Telemedicine App Development — AI Scribe Engineering Playbook

Published 2026-06-03 · 23 min read · By Nikolay Sapunov, CEO at Fora Soft

Why this matters

The market is large and moving fast: the global telemedicine market was about $141 billion in 2024 and is forecast to reach roughly $380 billion by 2030, and the narrower market for AI clinical-documentation tools is growing even faster off a smaller base. If you build, run, or are scoping a telehealth product — a virtual-visit app, a specialist consult platform, a behavioral-health service, or a remote-monitoring portal — "add an AI scribe" is now on the roadmap, and the questions behind it are real engineering decisions: do we embed a vendor, assemble our own on speech and language APIs, or build from open models; how do we keep it accurate; and what does the law require before a single visit is recorded. This playbook answers those for the telemedicine vertical specifically. It is written so a product manager can plan the feature and its risk posture without a medical or engineering degree, and so an engineer can see exactly where the scribe taps into a video call and where it can go wrong. The deeper lessons in this section are the per-component manuals; this is the vertical map that tells you which one to open.

What an AI medical scribe actually is

A medical scribe, in the original sense, is a person who sits in the exam room and writes the visit note so the doctor can keep their eyes on the patient instead of a keyboard. An AI medical scribe — also called an ambient scribe, because it listens to the room in the background — is software that does the same job: it captures the conversation, then produces a structured clinical note for the clinician to check and approve.

The reason this feature exists is a number that has not improved in years. For every eight hours of scheduled patient time, a primary-care physician spends an extra 5.3 hours in the electronic health record — the software system that stores a patient's chart, usually shortened to EHR — and about 2.1 of those hours go to writing notes. Much of it spills past the end of the day into what clinicians call "pajama time," the documentation done at home after dinner. As of 2025, roughly one in five physicians still spent more than eight hours a week in the EHR outside normal hours, the same share as in 2022.

That clerical load is a leading driver of burnout, and burnout is what an AI scribe is really sold against. A 2025 quality-improvement study published in JAMA Network Open followed 263 clinicians across six US health systems who used an ambient AI scribe for 30 days; the share reporting burnout fell from 51.9% to 38.8%, and the odds of burnout were 74% lower with the tool in use. The same study noted lower cognitive load and less after-hours documentation, and concluded that ambient AI may be scalable "at a lower cost than human scribes." That last phrase is the business case in five words.

Adoption has followed the pain. By 2025 more than 40,000 US clinicians were using an ambient AI scribe, up from fewer than 8,000 in 2022 — roughly a five-fold rise in three years. When a feature grows that fast, every telehealth product gets asked for it.

The telemedicine wrinkle: the audio is already in your pipeline

Here is the insight that ties "AI scribe" to "telemedicine app development," and it is the part most generic scribe articles miss. In a physical clinic, an ambient scribe is hard to do well: it needs a room microphone, and it has to fight far-field audio, two or three people talking over each other, and the clatter of equipment. A video visit hands you the opposite situation.

In a telemedicine app the conversation already travels as clean, digital, per-person audio through your real-time pipeline — the WebRTC stack that carries the call. Real-time communication on the web, WebRTC for short, sends each participant on a separate media track. That means the clinician's voice and the patient's voice arrive as two distinct streams, each captured close to the speaker's own microphone.

Three things fall out of that, all of them helpful. First, you get near-studio audio per person instead of one muddy room recording, so transcription is more accurate. Second, you get diarization — the task of labeling who said which words — almost for free, because each speaker is already a separate track rather than a voice your software has to tell apart after the fact. Third, you already own the natural place to tap the audio and the natural place to ask for consent: the call itself. A server-side participant that joins the session, or a hook in your selective forwarding unit — the media server, or SFU, that routes streams in a group call — can capture exactly the audio the scribe needs. The mechanics are the same ones we cover in the SFU-side ASR fan-out lesson.

The practical takeaway: a telehealth product is the easiest place to add an ambient scribe well, because you are not retrofitting a microphone into a noisy room — you are tapping a clean stream you already have.

The four-stage pipeline every scribe runs

Strip away the branding and Freed, Abridge, Suki, Dragon Copilot, and Heidi all run the same assembly line. Understanding the four stages is what lets you reason about accuracy, cost, and which stage to build versus buy.

Figure 1. The four-stage AI scribe pipeline. In a telemedicine app, stage zero is nearly free because the call already delivers clean, per-speaker audio.

