AI Agents for Dermatology PracticesAn Implementation Guide for 2026

Almost every automation failure we have seen in dermatology starts with the same modelling error: treating the practice as one business. It is two. There is an insurance-billed medical practice running on a per-encounter claim rail, with modifier and bundling exposure on every line. And there is a cash retail aesthetics business running on packages, deposits and promotions, with no payer anywhere in the transaction. They share a waiting room, a phone number, a schedule and frequently a physician.

The same patient can consume both in a single visit. That is where the compliance failures live: the cosmetic portion must be separately consented, separately priced and never bundled into the covered claim. A scheduling agent that does not know which pathway a visit belongs to before the patient arrives is not saving the front desk time — it is moving a determination that used to happen at the desk into a system that has no authority to make it.

The practical consequence for a build is that you do not have one intake workflow with a branch. You have two intake workflows with two rulebooks, two consent artefacts, two pricing disclosures and two definitions of a successful booking, and a triage step in front of them whose job is to route — not to decide coverage.

The second structural fact is volume. Dermatology runs a short-slot, high-throughput clinic — multiple exam rooms per physician, medical assistants rooming ahead — punctuated by long procedure blocks. High volume is what makes administrative automation attractive here. It is also what makes a small per-encounter error rate expensive, because you multiply it by a large denominator every single day.

One number worth keeping in view when you build the business case, with its limits attached: the Sinsky study in Annals of Internal Medicine reported that for every hour of direct clinical face time, physicians spent nearly two additional hours on EHR and desk work. That study observed 57 physicians across 430 hours in four specialties and four states, the authors noted the data came from self-selected, high-performing practices and may not be generalisable, and it was published in 2016. It is a real, citable finding about the shape of the problem. It is not a current dermatology benchmark, and we will not present it as one.

Dermatology's defining loop is biopsy to pathology to result to patient notification to follow-up booking. It is not a clinical problem. It is a tracking problem, and that is exactly why it is the right first automation target and the wrong place to let an agent finish the job.

Consider the structure. A specimen physically leaves the building. Days later the result comes back on a different rail than the one it left on — a lab interface, a fax, a portal login, sometimes all three for different labs. Somebody has to notice it arrived, read it, decide what it means, tell the patient, and convert a malignant result into a scheduled excision or a Mohs case. Every one of those handoffs is a place a result can sit unread while everyone assumes someone else has it.

Here is the loop broken into stages, with what an agent may legitimately do at each one and what stays human. Read the last column as the design contract, not as a suggestion.

Everything an agent can do on this loop runs on read access. Surfacing the backlog, ranking it by age and report category, reconciling sent against returned, detecting non-response to an outreach attempt — all of it computes over data your chart already holds. None of it requires a negotiated write path, which is why it is the cleanest first phase available to a dermatology practice on any certified EHR.

What it cannot do is commit. Creating the follow-up appointment, posting the result acknowledgement, and generating the patient letter into the chart are writes. An agent drafts; a human or a write-enabled integration commits. Hold that distinction and the rest of the architecture falls out of it.

Mohs micrographic surgery is where dermatology's scheduling problem stops being a volume problem and becomes a variance problem. It is a staged same-day procedure: the surgeon excises, processes and reads the tissue, and repeats until the margins are clear. The schedule therefore holds an open-ended block whose length is unknown at the time it is booked.

That unpredictability — not visit volume — is what breaks naive scheduling automation in this specialty. A self-scheduling or voice agent trained on fixed appointment durations will confidently place a Mohs case into a slot that cannot hold it, and the damage propagates down the rest of the day across two or three other patients who did nothing wrong.

The billing side carries its own constraints, and they are worth encoding as pre-submission checks rather than as automated corrections.

A single physician performs both the surgery and pathologic examination of the specimen(s). The Mohs micrographic surgery CPT codes include skin biopsy and excision services (CPT codes 11102-11107, 11600-11646, and 17260-17286) and pathology services (88300-88309, 88329-88332). Reporting these latter codes in addition to the Mohs micrographic surgery CPT codes is inappropriate.

— CMS National Correct Coding Initiative Policy Manual, revision date 1/1/2026, Chapter III

The general biopsy coding structure matters for the same reason. Since 2019 the biopsy codes have been 11102 and 11103 for tangential, 11104 and 11105 for punch, and 11106 and 11107 for incisional, replacing the deleted 11100 and 11101. Only one primary code is reported per encounter regardless of technique, and where techniques are mixed you report the single highest-value primary — incisional above punch above tangential — with add-ons for each additional lesion. We mark this one as consistent across professional restatements rather than quoted from a primary, because the AMA's CPT guidance itself is copyrighted; the code numbers and the 2019 replacement are confirmed in the NCCI manual text.

