You should not have to trust us.
How Scribeable actually works — architecture, BAAs, models, prompt structure, pricing logic, data retention, and the things Scribeable does not do well yet. Every claim on this page carries a date. Every claim is verifiable.
As of 2026-10-28. Reviewed quarterly.
Our Architecture
Your audio and PHI travel a specific path. Here is every step.
Audio capture (iOS / web / Apple Watch)
Recording happens on your device. Rounding Mode uploads and live web cockpit streaming are envelope-encrypted before they leave the device. Finished web cockpit recordings and single-visit iPhone recordings are sent over TLS, so they are encrypted in transit.
Encrypted upload
Audio and any PHI fields travel over envelope-encrypted channels (AES-256-GCM, ECDH P-256 key exchange). On those channels the CDN, edge, and reverse proxies see only ciphertext; calculator inputs and practice roster CSV uploads are protected by TLS only.
Transcription
Deepgram nova-3-medical (BAA covered) performs transcription; our servers decrypt the audio in memory to send it. Transcripts are encrypted at rest immediately after generation.
Two-stage LLM note generation
Anthropic Claude (BAA covered) runs a two-pass pipeline: Stage 1 drafts the note; Stage 2 validates calculators, checks clinical reasoning, and anchors billing codes to specific sentences in the draft.
Storage (Firestore + Firebase + KMS)
PHI is stored field-level-encrypted on Google Cloud (HIPAA-eligible, BAA executed). Each organization has its own data encryption key (DEK), wrapped by a master key encryption key (KEK) managed in Google Cloud KMS.
Retrieval & EHR insertion
Notes are decrypted only at the moment of retrieval, using keys negotiated via ECDH. Browser-extension insertion into the EHR happens on the clinician device. The one exception is eClinicalWorks: for a practice that connects its eClinicalWorks, Scribeable servers read the patient's chart from eClinicalWorks and file the finished note into it. eClinicalWorks is a trademark of eClinicalWorks, LLC; Scribeable is not affiliated with or certified by eClinicalWorks.
For the full encryption architecture — see the Security page — including the 5-layer envelope encryption model, ECDH P-256 key exchange, and enforcement-mode details.
For live uptime, latency, and incident history — see the Status page. Updated continuously.
Our BAAs
Every vendor that touches PHI has an executed Business Associate Agreement. Named counterparties — not a vague "we have BAAs in place."
| Counterparty | Scope | Status |
|---|---|---|
| Anthropic | LLM inference for all note-generation and review stages (Claude Sonnet 5.5, Haiku 5.5, Opus 5.5). | Active |
| OpenAI | The OpenAI API, with Modified Retention ("Eyes Off") active. As of this page's date, no production task sends patient data to OpenAI: order extraction and patient status classification were tested on GPT-6 Luna and stay on Claude Haiku 5.5. | Active |
| Google Cloud (Firebase / Firestore / Cloud KMS) | PHI storage, field-level encryption key management, authentication. | Active |
| Deepgram | Medical transcription (nova-3-medical model). Encrypted audio streaming. | Active |
| OVH Public Cloud (HIPAA) | Application server hosting for HIPAA-scoped infrastructure. | Active |
| Backblaze | B2 cloud storage: encrypted database backups and archived audit logs. | Active |
BAAs are executed directly between Scribeable LLC and each named vendor. Copies are available under NDA for covered-entity diligence — contact [email protected].
Our LLM Choices
Which models we use, for what, and what they may and may not do with your data.
Claude Sonnet 5.5
AnthropicDrafting notes that are not simple enough for Haiku 5.5. The verification pass when the note uses chart context, and for case reports, case presentations and Deep Consult notes. Formatting a note into your template. Building a template from a pasted note, and checking that template for leftover patient details. The second pass of real-time decision support.
Claude Haiku 5.5
AnthropicDrafting simpler notes. The verification pass when no chart context is used. Rounding Mode patient splitting and matching. After-visit summaries and letters. Extracting suggested orders from a note for your approval, and classifying patient status. Clarification answer suggestions, the clinical consistency check, and literature search queries. The first pass of real-time decision support.
Claude Opus 5.5
AnthropicThe first draft of case reports, case presentations, and Deep Consult notes.
Deepgram nova-3-medical
DeepgramMedical speech-to-text with clinical vocabulary and speaker diarization.
Our Prompt Architecture
How prompts are composed. Structure is public. Verbatim content is not — partial disclosure is a deliberate tradeoff between transparency and clinical risk.
