Even inside these groups, automation works best as a first draft. The software proposes codes, and a coder either accepts them or sends the chart for full review. The share of charts that pass untouched is a number each practice has to measure for itself, by encounter type, rather than assume.
Which coding cases need a certified coder?
Coding cases need a certified coder when the documentation calls for judgment: surgery and procedures with bundling rules, visits with more than one problem, modifiers, inpatient stays, and any note with uncertain or conflicting diagnoses. A certified coder reads the whole record, applies the guidelines and can query the provider when the note is unclear.
The credential matters because the work is classification under rules. AHIMA's Certified Coding Specialist (CCS) credential page describes the credential as designed for people "skilled in classifying medical data from medical records," and says coding specialists create the coded data that providers use to obtain reimbursement.
Cases that need a person usually involve at least one of these:
- Modifiers. A modifier changes how a payer treats a code. Modifier 25, for example, claims a separate evaluation and management service on the same day as a procedure. The HHS OIG guide to avoiding Medicare and Medicaid fraud and abuse warns that upcoding occurs when Modifier 25 is used for an E/M service that was not significant, not separately identifiable and not above and beyond the care usually associated with the procedure.
- Code combinations. The CMS NCCI edits page says procedure-to-procedure edits exist "to prevent improper payment when incorrect code combinations are reported," and Medically Unlikely Edits exist to prevent improper payment when services are reported with incorrect units of service. Knowing when a combination is allowed takes a coder.
- Uncertain diagnoses. The ICD-10-CM guidelines say that for outpatient services, diagnoses documented as "probable," "suspected," "questionable" or "rule out" should not be coded; the coder codes to the highest degree of certainty for that visit, such as symptoms or signs. For inpatient discharges, the same guidelines say to code the uncertain condition as if it existed. Software that applies one rule in both settings will be wrong in one of them.
- Conflicting documentation. When the problem list says one thing and the assessment says another, a coder queries the provider instead of picking one.
Rules-based versus AI-assisted medical coding automation
Rules-based medical coding automation maps documented facts to codes through explicit logic: if the order says this test, assign this code. AI-assisted automation reads free-text notes with a trained model and suggests codes with a confidence level. Rules are predictable and narrow; models are flexible and harder to explain.
Rules-based coding works where the input is already structured. A lab order with a test code, a charge sheet with checked boxes or an EHR template with picked values maps cleanly to a code. The logic can be read, tested and versioned. Its weakness is coverage: every new service or payer rule means new logic, and anything outside the rules falls through.
AI-assisted coding reads the narrative. It can find a procedure described in prose and suggest a code for it. Its weaknesses are the ones listed above: it may miss a modifier, code an uncertain diagnosis as confirmed, or propose a code pair that an edit forbids. Because a model's reasons are not always visible, every suggestion it makes should show the sentence in the note it relied on.
Code-set updates hit the two approaches differently. When CPT or ICD-10-CM codes are added, revised or deleted, rules-based logic needs someone to edit the mappings, and the change can be tested line by line. A model trained before the update may keep suggesting a deleted code or never suggest a new one until it is retrained or constrained by a current code list. Either way, the update date belongs on the audit calendar.
Most working setups combine them. Rules handle the structured encounters and enforce hard limits, such as known edits and required fields. A model drafts codes for narrative notes. A coder reviews anything the model was unsure about and anything the rules flagged. Reading diagnoses and procedures out of discharge summaries, notes and other clinical documents is its own subject, covered in the clinical records category.
How do coding audits catch automated coding mistakes?
Coding audits catch automated mistakes by pulling a sample of coded encounters, having a certified coder code each one independently from the clinical note, and comparing the two sets of codes. Disagreements are scored, sorted by type and traced back to the rule or model behavior that caused them.
Run the audit on a fixed schedule, such as monthly for a new setup, rather than waiting for a payer to complain. A schedule catches drift after template changes, software updates and new code-set releases, while a complaint-driven audit starts after the damage is done.
An example coding audit sample plan
This plan is a fictional example for a practice that has just turned on automated coding for office visits. The numbers are choices to adapt, not a standard.
| Step | Example choice | Why |
|---|---|---|
| Population | Last month's automated office visits | Audit what the software actually coded |
| Sample size | 30 charts per encounter type, chosen at random | Small enough for one coder, large enough to show patterns |
| Extra pull | 10 charts with the highest-level visit code | High levels carry the most payment risk |
| Reference | The signed clinical note for the date of service | Codes must be supported by documentation |
| Blinding | Auditor codes before seeing the software's codes | Prevents anchoring on the software's answer |
| Comparison | Each CPT code, each modifier, each ICD-10-CM code and its order | Errors hide in modifiers and sequencing |
| Scoring | Match, minor difference, or payment-changing difference | Separates harmless from costly errors |
| Follow-up | Trace each payment-changing difference to its cause | Fixing the rule or threshold prevents the next error, where fixing one chart does not |
Score agreement by line, not by chart. In a fictional audit of 30 charts, suppose the software assigned 84 codes and the auditor agreed with 78 of them. The agreement rate is 78 divided by 84, about 93%. Then look at the six disagreements one by one. If four were modifier errors and two were diagnosis sequencing, the modifier logic is the first thing to fix.
Track the results over time. A rising rate of payment-changing differences after a software update or a template change points straight at the change. A coding audit that finds a pattern also feeds denial work, because many denials trace back to a coding step. The guide to denial root cause analysis shows how to trace a denial to the step that caused it.
Schema Care does not assign codes. It reads the procedure and diagnosis codes already printed on CMS-1500 and UB-04 claims into fields, each with a confidence score and a link to its place on the page, which an auditor can use to compare a paper claim with the note. The review workflow on the homepage shows how a reviewer checks those fields.
Medical coding automation errors that trigger payer audits
Medical coding automation errors that trigger payer audits are the ones that repeat. A person who miscodes one visit makes one error. Software with a bad default makes the same error on every matching chart, and the pattern, such as a jump in high-level visit codes, is exactly what payer and government reviewers look for.
Upcoding by default settings is the most serious. The OIG guide describes upcoding as "using billing codes that reflect a more severe illness than actually existed or a more expensive treatment than was provided." It gives the example of billing a follow-up visit at a higher-level code as if a comprehensive new patient visit had been provided. Software that starts every visit at a high level, or that counts template text as documented work, can produce that pattern without anyone choosing it.
Payers and the government do review these patterns. An HHS OIG report, Improper Payments for Evaluation and Management Services Cost Medicare Billions in 2010, found that 42 percent of E/M claims in 2010 were incorrectly coded and 19 percent were lacking documentation, and that the incorrect coding included both upcoding and downcoding.
Other repeating errors to watch:
- Modifier appended by rule to every procedure-plus-visit day, whether or not the visit was separately identifiable.
- Units copied from a template instead of the note, which collides with Medically Unlikely Edits.
- Uncertain diagnoses coded as confirmed in outpatient visits.
- Downcoding by caution, where software or staff always choose a lower level to stay safe; it underreports care and is also an error.
Repeated coding errors often show up first as denials with matching reason codes, such as code 4, "The procedure code is inconsistent with the modifier used," or code 11, "The diagnosis is inconsistent with the procedure," on the X12 Claim Adjustment Reason Codes list. Routing those denials to the coding team rather than back to resubmission is covered in the guide to denial management automation.