One caution on a rule you will see cited constantly. A Medicare Administrative Contractor billing article is widely reported to add a CLIA certification number requirement to Mohs claims, and to require the record to show the physician acted in two integrated and distinct capacities, surgeon and pathologist. We were unable to fetch that article: the Medicare Coverage Database returns HTTP 403 to automated requests, re-confirmed on August 12, 2026. Open it in a browser and read it yourself before you encode it as a claim-scrubbing rule. A writer who follows the link with a script will get a 403, and a vendor who quotes it to you may never have opened it either.

This is the single most important architectural fact in medical practice automation, and it is the one most consistently omitted from vendor pitches. Certified API access to your EHR is read-only. ONC/ASTP's own Certification Companion Guide for §170.315(g)(10), last updated May 15, 2026, says so in plain language.

The § 170.315(g)(10) certification criterion requires Health IT Modules to support API-enabled "read" services for single and multiple patients. … These services specifically exclude "write" capabilities, where authenticated and authorized third-party applications would be able to create or modify EHI through a secure API.

— ONC/ASTP Certification Companion Guide, §170.315(g)(10) Standardized API for Patient and Population Services

The consequence is not subtle. An agent can always decide and draft. It cannot always commit. Everything on the commit side of that line — booking the slot, posting the charge, filing the note, generating the letter into the chart, updating the problem list — happens through one of three mechanisms, and you need to know which one you are buying.

The three mechanisms are: a vendor-discretionary write API, which is a commercial arrangement governed by the EHR vendor's app-review process and API terms and not by any certification right; robotic process automation, which is a robot typing into the EHR's user interface; or a human. There is no fourth option, and any specialty guide promising closed-loop automation without saying which of the three it means is selling something.

One forward-looking note, hedged because the source is hedged. ASTP/ONC published a deregulatory proposed rule in the Federal Register on December 29, 2025 — the "HTI-5" rule — which among other things proposes to remove the clinical decision support certification criterion and reserve §170.315(a)(9). Comments closed February 27, 2026 and we located no final rule. Note what it does not do: we found nothing in the sources we could verify that proposes to add a write capability to the certified API. Whether federal API requirements ever move beyond read is something we cannot tell you, and any vendor who tells you it is imminent should be asked for the document. The operating position for a practice buying in 2026 is unchanged: certified write access is not a right today.

Every dermatology agent build needs a written boundary document before the first sprint, and it needs to survive contact with a compliance officer. The organising principle is simpler than the vendor discourse suggests: federal law draws its sharpest line not between AI and not-AI, but between administrative data movement and clinical recommendation.

21 U.S.C. §360j(o)(1)(A) excludes from the device definition software intended for the administrative support of a health care facility, and it enumerates the ground broadly: processing and maintenance of financial records, claims or billing information, appointment schedules, business analytics, information about patient populations, admissions, practice and inventory management, determination of health benefit eligibility, population health management, and laboratory workflow. Scheduling, eligibility, claims data handling, referral logistics, recall lists and lab workflow are outside the device definition by statute — not by agency grace.

The moment software recommends to a clinician, §360j(o)(1)(E) applies and four criteria have to be cleared. And there is a further, harder limit that nothing in federal law relaxes: software cannot be the licensed professional. Medicare's incident-to rule defines auxiliary personnel as an individual who is not OIG-excluded and who meets state licensure requirements. Software is not an individual, cannot be excluded, and cannot hold a licence. There is no incident-to pathway for AI-performed work.

One nuance from FDA's 2026 revision that changes how you should design a dermatology triage flow. FDA now considers both the level of software automation and the time-critical nature of the clinician's decision-making when determining whether Criterion 4 is met — that is, whether the clinician can genuinely review the basis for a recommendation independently. FDA's own framing names automation bias: the propensity of humans to over-rely on a suggestion from an automated system, which increases under time pressure because there is not sufficient time to consider other information.

