Home/AI Suite/Evaluating Medical Coding AI Vendors
Buyer guide · 10 questions · 3 layers of evidence

How to evaluate medical coding AI vendors. The honest checklist.

Most vendor demos show you a beautifully tuned specialty on curated data. That is not what production looks like. This is the buyer's guide nobody else writes, because most people writing it are also selling. We are too. So we wrote down exactly what we would ask if we were on your side of the table, and the questions every honest vendor should be willing to answer in writing.

Code-level accuracy testing Specialty-by-specialty breakouts BAA, audit trail, exit terms
Section 01 · The accuracy trap

The 92 percent number is not the number.

Almost every medical coding AI vendor leads with a single accuracy figure. 92 percent. 94 percent. 96 percent. The figure is real in the narrow sense that the vendor measured something and got that result. The problem is what they measured.

Vendor accuracy claims are almost always pulled from a single tuned specialty, usually outpatient surgery, ED facility, or radiology professional, where coding patterns are highly templated and the model has been trained on tens of thousands of similar charts. Move the same model to ABA, behavioral health intake, FQHC sliding-scale primary care, hospital inpatient with comorbidities, or specialty drug J-codes, and the number falls. We have audited vendors who quote 94 percent in marketing and produce 68 percent code-level agreement on a buyer's own generalist data.

The defense is mechanical. Demand the test set composition: how many charts, what specialty mix, what payer mix, what date range, who coded the gold standard. Demand specialty-by-specialty breakouts on your own production data, not theirs. Demand code-level agreement (ICD-10-CM, CPT, HCPCS, modifiers, sequence) rather than encounter-level agreement, which can hide a missing modifier that costs you 18 percent of the reimbursement. AHIMA and AAPC both publish coder accuracy benchmarks (95 to 97 percent for credentialed coders on production work), and any AI claim above that floor on generalist data should trigger a hard look at the methodology.

Vendor demo claim vs real generalist accuracy
Vendor demo (single tuned specialty)92%
Radiology professional (templated)88%
Outpatient surgery84%
Primary care (FQHC mix)74%
Behavioral health intake71%
ABA session notes66%
Hospital inpatient w/ comorbidities63%
Code-level agreement, pre-coder review · industry observed range
Section 02 · The checklist

Ten questions every vendor must answer in writing.

Print this. Send it before the demo. The vendors who respond clearly and in writing are the ones worth the second meeting. The vendors who deflect to a slide deck are the ones who will deflect again at month six when accuracy slips and you need answers.

01

What is your accuracy on my data, not your demo data?

The single most important question. Demand a POC on 200 to 500 of your own de-identified charts, stratified by your specialty and payer mix. Demand code-level agreement, not encounter-level. Demand specialty-by-specialty breakouts.

→ Red flag: vendor refuses or wants to charge for the POC
02

Where is the audit trail for every code decision?

For every code the AI assigns, you need the source text it cited, the rule or guideline it applied, the confidence score, and the time stamp. Without an audit trail you cannot defend a payer audit or a RAC review. Compliance is yours, not the vendor's.

→ Red flag: explanations are post-hoc or generic
03

Is this a coder-in-loop workflow or a black box?

The best systems route low-confidence codes to a human coder with the AI's reasoning attached, and learn from every override. The worst systems output a final claim and expect you to trust it. Ask to see the coder UI in production, not a demo.

→ Red flag: no override mechanic or override is invisible to the model
04

How does it integrate with my EHR?

HL7 v2, FHIR R4, direct API, SFTP file drop, or a screen-scrape browser extension. Each has cost, latency, and reliability tradeoffs. Get the integration approach in writing, with the named EHRs they have shipped to production, the engineering hours estimated for your stack, and who owns the connector long term.

→ Red flag: "we will figure it out during onboarding"
05

What specialties are tuned vs untuned?

Every vendor has a tuned set and an untuned set. The tuned set is where their training data is dense; the untuned set is where you should expect 15 to 25 points lower accuracy. Demand the list. If your specialty mix overlaps the untuned set by more than 30 percent, that is the wrong vendor for you.

→ Red flag: "we work across all specialties equally"
06

What is the support model when a coder disagrees?

Coder pushback is the loudest signal you will get in the first 90 days. If a coder marks 30 charts as miscoded, who reviews the dispute, on what cadence, and does the model retrain on the override? Vendors who have no answer here will lose coder adoption and quietly disappear from the workflow.

→ Red flag: support is a ticketing queue with no SLA
07

What is the pricing structure?

