No microphone in the room. No PHI in someone else’s cloud.

• Drafts discharge summaries. Type or paste your in-stay shorthand and get a structured draft you can edit. The patient identifiers stay on your Mac, so nothing leaves the room with the audio still running.
• Reads referral letters. Summarise an inbound referral, draft an outbound one to the specialist, pull the relevant prior history into context.
• Drafts clinical correspondence. Letters to GPs, specialist replies, MDT meeting summaries, prior-authorisation appeals, insurance correspondence.
• Anonymises before you forward. Redact patient identifiers before sharing a case for teaching, M&M, or peer review. Each redaction logged in a Word comment with the original value so you can audit it later.
• Translates patient documents. For multilingual practice and international referrals, without sending the document to an online translator.
• Reads long PDFs. Discharge packs, complex referrals, journal articles applied to a current case, NICE/UpToDate-style guidance. Cited back to the page.

The 2025 wave of AI-scribe class actions in California anchors on the same fact pattern: the scribe recorded a patient without explicit consent, allegedly violating CIPA, the CMIA, and the Federal Wiretap Act. The lead case is Saucedo v. Sharp HealthCare, filed in San Diego Superior Court on 26 November 2025, alleging more than 100,000 patients were recorded without authorisation. A September 2024 incident at an Ontario hospital went further: a departed physician’s personal Otter.ai bot joined a hepatology rounds meeting via an unrescinded calendar invite, and the resulting transcript was sent to 65 recipients including 12 who had already left the institution. Seven patients’ PHI was exposed; the Ontario IPC published its findings in December 2024.

HHS OCR’s January 2025 proposed update to the HIPAA Security Rule explicitly puts AI tools inside the formal risk-analysis perimeter for the first time. Wolters Kluwer’s December 2025 Shadow AI in Healthcare survey found 17% of clinicians admit to using unauthorised AI tools and 10% used shadow AI for direct patient care. The pattern is the same in every case: the document or audio went somewhere it shouldn’t have.

For clinicians in Europe and the UK, the same exposure shows up under different rules: GDPR Article 9 special- category health data, the Schrems II transfer-impact assessment for any US-cloud processing, and the country-by-country secondary-use rules under the EHDS framework. On-device sidesteps the lot, because no transfer occurs.

Muet’s position is structural, not procedural. There is no microphone, no audio stream, no cloud round-trip. The patient identifier you type is processed on the same machine you typed it on and goes nowhere else.

• Discharge summary drafting from in-stay shorthand.
• Referral letters in and out, with relevant prior history pulled across.
• Operative note and procedure-summary drafts.
• MDT meeting minutes and case-conference summaries.
• Anonymising case material for teaching, M&M, or publication.
• Insurance and prior-authorisation appeal letters.
• Reading long discharge packs and applying current guidance to a case.
• Bilingual practice where intake forms or referrals arrive in another language.

Abridge, Suki, Ambience, Nuance DAX, DeepCura, and the new ChatGPT for Healthcare all sign BAAs and most integrate with major EHRs. They also stream audio to their cloud. A BAA shifts liability if the data is breached at the vendor; it does not stop the data from being there in the first place. The 2025 California consent lawsuits are about exactly that distinction.

Heidi Health’s “Heidi Remote” mode is the only major competitor moving toward on-device, with an encrypted wearable mic. It still sends transcripts to the cloud for the LLM step. Muet does the LLM step on the Mac too, so neither the audio nor the text leaves.

Cheaper than every BAA-tier cloud scribe per provider per month, and there is no per-encounter cost. A solo practitioner who drafts 30 letters a week recoups the licence in a fortnight.