Read that qualifier carefully, because it is a consideration and not a test. Automation level and time-criticality feed the Criterion 4 analysis; they do not by themselves make a function a device. In FDA's worked examples, the time-critical workflow is always paired with a specific or directive output. The defensible design conclusion is analytical rather than regulatory: the more autonomous and the more time-pressured the workflow, the harder it becomes to claim the clinician is independently reviewing. A recommendation surfaced inside a thirty-second rooming decision is in worse shape than the same recommendation surfaced in an asynchronous inbox.

This is the section the article exists for. What follows is a sequence, not a menu. Each phase has an owner, an entry criterion, an exit criterion and a defined response to failure. Phases overlap slightly by design — Phase 5 starts before Phase 4 finishes, because prior-authorization work is owned by a different person and does not compete for the same attention.

A few notes on why the order is what it is.

• Phase 0 is not optional and it is not a formality: If a practice cannot produce specimens sent, results returned, results unread past its own threshold and notifications outstanding from its own systems, that is the most important finding of the entire engagement. Buying software on top of an uncounted process just makes the count harder later.
• Phase 1 is read-only on purpose: It needs no negotiated write path, so it cannot be blocked by an EHR vendor's commercial terms. It also produces an artefact — a reconciling register — that every later phase depends on for its own quality control.
• Phase 2 comes before anything patient-facing: The medical-versus-cosmetic routing decision touches consent, pricing and claim integrity. Getting the rule set written and disagreement-rate-tested while it is still a back-office flag is far cheaper than discovering it is wrong through a payer audit.
• Phase 3 is the first phase that can hurt a patient: It does not start until Phase 1 has reconciled cleanly for ten business days, because a notification queue built on an unreliable register is worse than no queue at all — it creates a false sense of coverage.
• Phase 4 is where clinical authority must be explicit: Recall intervals are clinical decisions. They get set by clinicians, in writing, per diagnosis, and they do not get adjusted by anyone optimising outreach volume.
• Phase 5 runs on a separate owner and a separate clock: Biologic and phototherapy authorization lives with the biller, not the front desk. Running it in parallel keeps the front-office change budget from being consumed twice over.
• Phase 6 is last because it is hardest: Mohs forecasting depends on stable data from Phases 1 and 3 and on a coordinator willing to test the forecast against their own judgement. If the coordinator wins, the coordinator wins — and you use the agent for packet completeness only.

Vendors sell the happy path. What determines whether a dermatology deployment survives its second quarter is whether somebody wrote down, in advance, what the failure looks like, what signal detects it, and what you do the day it happens. These are the failure modes specific to this specialty, in roughly the order we expect to see them.

On omissions specifically: the evidence says this is the dominant failure mode in ambient documentation, not fabrication. In the OHSU/MedStar simulated-encounter evaluation, 76.3% of errors were omissions, and roughly 19.5% of raw transcription errors survived into the finished note. The MedStar/Georgetown instrument-validation study found 127 errors across 44 draft notes, with 70% of notes containing at least one error and omissions accounting for 83% of one product's errors and 54% of the other's. The authors make the point that matters: omissions are the hardest error for a clinician to catch, because catching one requires recalling a detail from the encounter that the note simply does not mention.

In dermatology that generalises to a specific, checkable risk. Lesion site, laterality and count are exactly the elements a note can quietly drop, and they are exactly the elements that convert a documentation defect into a wrong-site risk. If you run ambient documentation in a procedure room, run a standing audit sample of notes against the procedure log for those three fields, and treat a miss as an incident.

On prompt injection: frame it as blast-radius reduction, never as a solved problem. The controlled JAMA Network Open study found injection attacks succeeded in 94.4% of evaluations at turn 4 and persisted in 69.4% of follow-ups, including 91.7% success in extremely high-harm scenarios. It is a simulation with named limits — the main experiment used lightweight models with a small proof-of-concept on flagship models, and three co-authors disclose company roles — and it is still the best-sourced evidence available, and it is not a vendor claim. The Nature Communications work on vision-language models in oncology adds the finding that breaks the standard mitigation story: sub-visual prompts embedded in medical imaging data can cause harmful output and are non-obvious to human observers. A human-in-the-loop who cannot see the injection cannot review it away.

What is real as mitigation is architectural, because it does not depend on detecting the attack: least privilege enforced at the API boundary, read-only by default (an agent that cannot write cannot be injected into writing), a human signature as the only path from draft to record, unique agent identity plus audit controls so a successful injection is reconstructable afterwards, and out-of-band confirmation for irreversible actions. What is marketing is any vendor-published injection detection or guardrail accuracy rate. We could locate no independent benchmark for clinical prompt-injection defence, and every figure we encountered traced to a vendor.