Stage one is transcription. Automatic speech recognition — the technology that turns speech into text, abbreviated ASR — converts each audio track into a written transcript with timestamps. Medical transcription is harder than ordinary dictation: the system has to get drug names and dosages right, expand clinical abbreviations, cope with accents, and keep speakers apart. The engineering of production speech recognition is its own subject, covered in the streaming ASR lesson, and the speaker-labeling step in the WhisperX diarization lesson and the Pyannote diarization lesson.

Stage two is structuring — turning a raw transcript into a clinical note. The safe way to do this is in two steps, not one. First a model extracts the clinical facts — symptoms, findings, medications, the plan — and ties each one to the exact span of transcript it came from. Then a language model writes the note from those extracted facts, usually into the standard SOAP format: Subjective (what the patient reports), Objective (what the clinician observes), Assessment (the diagnosis), and Plan (next steps). Extracting first and generating second matters because every line in the finished note can then be traced back to something that was actually said. We return to why that ordering is the difference between safe and unsafe below.

Stage three is review and sign. The clinician reads the draft, corrects anything wrong, and signs it. Only then does the note become part of the medical record, and only then is it pushed into the EHR. This stage is not a formality bolted on at the end — it is the load-bearing wall of the whole design, which is the next section.

The one rule that runs through everything: the model drafts, the clinician signs

In a video call, the decision that drives the architecture is where a feature runs, because latency is the binding constraint. In an OTT catalog it is when a feature runs, because cost is the constraint. In a medical scribe the binding constraint is trust, and the rule that follows is blunt: the model drafts, the clinician signs, and nothing is filed automatically.

Figure 2. The sign-off gate. The draft is never the record; the clinician's signature is. Three consequences — liability, regulatory status, and accuracy — all hang on this one gate.

Three consequences hang on that single gate, and they are why it is non-negotiable rather than a nicety.

The first is liability. The clinician who signs the note owns it, exactly as they would own a note they typed themselves. Responsibility for what is in the chart stays with a licensed human, not with a vendor's model. Design the product so that signing is a deliberate act, not a default that scrolls by.

The second is regulatory status. Because the scribe only assists and the clinician makes every decision, most AI scribes today are treated as administrative tools rather than medical devices, which keeps them outside the US Food and Drug Administration's device oversight. That status is conditional, not permanent: the moment a scribe starts suggesting diagnoses or treatments rather than transcribing what was decided, it edges toward being regulated as software as a medical device. The sign-off gate, and a careful line about what the model is allowed to assert, are what hold the tool on the safe side of that line.

The third is accuracy, and it is the one engineers underestimate. The review step is not quality theater — it is the actual safety control. A language model will occasionally write something fluent and wrong, and the only reliable defense is a human who knows the patient reading the draft before it counts. Everything earlier in the pipeline — clean per-speaker audio, extract-then-generate, traceable evidence — exists to make that review fast and trustworthy, not to replace it.

The accuracy problem you must design around

A note that reads perfectly and contains a drug the patient never mentioned is more dangerous than an obviously messy one, because it invites trust it has not earned. Two failure modes matter, and a telemedicine product has to be built around both.

The first is hallucination at the transcription stage — the model inventing words that were never spoken. This is not hypothetical. A 2024 study presented at the ACM Conference on Fairness, Accountability, and Transparency ran clinical-style audio through OpenAI's widely used Whisper model and found that about 1% of segments contained hallucinated text — fabricated sentences, and in some cases invented medication names or harmful phrases that no one had said. Whisper was, at the time, in use by an estimated 30,000 clinicians across 40 health systems, which is how a 1% rate turns into tens of thousands of corrupted transcripts at scale. The researchers noted that several other commercial speech engines did not show the same behavior — a reminder that the choice of ASR engine is a safety decision, not just a price comparison.

The second is the language model's own errors at the structuring stage — hallucinations, but also omissions, where a real and important detail is dropped. Studies of LLM-based clinical summaries put overall error rates in the low single digits; one analysis reported a 1.47% hallucination rate alongside a 3.45% omission rate. Omissions are sneakier than hallucinations because nothing on the page looks wrong; a fact is simply missing.

The engineering answer to both is the extract-then-generate ordering from stage two. When the model first pulls out structured facts tied to specific transcript spans, and only then writes prose from those facts, you gain two defenses at once: the note is harder to fabricate from, because it is built from extracted evidence rather than free association, and every line can show the clinician the words it came from, which makes review fast instead of a re-listen.