Per-chart, per-coder, per-encounter, per-month flat, or per-percentage-of-collections. Each has incentive distortions. Per-chart pricing punishes complex cases. Per-percentage aligns vendor incentives but caps your margin. Demand a written quote that survives volume growth and specialty expansion without renegotiation.

→ Red flag: pricing only "on request"
08

How fast is the feedback loop on payer policy changes?

CMS publishes thousands of LCD and NCD updates per year. Commercial payers publish hundreds more. When a policy changes, how long until the model reflects it, and who is responsible for catching the change? AHIMA and HFMA both track this as a top compliance risk for autonomous coding.

→ Red flag: no defined retraining cadence or policy monitoring team
09

What is your hallucination rate and how do you measure it?

LLM-based coding tools can fabricate codes that look plausible but are not in the documentation. The measure is rate of codes assigned where no source text justifies them. Less than 0.5 percent is acceptable for a coder-assist tool; less than 0.1 percent for an autonomous one. If the vendor does not measure it, assume it is high.

→ Red flag: "we do not use LLMs" with no architectural detail
10

Who owns the data and what happens at contract end?

Three sub-questions. Does the vendor train on your data and benefit other customers from your patterns? On termination, do you get a complete export of every coded chart in a usable format? Is the model itself escrowed so a vendor failure does not strand your operation? These are the terms you negotiate before signing, not after.

→ Red flag: data ownership ambiguous or training rights asymmetric
Section 03 · The comparison framework

How to score vendors side by side.

The matrix below is the scorecard structure we recommend. Vendor A, B, and C represent typical archetypes we have audited on behalf of buyers. We list ourselves in the last column because we are also a vendor and the buyer's framing is to evaluate everyone, including us, against the same bar.

Criterion
Vendor A · slide-deck
Vendor B · enterprise
Vendor C · startup
ASP-RCM
Accuracy disclosure
OpaqueSingle number, no test set composition
PartialAggregate accuracy by service line, no payer breakout
PartialWill run a POC but on their curated set
OpenPOC on your data, code-level agreement, specialty breakout shared in writing
Audit trail depth
ThinConfidence score only, no source citation
DeepSource citation, rule cited, time stamp, override history
MidSource citation, no override history
DeepFull chart-to-codes lineage, every code tied to source text and rule, exportable
Coder workflow
Black boxAuto-submit only, no review queue
Review-firstConfidence-thresholded queue, batch override
MidReview queue but override is local, model does not learn
Coder-in-loopConfidence-thresholded review, override telemetry, monthly retraining cycle
EHR integration
LimitedScreen scrape or file drop, 2 to 4 named EHRs
BroadHL7 v2 + FHIR R4 + direct API across 15+ EHRs
LimitedSingle API, "EHR-agnostic via abstraction layer" claims
BroadHL7 v2, FHIR R4, direct API, billable connector engineering hours quoted upfront
Specialty coverage
NarrowOne tuned specialty marketed as universal
Broad-shallow15+ specialties but quality varies; refuses to disclose tuned vs untuned
Niche-deep2 to 3 specialties tuned well, honest about the rest
HonestTuned list published; ABA, HCC, BH, FQHC primary care, hospital outpatient documented as production-grade
Support SLA
Ticketing onlyNo named owner, no response SLA
EnterpriseNamed account team, 4-hour response on P1
Founder-ledFast but unscalable
Named coder opsSenior coder owner per client, weekly accuracy review, written escalation path
Pricing model
Per-chartPunishes complexity, opaque markup
Per-coder seatVolume-blind, expensive at scale
Flat monthlyVolume-fair, predictable
Per-percentageAligned incentives, written floor and ceiling, no per-chart surprises
Section 04 · The vendor disclosure

We are a vendor. Here is what we publish honestly.

We are not above the comparison. We are in it. ASP-RCM sells medical coding AI as part of our broader revenue cycle stack, and the same checklist applies to us. If we cannot give you specialty-by-specialty accuracy on your data, written pricing, an audit trail screenshot, and a 90-day pilot structure inside the first 48 hours, you should drop us from the shortlist.

What we publish: the engine architecture (deterministic HCC engine with a curated V28 crosswalk and hierarchies, not a black-box LLM that hallucinates plausible codes), the tuned specialty list (ABA, HCC risk adjustment, behavioral health, FQHC primary care, hospital outpatient), the coder-in-loop workflow with override telemetry, and the audit trail design that ties every code to its source text, rule citation, and confidence score.

What we do not claim: that we are best on every specialty, that accuracy is 95 percent across the board, or that any AI removes the need for credentialed coders. AHIMA and AAPC are right that human coder review remains the standard of care for production claims. We built our workflow around that, not against it.

Product · Coding engine

Our chart to codes engine.