Ambient documentation is usually the first AI product a dermatology practice is offered, and it sits slightly outside the administrative agent story — but the governance questions overlap enough that you should settle them once.

Start with what is actually known. The strongest independent evidence in this category is a randomized trial published in NEJM AI in December 2025 at UCLA Health: 238 outpatient physicians across 14 specialties, three arms comparing two named products against usual care. The primary endpoint was change in log writing time-in-note. One product showed a statistically significant reduction of 9.5% versus control; the other showed −1.7% and was not statistically significant. Accuracy in that trial was self-reported on a five-point scale, not measured. Read that carefully: the strongest evidence in the category found that a market-leading enterprise product produced no statistically significant reduction in note-writing time, and neither arm measured whether the note was correct.

Now the part that is not a matter of opinion. CMS has already answered who signs, and it names AI explicitly. The Medicare Program Integrity Manual, Chapter 3, §3.3.2.4, effective January 17, 2025, states that when a scribe documents medical record entries CMS does not require the scribe to sign, because the treating practitioner's signature indicates the practitioner affirms the note adequately documents the care provided — and it adds that this type of practitioner concurrence is also required when using artificial intelligence technology to capture the transcription of medical record entries.

The same section closes every workaround. Reviewers will not consider an attestation statement from someone other than the author of the entry, even within the same group. They will not consider an attestation with no associated medical record entry. And an attestation cannot be used to backdate a plan of care. The signature is the act that converts a draft into a billable, legally operative record, and it is a personal, credentialed, non-delegable act. There is no configuration in which an agent attests.

Consent is the live legal risk, and it is separate from HIPAA. Recording a clinical encounter is governed by state wiretap and eavesdropping statutes; HIPAA can permit the use under a BAA while state law still makes the recording itself unlawful. Two putative class actions over ambient recording consent — one filed in San Diego in late November 2025, one filed in the Northern District of California on April 7, 2026 — name health systems, not the vendor, as defendants. The 2026 complaint pleads the California Invasion of Privacy Act, the Confidentiality of Medical Information Act and the state unfair-competition law alongside the federal Wiretap Act, on the theory that the violation occurs at the moment of interception rather than at any later use of the data. We are relying on law-firm and trade analysis for the docket details, not on the dockets themselves; we did not verify the current status of either matter, and we cite no case number. The operational lesson is narrow and useful: the practice carries the consent risk, and a consent checkbox defaulted on inside the vendor's tool is not patient consent.

Four things a dermatology practice should have straight before a vendor tells them otherwise.

FDA revised its Clinical Decision Support guidance, but it did not bless AI. The operative document is dated January 29, 2026 and supersedes a January 6, 2026 version, which itself replaced the September 2022 guidance. FDA did not change the statute or reclassify single-output software as non-device: software providing a specific directive output still fails Criterion 3. What FDA announced is an enforcement discretion policy over that failure — in its own words at its March 11, 2026 town hall, if only one recommendation is clinically appropriate and the function otherwise meets all the §520(o)(1)(E) criteria, FDA does not intend to enforce. Discretion is revocable without notice-and-comment; a statutory exclusion is not. Write it as discretion, never as approval or clearance. FDA also said nothing new about AI or large language models specifically, and nothing about patient-facing tools.

One FDA example is worth memorising for any dermatology practice contemplating a clinical layer: a function that generates a clinical recommendation from information that cannot be verified to come from well-understood and accepted sources fails Criterion 2 and is a device. A model producing clinical recommendations from its own parametric memory, rather than from retrievable guidelines and peer-reviewed literature, is FDA's stated example of exactly that.

The HIPAA Security Rule you are held to is the old one.The NPRM published January 6, 2025 remains proposed; it sits on the Unified Agenda's Long-Term Actions list under RIN 0945-AA22 with final action projected for July 2027. Much of the market writes as though mandatory MFA, asset inventory and encryption at rest are already law. They are not. What binds you today is 45 CFR §164.312 — access control, unique user identification, audit controls, integrity, person-or-entity authentication and transmission security — with encryption still addressable rather than required. Do not let a vendor sell you compliance with a rule that does not exist yet, and do not skip the controls that do.