Common pitfall: treating the AI note as the record instead of a draft. A demo where the note comes out clean ninety-nine times in a hundred tempts a team to auto-file it and save the clinician a click. Then the hundredth note carries a wrong dose or a fabricated symptom into a real chart, and a convenience feature becomes a patient-safety incident and a legal exposure. The fix is structural, not a disclaimer: make the signed note the only thing that ever reaches the EHR, make the draft visibly a draft, show the source text behind each line, and never let the system file on its own. Ask of every scribe feature, "could this reach the chart without a human signing it?" — and if the answer is yes, that path is the bug.

The tools buyers name: Freed, Abridge, Suki, Dragon Copilot, Heidi

Product teams rarely arrive asking for "an ambient documentation system." They arrive asking how you compare to a named brand. Five come up most, and they sort cleanly by who they are built for.

Figure 3. The AI scribe landscape, sorted by buyer. The left end self-serves individual clinicians; the right end sells deep EHR integration to health systems.

Freed AI is the name behind the searches. Founded in 2022 and based in San Francisco, Freed is the self-serve end of the market: a clinician can sign up and be generating notes in minutes with no IT involvement, the note arrives one to two minutes after the visit ends, and the clinician reviews it and copies it into their EHR. Its 2026 pricing runs from about $39 a month for a capped starter plan to roughly $79 for unlimited notes and around $104–$119 for a tier that adds EHR push and billing codes. It is built for the solo practitioner and the small practice, which is exactly why "freed ai" draws on the order of tens of thousands of searches a month — individual clinicians shopping for themselves.

Abridge is the enterprise leader. It sells to large health systems and is deployed at Kaiser Permanente, Mayo Clinic, Johns Hopkins, Duke Health, and well over two hundred others; Kaiser's rollout alone reached tens of thousands of physicians. Investors have noticed: a 2025 Series E put its valuation at $5.3 billion, and it has been named Best in KLAS for ambient AI two years running. Where Freed optimizes for instant individual setup, Abridge optimizes for system-wide deployment, billing-grade notes, and integration depth.

Suki is the EHR-integration specialist. Marketed as an AI assistant rather than just a scribe, Suki offers bi-directional, real-time integration with the major EHRs — Epic, Oracle Health (formerly Cerner), athenahealth, and MEDITECH — so finished notes write back into the chart without copy-and-paste, and it can embed directly inside Epic's mobile and desktop apps. It is used across several hundred health systems and is the right reference point when a buyer's first question is "how deep does it go into our EHR?"

Microsoft Dragon Copilot is the answer for shops already standardized on Microsoft and Nuance. In March 2025 Nuance's DAX Copilot merged with Dragon Medical One under the Dragon Copilot brand. It produces specialty-aware drafts, surfaces order suggestions you can manage inside Epic, and generates referral letters and after-visit summaries; pricing sits in the hundreds of dollars per provider per month. When a buyer says "Nuance DAX," this is what they mean.

Heidi Health is the one most relevant to a telehealth product, because it was built to work in both in-person and video visits, across 200-plus specialties and multiple languages, with a free tier beneath a paid clinician plan. It is a useful proof that the telemedicine use case is mainstream — but also a cautionary one: in 2026 its public reviews carried a run of complaints about lost recordings and dropped sessions, a reminder that reliability of the capture stage is a feature, not an afterthought.

Three ways to add a scribe to your telemedicine product

If you are building the telehealth product itself, "add an AI scribe" resolves to one of three routes, and they trade speed against control the same way platform decisions always do.

Figure 4. Three routes to a scribe inside your product. Embedding validates fastest; building keeps the patient audio, the notes, and the model inside your perimeter.

The first route is to embed a vendor — plug a scribe company's software development kit or API into your app so their pipeline runs inside your experience. This is the fastest path, often days to a few weeks, and you inherit the vendor's accuracy work, their EHR connectors, and a signed legal agreement covering patient data. The cost is that the notes, the patient data, and the per-visit margin live in the vendor's system, and you customize only within the limits they expose. It is the right choice when you want to validate that your users will use the feature before you invest in owning it.

The second route is to assemble on managed AI APIs — wire a hosted speech-to-text service to a hosted language model and write the note logic, the SOAP formatting, and the review screen yourself. This is a step up in effort, on the order of weeks to a few months, and it buys real control: you decide the note format, you choose each model, and you control where data flows. The trade is that integration, accuracy tuning, and every compliance agreement are now yours to own and operate.

