Home/Technology/Coding AI Engine
92% auto-accept · sub-second per line · CPC coder-in-loop

Coding AI tuned to your specialty.

A deterministic rule engine across ICD-10, CPT, HCPCS, HCC V28, and modifiers. Two runs of the same chart produce the same codes. Every flagged line is reviewed by a CPC-credentialed coder by specialty, with a full audit trail per line. This is the engine behind our hospital, ABA, FQHC, and physician-group coding books.

Deterministic, reproducible coding HCC V28 crosswalk, refreshed quarterly ABA 97151 family tuned
The coverage

Five coding domains. One engine.

Most coding tools handle CPT and ICD-10 and stop. We carry full coverage across the five domains a complete claim line actually needs. HCC V28 capture with hierarchy and disease-interaction logic. HCPCS with NDC mapping for J-codes and infusions. Modifier validation including 25, 59, X-modifiers, and laterality. ICD-10 sequencing tied to CPT validation so the diagnosis supports the procedure.

The same engine powers our HCC dashboard and our chart-to-codes pipeline in RCM_DEV. The curated V28 crosswalk is maintained against CMS updates within 30 days of each release. Two runs of the same chart produce the same codes. That property is what our CPC auditors and the payer review teams care about.

Five domains · in-production coverage · curated rule packs
ICD-10-CM
Diagnosis
2024 set · sequencing · CPT validation
CPT
Procedure
E/M, surgical, OB, ABA 97x series
HCPCS Lvl II
Supply · Drug
J-codes · NDC map · DME
HCC V28
Risk Adjustment
Hierarchy · trumping · interaction
Modifiers
25, 59, X-modifiers, laterality, anatomic, professional/technical
Payer-specific rules · audit-trail per line
Curated rule library · refreshed quarterly · CMS V28 within 30 days of release
What the coder sees

One screen. Every line traceable.

The coder-in-loop view. Every proposed code shows the rule that fired, the supporting documentation snippet, the confidence score, and the modifier sequence. Below-threshold lines surface for review. Above-threshold lines post to the claim without intervention. Same view, same audit trail, every chart.

CodingCase · Encounter 2026-06-19 · Patient ID redacted · Specialty: Behavioral Health
97153
F84.0
Adaptive behavior treatment by protocol · 15 min units
Rule BH-97153-AUTH-V3 · units within active auth · rendered by RBT under BCBA supervision · documentation snippet: "session 9:00-10:30, behavior plan goals 1, 3, 4 worked"
0.97 auto-accept
97155
F84.0
Adaptive behavior treatment protocol modification
Rule BH-97155-BCBA-V2 · BCBA credential validated · supervision concurrent · 30 min documented
0.94 auto-accept
97156
F84.0
Family adaptive behavior treatment guidance · with caregiver
Rule BH-97156-FAM-V1 · caregiver present documented · documentation gap on goal-linkage detail
0.71 review
99213
F84.0, F90.2
Office visit · E/M established · low complexity
Rule EM-99213-MDM-V4 · MDM components present · documentation supports level · modifier 25 not required
0.96 auto-accept
The capabilities

Six capabilities. Every line auditable.

Each capability is real production code. The same engine processes ABA, FQHC, physician-group, and hospital E/M charts. The rule library carries the specialty differences. The engine applies them deterministically and posts the full audit trail per line.

01 · Deterministic rule engine

Two runs, same codes.

Rule-based code assignment with a curated rule library per specialty per payer. LLMs assist with documentation parsing; the assignment itself is deterministic. Output is reproducible, auditable, and traceable to the rule that fired. That is the property our CPC auditors and payer review teams require.

02 · HCC V28 current

V28 crosswalk, 30-day refresh.

The V28 model with hierarchy, trumping rules, and disease-interaction logic encoded. V24 retained for retrospective sweeps. Crosswalk refreshed against CMS updates within 30 days of release. Output reflects the same logic CMS uses for risk-adjustment payment, not a generic approximation.

03 · ABA 97151 family tuned

97151, 97153, 97155, 97156, 97158.

The ABA family is one of our heaviest specialty packs. Authorization unit tracking, supervision ratios, three-way match between assessment, treatment plan, and rendered session, and the BCBA-versus-RBT credentialing requirement per code per state. The only RCM vendor with a BHCOE channel partnership.

04 · Three-way match

Auth · credential · documentation.

Every line validates against three things in parallel: the rendered service matches the active authorization, the rendering provider has the correct credential for the code, and the documentation supports the level billed. A miss on any of the three flags the line before submission. Biggest source of recovered revenue on ABA and behavioral-health books.

05 · Audit trail per line

Rule, snippet, actor, timestamp.

Every coded line carries a structured trace: the rule ID that fired, the documentation snippet it matched, the confidence score, the coder who reviewed (if applicable), timestamp, actor, and IP. Exportable for payer audit, CPC quarterly review, or internal QA. Same audit-trail design as Credential OS.

06 · Coder-in-loop validation

Judgment cycles, not data entry.

Below-threshold lines route to a specialty-keyed queue. CPC-credentialed coders review, accept or modify, and the modification feeds back into the rule library as precedent. Continuous loop, not batch. Independent quarterly CPC audit on a meaningful sample confirms accuracy holds. Coders spend judgment cycles, not data-entry cycles.

How a chart runs

Four steps. Sub-second per line.

From EHR push to finalized claim. The engine handles the volume; the CPC coders handle judgment. Below is the workflow per chart, with the latency budget on each step.

Step 01

Ingest chart.