CMS-0057-F binds payers, not you.Every obligation in 42 CFR §422.122 runs to a Medicare Advantage organization. Live now: specific denial reasons since January 1, 2026, public prior-authorization metrics, and decision timeframes of seven calendar days standard and 72 hours expedited. The FHIR prior-authorization API is a January 1, 2027 obligation — on the payer. Your practice's only relationship to this rule is as the recipient of a denial reason and, from 2027, as a potential consumer of a payer's API. And note the scope limit that matters most in dermatology: commercial and ERISA plans are outside the rule entirely, so if your biologic authorization pain is commercial, this rule changes nothing for you.

There is a federal nondiscrimination duty most practices have never heard of. 45 CFR §92.210 is in the current CFR and prohibits a covered entity from discriminating through the use of patient care decision support tools, imposes an ongoing duty to make reasonable efforts to identify tools using input variables measuring race, colour, national origin, sex, age or disability, and requires reasonable efforts to mitigate identified risk. It applies whether or not the tool is AI and whether or not FDA calls it a device. The reported compliance date is May 1, 2025 and we could not verify whether the section is subject to any injunction or pending rescission — we assert only that it appears in the current CFR. For dermatology this is not abstract: any tool that stratifies skin-cancer risk or prioritises surveillance recall is a patient care decision support tool, and you owe an inventory and a mitigation effort.

The state layer is mostly disclosure. California's AB 3030, in force since January 1, 2025, requires a GenAI disclaimer and human-contact instructions on communications pertaining to patient clinical information, with two exits that matter operationally: it does not apply if a licensed provider read and reviewed the communication, and "patient clinical information" expressly excludes administrative matters including appointment scheduling and billing. So an agent confirming an appointment needs no disclaimer; an agent answering a question about a rash does, unless a licensed human read it first. Texas HB 149 §552.051(f), in force since January 1, 2026, is broader on its face and has no human-review exemption written into it. Texas SB 1188 permits diagnostic AI use conditioned on the practitioner acting within their licence regardless of the use of AI, reviewing all records created with AI, and disclosing the use to patients. That phrase — regardless of the use of artificial intelligence — is the whole doctrine in five words.

Naming what we could not confirm is more useful to an operator than filling the gap with something that sounds authoritative. Here is the honest list for this article, as of August 12, 2026.

• Dermatology cosmetic-versus-medical revenue mix: We found no non-vendor source for any percentage. We describe the structure and refuse the ratio.
• The Mohs CLIA and dual-capacity requirements: Widely reported from a MAC billing article we could not fetch: the Medicare Coverage Database returns HTTP 403 to automated requests, re-confirmed today. Open the article in a browser and verify by hand before encoding it as a rule.
• Any independent accuracy benchmark for dermatology AI: None located, for scribes, agents or coding. The best independent scribe evaluations anonymised the products and none was dermatology-specific.
• Any independent evaluation of administrative agent platforms: We searched for randomized, peer-reviewed or third-party evaluations of multi-step administrative agents in medical practices and found none. Every automation rate, resolution rate and touchless rate in the market is vendor-defined.
• Exact Medicare payment amounts: We print no Medicare dollar figures for any code in this article. Physician Fee Schedule Addendum B is the authoritative source and we did not fetch it.
• Whether any specific vendor will sign a BAA, and on which tier: BAA availability is sourced only to each vendor's own terms and changes without notice. Re-check on the day you sign, in writing.
• Prompt injection through faxes, referral packets, portal messages and payer portals: The published literature covers patient dialogue and imaging. We found no study of injection through those practice channels in a live setting. Extrapolating to them is analysis, not evidence, and we have written it that way.
• Colorado's enforcement pause, as distinct from its statute: The statute itself is settled and stated in the FAQ: SB 24-205 took effect June 30, 2026 and SB 26-189 repeals and reenacts it effective January 1, 2027 under enrolled SECTION 5, with a safety clause at SECTION 6. What we could not verify at source is the enforcement posture — the April 27, 2026 federal court order reported to have granted a joint motion in which the Attorney General stated he does not intend to enforce SB 24-205 until rulemaking concludes. We did not fetch that order or its docket and we name no case. The delayed June 30, 2026 start date likewise rests on law-firm trackers rather than on a primary we pulled ourselves.
• Whether 45 CFR §92.210 is subject to an injunction or pending rescission: We could not verify either way. We assert only that it appears in the current CFR.
• Any claim that FDA approved, cleared or blessed AI documentation or CDS: FDA granted enforcement discretion over a criterion failure. That is not clearance, and we refuse the word.