The third route, and the only one that keeps the product fully yours, is to build on open-weights models — run open speech and language models on your own infrastructure that meets healthcare data rules. This takes the most engineering up front, typically months, but the patient audio and the finished notes never leave your perimeter, there is no per-visit fee to a vendor, and you control the model end to end. It is the route for a telehealth company whose patient data, note quality, and clinical model are the business itself. The per-feature cost method behind all three routes — the token math for the language-model step in particular — is in the real cost of AI in video lesson, and domain adaptation for a clinical model is in the fine-tuning a video VLM lesson.

A worked cost example: per visit is the unit that matters

The right way to reason about scribe cost is per visit, because that is what scales with a growing telehealth practice. The arithmetic is simple and it decides the build-versus-buy conversation, so do it once.

Take a clinician who sees about 20 patients a day, 20 working days a month — roughly 400 visits a month. A human scribe service costs in the rough range of $2,000–$4,000 a month for that clinician. Put the middle of that range over the visit count:

human scribe:  $3,000 / month ÷ 400 visits = $7.50 per visit

Now an embedded AI scribe on an unlimited plan, say $99 a month for that clinician:

embedded AI:   $99 / month ÷ 400 visits ≈ $0.25 per visit

That is the roughly thirty-fold gap that the JAMA authors meant by "scalable at a lower cost than human scribes." If instead you build on your own APIs, the marginal cost is the compute for one visit: transcribing about 15 minutes of audio costs a few cents, and the language-model pass that structures the note costs a few cents more, so the per-visit floor lands in the low tens of cents — similar to the embedded per-visit number, but with the vendor margin and the data-residency question removed.

The numbers above are illustrative and move with vendor plans and API prices; the point is the shape. A human scribe is the expensive unit; any AI route is one to two orders of magnitude cheaper per visit, and the choice among the AI routes is decided less by cost than by who has to hold the patient data and how much of the note you need to control.

The gate every visit passes: consent, HIPAA, and disclosure

This is where a telehealth roadmap quietly becomes a legal one, and where more scribe projects get stopped than for any technical reason. Treat what follows as engineering-relevant context, not legal advice — confirm specifics with a qualified lawyer for the jurisdictions you operate in.

Consent to record is the first gate, and telemedicine makes it harder, not easier. Capturing the audio is a recording, and recording law is separate from health-privacy law. US federal law sets a one-party-consent floor, but a dozen states — including California, Illinois, Florida, Pennsylvania, and Washington — require all parties to consent before a conversation is recorded. A video visit routinely crosses state lines: the clinician may sit in a one-party state while the patient dials in from an all-party state. The cautious and common practice is to apply the stricter of the two, which in practice means getting explicit patient consent before the scribe starts listening, every time. The stakes are real — in Florida, recording without all-party consent is a felony. Engineering consequence: consent is a first-class step in your call flow, captured and logged before the first second of audio is tapped, not a checkbox buried in a terms-of-service page.

HIPAA is the second gate. The Health Insurance Portability and Accountability Act is the US law governing protected health information, or PHI. An AI scribe handles PHI at every stage — it receives audio of the visit, creates a note, stores both, and transmits the note to the EHR — which makes any scribe vendor a "business associate" under HIPAA and requires a signed Business Associate Agreement (BAA) before a single real patient is recorded. A scribe-specific BAA should do more than the generic template: it should explicitly forbid using your patients' data to train or improve the vendor's models, commit to processing only the minimum data necessary, require HIPAA Security Rule safeguards, and ensure that any subcontractor touching the audio is itself under a BAA. If you build rather than buy, you are the one who must implement those safeguards.

Disclosure and transparency are the third gate, and they are where regulation is still settling. Under the EU AI Act — Regulation (EU) 2024/1689 — a system that interacts directly with a person must let them know they are dealing with AI, and its transparency rules take effect on 2 August 2026. Encouragingly for scribes, routine administrative documentation is generally not treated as "high-risk" under the Act's Annex III, and the Act explicitly lightens disclosure duties where AI-generated content "has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility" — which is precisely the sign-off gate from Figure 2. In other words, the design that keeps the tool safe also keeps it on the lighter side of the regulation. The broader regulatory engineering is covered in the EU AI Act lesson and the disclosure and C2PA lesson.

The rule across all three gates is the same one that governs the whole article: the human stays in control. Consent is the patient's control over being recorded; the BAA is your control over where the data goes; the sign-off is the clinician's control over what enters the chart. A scribe that respects all three is a feature; one that skips any of them is a liability.

