AI CPT Coding in Dermatology: What It Does — and What It Doesn't
I spent years as a dermatology practice administrator before I built software, and I've sat through enough denial reviews to be skeptical of any tool promising to "automate coding." I've also watched a medical assistant hold up a schedule for eleven minutes deciding whether a lesion sent to pathology was a shave biopsy or a shave removal. Both things are true: AI is genuinely useful in derm billing, and most of what's marketed as AI CPT coding oversells what the technology can responsibly do. Here's the honest version — where AI helps in AI medical billing dermatology work, where it fails, how our own assistant is built, and what the HIPAA answer actually is.
Why CPT Coding Is So Hard and Error-Prone in Derm Billing Today
Dermatology has an unusually punishing code set for its visit volume. A practice seeing 45 patients a day may touch destruction, biopsy, shave, excision, repair, Mohs stage, and injection codes before lunch, and the distinctions that determine payment are small and physical: how the lesion was removed, whether it went to pathology, site, diameter plus margins, repair complexity, lesion count. Three structural problems make it worse than it looks.
The code families changed and the habits didn't. The 2019 biopsy restructure split biopsies into tangential, punch, and incisional families with separate add-on codes for additional lesions. Years later I still see practices billing the old way out of muscle memory.
The person coding usually isn't a coder. The front desk quotes a price, a medical assistant drops charges, and the provider signs off between rooms. There may be one certified coder covering four locations — or a billing company that surfaces problems 45 days later, as denials.
Errors now cost money twice. They cost you the claim, and — since the No Surprises Act — the accuracy of your patient estimate. A Good Faith Estimate built on the wrong CPT code produces a number the patient can dispute when the real bill arrives; the GFE template walkthrough covers those mechanics, and our No Surprises Act compliance guide covers the deadlines and penalties tied to getting it wrong.
So the real question isn't whether AI replaces a coder. It's whether it can shorten the gap between "I know what we did" and "I know what to call it" for the non-coder at the desk right now.
What AI Does Well for Medical Coding Assistance
Used narrowly, AI is very good at a few things that eat staff time.
- Translating clinical language into code candidates. Staff describe procedures the way clinicians say them out loud — "froze six spots on the scalp," "punch off the forearm, sent to path." AI handles that fuzzy mapping well; keyword search handles it badly.
- Explaining descriptors in plain English. AMA descriptors are written for precision, not comprehension. The difference between 11102 and 11302 in one sentence is worth more to a front-desk employee than the descriptor itself.
- Surfacing the question that determines the code. A good assistant answers "it depends on whether the specimen went to pathology" instead of guessing. That prompt alone prevents a lot of misbilling.
- Flagging predictable coverage risk. Cosmetic exclusions, diagnosis requirements for benign lesion destruction, medical-necessity documentation for wart treatment — patterns stable enough that a note at quoting time is preventive.
That's the honest scope of how AI helps derm billing. It compresses lookup and explanation time. It does not make the coding decision.
What AI Cannot (and Should Not) Do in Coding and Billing
These limitations aren't temporary gaps waiting on a better model. Several are structural, and ignoring them means taking on unseen compliance risk.
-
It cannot see the chart, so it cannot verify what happened
An assistant only knows what someone typed into it. If the description says "excision" but the note documents a shave, it can't catch the discrepancy. Code selection must be reconciled against the documentation by a human — always.
-
General-purpose AI invents codes that sound right
Ask a consumer chatbot for a CPT code and you'll get a plausible five-digit number with a confident descriptor. Sometimes it's right, sometimes it's a real code for the wrong procedure, sometimes it doesn't exist — and a non-coder can't tell which. Grounding matters more than model quality.
-
It should not be the final compliance authority
Modifier decisions, medical-necessity judgments, NCCI bundling calls, and anything payer-specific belong to a certified coder. AI output is a starting point for that judgment, never a replacement.
-
It should not auto-submit anything
A suggestion that silently becomes a claim removes the step where errors get caught. Bill without a human confirming and you've automated your error rate along with your throughput.