Use this as an evaluation grid, not as a list of villains. Most of these appear in otherwise reasonable products, and most are answerable — the point is to get the answer before you sign rather than after.

Two further durability checks that have nothing to do with the software. First, corporate stability from public record. This market has already produced a well-funded platform that wound down and sold its assets in pieces, and an administrative automation product whose standalone services were discontinued after a private-equity roll-up — reported by a trade outlet rather than confirmed by the company, but with customers given a short window to migrate. Nothing public suggested that product had failed technically. A practice's continuity risk in this category is often a cap-table risk, not a technology risk. Ask who owns the vendor, what else that owner owns, and what happens to your workflow if the product is folded into a portfolio.

Second, ask what your own EHR already ships. In 2026 the EHR vendors moved into the agent layer themselves, which changes the buying question from "which agent platform" to "what does my EHR already include, when, and at what price — and what is genuinely left over for a third party?" For practices on a specialty EHR the leftover is usually multi-system orchestration: payer portals, faxes, phones and lab rails that the EHR does not touch. That is a real answer, and it is a better basis for a purchase than a feature grid.

These are Frenchy Digital's ranges for build engagements in regulated healthcare environments. They are ranges because scope genuinely varies; they are published because a practice should not have to enter a sales cycle to find out whether a conversation is worth having.

• Senior-led delivery at $150–$225 per hour: We do not staff junior engineers on regulated healthcare work.
• Retainer at $2,500–$9,500 per month: Optional after launch: interface monitoring, payer policy change tracking, override-rate reporting and audit-log review.
• 30-day post-launch warranty: Included on every engagement.
• Written fixed-price phased proposal within 5 business days: Phases are priced individually so you can stop after any one of them.
• Full source-code and IP ownership transfers to you: Code, prompts, policies, documentation and infrastructure accounts. No lock-in and no hostage data.

Booking is at calendly.com/frenchydigital/discovery-call or +1 (424) 272-5601.

Frenchy Digital is a senior-led, Black-owned Los Angeles agency building administrative automation for medical practices, clinics and healthcare operators. For a dermatology practice, an engagement looks like this.

• Discovery that produces your four numbers: A structured call followed by a workflow audit that counts specimens sent, results returned, results unread past your own threshold, and notifications outstanding — from your data, not a national average, and put in writing.
• The write-path answer, in writing, from your EHR vendor: We send the question, we chase the answer, and we design against the answer you actually get rather than the one a vendor implied.
• A boundary document before the first sprint: Which decisions the agent may take alone, which need review, which it must never touch — signed by your clinical lead and reviewable by your compliance officer.
• Read-only first, always: Phase 1 ships without a negotiated write path so it cannot be held hostage by commercial terms, and so the first thing you learn is whether your data reconciles.
• Compliance controls designed in: Signed BAA before data moves, minimum-necessary handling in prompts and logs, unique agent identity, audit logging on every action, and human release on every clinical communication.
• A written limitations statement at delivery: An explicit statement of what the agent does not do and must never decide. It is what your compliance officer will ask for, and it is what keeps the boundary enforceable after we leave.

Book a free 60-minute discovery call or call +1 (424) 272-5601. You will leave with a clear read on which parts of your dermatology workflow are worth automating in 2026, which ones are waiting on a write path you do not control, and which ones should stay human permanently.

Dermatology is a good candidate for administrative automation for an unglamorous reason: it has a large volume of genuinely clerical work sitting on top of a tracking problem with real safety consequences. The pathology loop is where an agent earns its keep, and it earns it by making outstanding work visible rather than by making decisions.

Build in this order. Count first. Then a read-only register that reconciles. Then the routing flag that keeps the medical and cosmetic funnels apart. Then notification drafting with a named human on the release. Then recall. Then authorization. Then, only when the earlier phases are stable, Mohs forecasting. Every phase carries a BAA, minimum-necessary discipline, a unique agent identity, an audit log and override telemetry — regardless of what it builds.

And hold three lines. Certified API access is read-only, so an agent can decide and draft but not always commit — find out which writes your EHR vendor will actually permit before you design around them. The medical-versus-cosmetic coverage determination stays human, because it is a coverage judgement and a directive output. And the signature stays human, because CMS has already said so in a document that names artificial intelligence by name, and nothing in a vendor contract moves it.

The rest is measurable. Refuse the accuracy numbers, measure your own baseline, and let the four counts from Phase 0 be the business case. That is a harder sell than a percentage on a slide, and it is the only version that survives contact with your own data.