The playbook: from "add a scribe" to shipped feature

Put the pieces together and adding an AI scribe to a telemedicine product reduces to four questions, asked in order.

Figure 5. The playbook in one path. Capture from the call you already have, decide build-versus-buy on data ownership, enforce extract-then-generate, and pass every visit through the consent-and-sign-off gate.

First, capture: because this is a video product, tap the clean, per-speaker audio already flowing through your call rather than retrofitting a microphone — it is the single biggest quality advantage telemedicine has over in-clinic scribing. Second, build or buy: if the patient audio and the notes must live inside your product, build on open models or assemble on AI APIs; if you are validating the feature or do not need to own the data, embed a vendor and ship in days. Third, the accuracy rule: extract clinical facts tied to transcript spans first, then generate the note, so every line traces back to evidence and review is fast. Fourth, and without exception, the compliance gate: get recording consent before any audio is captured, apply the stricter state's rule on a cross-state visit, sign a BAA that forbids training on your patients' data, and require a clinician signature before anything reaches the chart.

That is the entire playbook. The deeper lessons in this section are the manuals for each box — streaming ASR for the transcription stage, diarization for telling speakers apart, SFU-side ASR fan-out for tapping the call, and the AI in video conferencing playbook for the sibling meeting-notes use case that shares most of this pipeline.

Where Fora Soft fits in

We build the telemedicine and video-conferencing platforms that AI scribes live inside — virtual-visit apps, specialist-consult products, behavioral-health services, and the WebRTC pipelines underneath them — so we run this playbook with clients regularly. When a client wants to validate the feature, we embed a vetted scribe vendor into their visit flow and wire the consent step into the call UI first. When the patient data and the note quality are the product, we build the scribe on an owned pipeline — tapping the clean per-speaker audio from the call, structuring notes with extract-then-generate, and keeping audio and notes inside the client's HIPAA-eligible perimeter — with consent capture and clinician sign-off designed into the flow from the first sprint. The four questions in this playbook are the same ones we weigh in scoping calls when a telehealth client asks whether to buy a scribe or build one.

What to read next

• Streaming ASR in production — Deepgram, Whisper, AssemblyAI
• Live captions — the SFU-side ASR fan-out pattern
• AI in video conferencing — engineering playbook

Talk to us / See our work / Download

• Talk to a video engineer — scope an AI scribe into your telemedicine or telehealth product: book a 30-minute call.
• See our work — telemedicine and video-conferencing platforms we have shipped: our telemedicine app development work.
• Download the Telemedicine AI Scribe Engineering & Compliance Decision Sheet — the four-stage pipeline, the draft-never-file rule, the vendor landscape, the build-versus-buy split, the per-visit cost math, and the consent-and-HIPAA gate on one page: Telemedicine AI Scribe — Engineering & Compliance Decision Sheet.