EHR pushes encounter payload (clinical notes, problem list, orders, metadata). Documentation parser extracts concepts. Specialty pack and payer rule library load for the encounter context. Latency budget: under 200ms.

Step 02

Code the line.

Rule engine assigns ICD-10, CPT, HCPCS, HCC mappings, and modifiers per line. Confidence score per line. Audit trace written. Latency budget: sub-second per line at P95 on production volume.

Step 03

Validate three ways.

Three-way match against active authorization, provider credential, and supporting documentation. Lines that clear all three at high confidence auto-accept. Lines that fail any check route to coder review with the specific failure surfaced.

Step 04

Review exceptions.

CPC coder by specialty reviews flagged lines. Accept, modify, or escalate. Modifications feed back into the rule library as precedent for similar charts. Final claim posts with the full audit trail per line attached. Quarterly external CPC audit on a sample.

What clients see

Measured outcomes from the coding engine.

Across our active book. Anonymized; individual results depend on specialty mix, payer mix, and pre-engagement coding baseline. The numbers below are honest, not headline.

92%
Per-line auto-accept
Lines clearing the confidence threshold and posting without coder review. The other 8 percent route to coder-in-loop. Roughly half of those are documentation gaps; half are complex multi-line surgical, OB, or HCC capture cases needing rule precedence judgment.
3x
Coder throughput vs manual
Coders working the coder-in-loop queue process roughly three times the chart volume of pure-manual coding because they see only the lines needing judgment, not the data entry. Same accuracy on independent quarterly CPC audit. Reduces the coder-shortage pressure most groups feel.
V28
HCC capture, current model
V28 is the active CMS model for payment year 2025. The crosswalk we maintain refreshes within 30 days of CMS updates. V24 retained for retrospective sweeps and audit defense. Hierarchy, trumping, and disease-interaction logic encoded against the same rules CMS applies.
Common questions

Frequently asked: the coding engine.

What does the coding engine actually do per chart?
The engine accepts a chart payload (clinical notes, problem list, orders, encounter metadata) and returns a structured coding case: ranked ICD-10 candidates, CPT and HCPCS line items, modifiers, sequencing, HCC mappings under V28, and a confidence score per line. Each line carries the supporting documentation snippet that justified it. A CPC-credentialed coder reviews any line under the configured confidence threshold before the claim is finalized.
Is this an LLM doing the coding?
No. The core is a deterministic rule engine with a curated V28 HCC crosswalk, HCC hierarchies, and ICD-10 to CPT validation logic baked in. LLMs assist with documentation parsing and concept extraction, but the code assignment itself is rule-based so the output is auditable, reproducible, and traceable to the rule that fired. Two runs of the same chart produce the same codes. That is the property our CPC auditors and payer review teams care about.
What is the 92 percent auto-accept rate?
Across our active book, 92 percent of coded lines clear the confidence threshold and post without coder review. The other 8 percent route to a coder-in-loop queue. Of that 8 percent, roughly half are flagged for documentation gap (the chart does not support the code level), and half are complex multi-line surgical, OB, or HCC capture cases where rule precedence needs human judgment. The 92 percent is honest measurement, not a marketing number.
How current is the HCC V28 crosswalk?
The V28 crosswalk we maintain is refreshed against CMS updates within 30 days of each release. V28 is the active model for payment year 2025, and the engine carries both V24 and V28 to support retrospective sweeps. HCC hierarchies, trumping rules, and disease interaction logic are encoded in the rule library so the model output reflects the same logic CMS uses for risk adjustment payment.
Is ABA coding actually supported?
Yes. The ABA family (97151 assessment, 97153 protocol implementation, 97155 protocol modification, 97156 family guidance, 97158 group) is one of our most heavily tuned specialty packs. Authorization unit tracking, supervision ratios, three-way match between assessment, treatment plan, and rendered session, and the BCBA-versus-RBT credentialing requirement per code are all enforced at the coding gate. ASP-RCM is the only RCM vendor with a BHCOE channel partnership and our team co-authored Essential First Step for ABA billing.
What is the three-way match capability?
For every claim line, the engine validates three things in parallel: the rendered service matches the active authorization (CPT, units, date range), the rendering provider has the correct credential for the code (BCBA, RBT, BHCOM, LCSW, MD, NP per state and payer), and the documentation supports the code level billed. A miss on any of the three flags the line before submission. This is the single biggest source of recovered revenue on ABA and behavioral-health books.
What is the audit trail per line?
Every coded line carries a structured trace: the rule ID that fired, the documentation snippet it matched against, the confidence score, the coder who reviewed it (if applicable), the timestamp, the actor, and the IP. The trace persists for the duration of the engagement and is exportable for payer audit, CPC quarterly review, or internal QA. The same audit-trail design powers Credential OS, our credentialing platform.
How does the coder-in-loop validation work?
Charts below the per-line confidence threshold land in a queue keyed by specialty. CPC-credentialed coders by specialty review flagged lines, accept or modify, and the modification feeds back into the rule library as a precedent for similar charts. The loop is continuous, not batch. Coders see the AI proposal with the supporting documentation and the rule that produced it, so they spend judgment cycles, not data-entry cycles. Independent quarterly CPC audit on a meaningful sample confirms the engine and coder accuracy hold.

Send 30 days of coded charts. We send back the audit.

A free 30-day coding audit. Drop your last 30 days of finalized claims and supporting documentation. We return a four-page audit covering per-specialty per-payer accuracy, undercoding and overcoding flags by dollar value, HCC capture gaps under V28, and a 90-day fix plan. An ASP-RCM senior partner on the call.