-
It cannot resolve genuine ambiguity
Some cases are unclear until the provider clarifies intent or pathology returns. The right answer is "ask the provider," and a well-designed assistant says so instead of picking a code to seem helpful.
How DermEstimator's AI Coding Assistant Works
Because generic AI fails in one specific way — inventing codes — we designed our assistant to make that failure mode impossible rather than unlikely. The AI coding and billing assistant isn't a chatbot with a dermatology prompt around it. It's a search over our own fee schedule database with an explanation layer on top.
No CPT lookup, no memorization. Staff type it the way they'd say it to a coworker: procedure type, body site, and anything that changes the code — "sent to pathology," lesion count.
Your description is matched against our 525+ code dermatology database before any response is generated, and the assistant may only discuss codes in that shortlist. It cannot cite a code from general AI knowledge or invent one that isn't in the system. If nothing matches, it says so instead of guessing.
Each candidate returns with its full AMA descriptor, a plain-language explanation of what distinguishes it from neighboring codes, and where relevant a coverage or denial-risk note.
Nothing is added to an estimate automatically. Staff decide which code is right and add it manually. The assistant suggests and explains; it never bills, never submits a claim, never touches a clearinghouse.
That design has a cost worth naming: the assistant is narrow and useless outside dermatology. It's the right trade — a tool that confidently answers everything is a tool that confidently answers wrong. Confirmed codes flow into the estimating engine on our features page, which applies locality-adjusted rates and MPPR. For a specific, frequently underpaid example of why grounded code detail matters, see our breakdown of Mohs add-on codes 17312 & 17314.
The Privacy Question: HIPAA and Patient Data
This is the first question thoughtful administrators ask, and it deserves a direct answer rather than a badge.
The assistant is built for de-identified procedure descriptions — not real patient chart notes. Describe the procedure, not the patient. "Punch biopsy, left forearm, sent to path" is exactly right. Pasting a chart note containing a name, date of birth, MRN, appointment date, or narrative history isn't how the feature is meant to be used, and the interface prompts staff not to do it.
The reasoning is the principle that governs the rest of a HIPAA program: minimum necessary. Choosing the right CPT code requires the clinical facts of the procedure; it never requires knowing who the patient is. If a detail doesn't change the code, it doesn't belong in the box — and that's a staff habit, not a software setting.
Practical rule for staff: if you'd be uncomfortable reading the sentence out loud in a full waiting room, don't type it into the assistant. Procedure details are fine; identifiers are not. For our full stance on encryption and access controls, see the HIPAA & Security page.
One broader point on AI billing software HIPAA diligence: ask any vendor whether prompts train models and whether a BAA covers components that could touch PHI. One who won't answer in writing has told you something.
How to Use AI Coding Well
Rolling this out to staff, these habits separate practices that benefit from practices that create new problems.
- Include the code-determining details. Procedure type, site, lesion count, size if known, pathology disposition. Vague input produces vague output — "removed a mole" spans four code families.
- Treat every suggestion as a candidate. Read the descriptor and confirm it matches what was documented. If the descriptor and the note disagree, the note wins.
- Never paste identifiers. No names, dates of birth, MRNs, appointment dates, or copied chart text. Put this in your onboarding checklist.
- Escalate ambiguity instead of resolving it yourself. If the answer depends on something you don't know, that's a question for the provider — not a prompt to rephrase until you get one answer.
- Keep a coder on the unusual cases. Complex repairs, atypical Mohs stage counts, multi-site excisions, and anything involving modifiers still need coder review before submission.
- Audit a sample monthly. Compare ten AI-assisted estimates against the final billed claims. Twenty minutes tells you whether the gap is the tool, the descriptions, or training.
- Use it to teach. Staff who read why 17110 differs from 17000 by lesion count learn the logic instead of a cheat sheet that goes stale.
The framing I'd offer: AI coding assistance is a fast, patient reference desk — there at 4:50pm on a Friday, and honest when it doesn't know. What it isn't is the person who signs off on the claim. It's on the Derm Pro plan; see plans and pricing, or browse the blog for more on derm coding.