The actual production page for our HCC and risk-adjustment coding engine. Architecture, V28 crosswalk, UAT results, and the audit trail screenshots.

Product · ABA billing AI

ABA-specific coding and billing AI.

The tuned ABA workflow: 97151 to 97158 code assignment, supervision ratio enforcement, three-way match between auth, session, and claim.

Free buyer scorecard

Get the weighted scorecard.

Send your last 30 days of charts and the vendor list you are evaluating. We return the scorecard filled in for every vendor including us, with evidence per criterion.

Section 05 · Common questions

Frequently asked questions: evaluating coding AI.

How do I run a fair test of a medical coding AI vendor?
Build a blind test set from your own production charts, not the vendor's demo set. Pull 200 to 500 de-identified encounters stratified by specialty, payer, and complexity. Have two of your senior coders independently code the set and reconcile disagreements; that is the gold standard. Then feed the same charts to the vendor and measure agreement at the code level (ICD-10-CM, CPT, HCPCS, modifiers, sequence). Track accuracy by specialty, not aggregate. Demand the vendor give you a flat-rate or no-cost POC so you are not paying to evaluate.
What accuracy number is realistic for medical coding AI?
On a single tuned specialty (often outpatient surgery, ED, or radiology) a competent vendor can show 88 to 94 percent agreement at the code level. Across a generalist book of mixed specialties (ABA, behavioral health, FQHC primary care, hospital inpatient, specialty drugs), realistic generalist accuracy is 65 to 78 percent before human review. Anyone quoting 95 percent or higher on generalist data either has a narrow definition of accuracy (encounter level instead of code level), a curated test set, or is overstating. Ask for the denominator and the specialty mix every time.
Should I ask for a proof of concept or a production pilot?
Both, in sequence. A POC on 200 to 500 charts takes 2 to 4 weeks and tells you whether the model can read your documentation patterns. A production pilot on one specialty or one location for 60 to 90 days tells you whether the workflow holds up under real coder pushback, payer denials, and your EHR's quirks. Skip the POC and you sign a contract on a slide deck. Skip the production pilot and you sign a contract on a sample.
What about HIPAA and PHI in test data?
Three options. First, fully de-identified data per HIPAA Safe Harbor (18 identifier removal) is the cleanest and what most vendors should accept for a POC. Second, a Limited Data Set with a signed Data Use Agreement works if you need dates of service preserved for sequence testing. Third, full PHI under a signed BAA is only appropriate for a production pilot after the vendor passes contract review. Demand the BAA before any PHI moves; do not let urgency push you past the legal step.
How long should a medical coding AI pilot run?
At minimum 60 days, ideally 90. The first 30 days are calibration: the model learns your documentation patterns, your coders learn the workflow, and exception rates settle. The next 30 to 60 days are measurement: stable accuracy by specialty, denial rate compared to your historical baseline, coder time per chart, and payer remit cycle impact. A 30-day pilot tells you almost nothing about real performance because the operational signal has not stabilized yet.
How do I score vendors against each other?
Use a weighted scorecard with five categories: accuracy at the code level on your data (35 percent), audit trail and explainability (20 percent), coder workflow and override mechanics (15 percent), EHR and clearinghouse integration (15 percent), and total cost including coder time savings (15 percent). Score each on a 1 to 5 scale with written evidence per criterion. The vendors that win on slide decks usually lose on audit trail and override mechanics, which is where coder adoption is actually decided.
What happens if AI accuracy degrades after go-live?
Three failure modes are common. First, payer policy drift: an LCD or NCD update changes what is codable, and the model has not been retrained. Second, documentation drift: a new EHR template or a new clinician changes how findings are recorded. Third, specialty expansion: you added a new service line the model was never tuned for. Your contract should require quarterly accuracy reporting by specialty, a defined retraining cadence, and a service credit or termination right if accuracy falls below an agreed floor for two consecutive months.
Who in my organization should own vendor evaluation?
A three-person evaluation team works best. A senior coder or coding manager owns accuracy testing and the gold-standard reconciliation. A revenue cycle leader owns the workflow, denial, and financial impact analysis. An IT or compliance lead owns the security review, BAA, EHR integration, and data ownership terms. The CFO signs off on total cost. Avoid letting a single owner run the evaluation; medical coding AI failures usually show up in a discipline the single owner does not understand.

Get the buyer scorecard. Free.

Send the names of the vendors you are evaluating and 30 days of de-identified charts. We return the weighted scorecard filled in for every vendor on the list, including us, with evidence per criterion and the questions still open against each. A senior coder on the review call. If we are not the right fit, we will tell you.