References

1. American Medical Association — "Burnout on the way down, but 'pajama time' stands still" (2025). Physician burnout 41.9% in 2025 (down from 43.2% in 2024, 48.2% in 2023); ~20.9% of physicians spend >8 hrs/week on the EHR outside work hours, unchanged from 2022. Tier 5 (professional-body survey reporting). https://www.ama-assn.org/practice-management/physician-health/burnout-way-down-pajama-time-stands-still
2. Tai-Seale et al. / health-system EHR time studies, summarized by Tebra "The Intake." For every 8 hours of patient visits, primary-care physicians spend ~5.3 additional hours in the EHR, ~2.1 of them on documentation; ~1.77 hrs/day of after-hours "pajama time." Tier 6 (educational summary of peer-reviewed time-motion studies). https://www.tebra.com/theintake/ehr-emr/how-documentation-became-top-cause-of-physician-burnout
3. JAMA Network Open — "Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout" (published online 2 Oct 2025). Multicenter quality-improvement study, 263 clinicians across 6 US health systems; burnout fell 51.9% → 38.8% over 30 days, odds of burnout 74% lower; ambient AI "may be scalable at a lower cost than human scribes." Tier 5 (peer-reviewed primary). https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2839542
4. Koenecke et al. — "Careless Whisper: Speech-to-Text Hallucination Harms," ACM Conference on Fairness, Accountability, and Transparency (FAccT 2024). Whisper hallucinated text in ~1% of clinical-style audio segments (187 of 13,140), including fabricated medications and harmful content; comparable Google/Amazon/AssemblyAI/RevAI engines did not. Tier 5 (peer-reviewed conference paper). https://www.science.org/content/article/ai-transcription-tools-hallucinate-too
5. "Beyond human ears: navigating the uncharted risks of AI scribes in clinical practice," npj Digital Medicine (2025). LLM clinical-note error modes — hallucination, omission, misattribution; one analysis reported ~1.47% hallucination and ~3.45% omission; most scribes are marketed as administrative tools outside FDA device oversight. Tier 5 (peer-reviewed). https://www.nature.com/articles/s41746-025-01895-6
6. Regulation (EU) 2024/1689 (EU AI Act), Article 50 — Transparency obligations (Official Journal version, 13 June 2024; Article 50 applies from 2 Aug 2026). §50(1) duty to inform a person they are interacting with an AI system; §50(4) lighter disclosure where AI output undergoes human review/editorial control with a responsible person. Read directly from the consolidated article text. Tier 1 (official EU regulation). https://artificialintelligenceact.eu/article/50/
7. EU AI Act — Annex III (High-Risk AI Systems) and Article 6 (classification). Routine administrative tasks such as medical text classification or structuring are generally not high-risk unless the system is a medical-device safety component. Tier 1 (official EU regulation). https://artificialintelligenceact.eu/annex/3/
8. U.S. Department of Health & Human Services — HIPAA Business Associate Contracts (45 CFR §164.504(e)) and the Security Rule (45 CFR Part 164, Subpart C). A vendor that receives, creates, maintains, or transmits PHI is a business associate requiring a BAA and safeguards. Tier 1 (official US regulation / regulator guidance). https://www.hhs.gov/hipaa/for-professionals/covered-entities/sample-business-associate-agreement-provisions/index.html
9. 18 U.S.C. § 2511 (Federal Wiretap Act) and state all-party-consent statutes (e.g., California, Illinois, Florida, Pennsylvania, Washington). Federal one-party floor; all-party states require every party's consent before recording; cross-state telehealth visits typically apply the stricter rule. Tier 1 (official US statute), with state-survey context. https://www.law.cornell.edu/uscode/text/18/2511
10. Freed AI — product and 2026 pricing (getfreed.ai and independent reviews). Ambient scribe founded 2022 (San Francisco); setup in minutes, note in 1–2 minutes; 2026 tiers ~$39 (capped) / ~$79 (unlimited) / ~$104–$119 (adds EHR push, ICD-10); ~29K monthly searches for "freed ai." Tier 7 (vendor + trade reviews). https://www.getfreed.ai/resources/cost-of-ai-scribes
11. TechCrunch / Fierce Healthcare — Abridge funding and deployments (2025). $300M Series E at $5.3B valuation (a16z, Khosla); deployed at Kaiser Permanente, Mayo Clinic, Johns Hopkins, Duke and 250+ systems; Best in KLAS Ambient AI 2025 and 2026. Tier 7 (trade press). https://techcrunch.com/2025/06/24/in-just-4-months-ai-medical-scribe-abridge-doubles-valuation-to-5-3b/
12. Suki — product and EHR integration (suki.ai); Microsoft for Healthcare — Dragon Copilot (formerly Nuance DAX Copilot). Suki: bi-directional integration with Epic, Oracle Health, athenahealth, MEDITECH; 400+ health systems; $70M Series D (Oct 2024). Dragon Copilot: DAX + Dragon Medical One merged Mar 2025; specialty-aware drafts, order suggestions in Epic; ~$369–$830+/provider/mo. Tier 4/7 (vendor). https://www.microsoft.com/en-us/health-solutions/clinical-workflow/dragon-copilot
13. Grand View Research — telemedicine market size; telemedicine app development cost surveys (2026). Telemedicine market ~$141B (2024) → ~$380B (2030), ~17.5% CAGR; compliant telemedicine MVP ~$110K–$230K, enterprise $450K–$1M+; HIPAA cloud hosting ~$2K–$10K/mo. Analyst/agency estimates vary by definition. Tier 7. https://www.grandviewresearch.com/industry-analysis/telemedicine-market
14. Towards Healthcare / Fortune Business Insights — AI medical scribing market size and adoption (2026). AI medical scribing market ~$1.39B (2025) → ~$1.67B (2026) → ~$8.93B (2035), ~20.5% CAGR; >40,000 US clinicians using an ambient scribe in 2025, up from <8,000 in 2022. Estimates vary widely by source and definition. Tier 7 (analyst). https://www.towardshealthcare.com/insights/ai-in-medical-scribing-market-sizing