Per-specialty prompt composition
Prompts are composed per encounter from (a) a specialty-specific clinical scaffold, (b) a care-setting layer (outpatient/inpatient/ED), (c) a note-type layer (H&P, progress, ED, SOAP, etc.), and (d) the organization's quality configuration.
Two-pass verification
Stage 1 generates. Stage 2 validates — calculators scored, HCC/MIPS codes anchored to specific draft sentences, quality-gap opportunities surfaced. Stage 2 outputs corrections that the Stage 1 model cannot overrule.
Chart-aware context
Prior encounters, medications, allergies, recent labs, and open quality gaps are injected as context. Context carries across encounters for the same patient.
What we do not publish
Specific prompt text is not published. Publishing verbatim prompts creates clinical risk (competitors produce lookalikes without clinical review) and no educational value. Structure is public; content is protected.
Our Benchmarks
Honest about where we are. Published benchmarks will appear at scribeable.ai/research with full methodology.
Calculator validation accuracy
Methodology: inputs extracted by AI vs. inputs verified by attending; scored against reference implementation. Full methodology and dataset will publish with the study.
Note-quality blinded review
Blinded attending review: AI-drafted vs. attending-written notes, from early users who consented to participate. Results will be published on scribeable.ai/research with full methodology when complete.
Hallucination rate
We have not yet published a hallucination benchmark. We will when the methodology is reproducible — "trust me" metrics get challenged and should.
Our Pricing Logic
Every tier, to the dollar. Every cost driver, named.
Free Trial
14 daysFull Pro features for 14 days, capped at 150 notes. No credit card is required to start. If you don't subscribe when the trial ends, nothing is charged and your account goes read-only, with every note kept, until you do.
Lite
/mo (or $390/yr — 17% off)100 AI notes/month. All note formats. HIPAA + BAA included. Browser extension for any EHR. No Rounding Mode.
Pro
/mo (or $790/yr — 17% off)Unlimited AI notes. After-visit summaries. 35+ document types. ICD-10 & HCC coding. Template marketplace. 2 rounding sessions/week (unlimited during the 14-day trial, fair use applies).
Practice
per providerUnlimited notes, every feature, custom templates, quality reporting and population-health dashboards, custom EHR integration, SSO, dedicated founder support. Per-provider pricing negotiated per engagement.
Where the money goes
LLM inference
Two-pass generation runs primarily on Claude Sonnet 5.5 and Claude Haiku 5.5, with Claude Opus 5.5 for case reports and Deep Consult notes. Cost per note scales with encounter length, specialty complexity, and how many chart-context passes run.
Transcription
Deepgram nova-3-medical charges per minute of audio. Rounding Mode amortizes this across many patients.
Storage + compute
Firestore, Cloud Functions, KMS, and HIPAA-scoped hosting on OVH. Costs grow linearly with stored encounters and user count.
Support
Support is answered by the founder directly through 2026. Scales to a support team as usage grows.
Gross margin: We will publish a gross-margin range in the first quarterly transparency report (Q4 2026). We are not yet publishing a number because we want it to be one we can stand behind with audited math — not an estimate. The honest answer today is that per-note margin is positive across all tiers; precise tier-level numbers will appear in the quarterly report.
Our Data Retention
What we keep, for how long, how you export, how you delete.
Encounter audio
iPhone: the recording stays on your iPhone until you delete the note or sign out. Web: your browser keeps a backup copy on your computer for up to 2 hours so an interrupted upload can be retried. Our servers process single-visit audio in memory and do not store it. Rounding Mode: audio is stored on our servers, encrypted, until it is transcribed, and is deleted within 24 hours after transcription. Audio that fails to transcribe is deleted after 7 days. If a technical fault stops us reading a recording, we hold it, encrypted, until we can retry it or destroy it. Fixing the fault does not involve listening to it. There is no option to save a recording. When you close your account, recordings on our servers are destroyed at once.
Generated notes
Stored encrypted for the life of the account. Exportable at any time in structured JSON or plain text.
Transcripts
Stored encrypted, linked to the associated encounter. Deletable with the encounter.
Account deletion
When you close your account, voice recordings, rounding data and cached EHR data are destroyed at once. For your patient records, you choose. On an individual account you can choose destruction: the records are destroyed from our live systems on or after the 30th day after closure, backup copies expire within 28 days after that, and we email you a certificate. If you do not choose destruction, we keep the records (Privacy Policy §5.2), and you can contact support for a copy or to have them destroyed. If your account belongs to a practice, the practice keeps the records. Audit logs, consent records and billing records are kept either way, as the law requires.
Export
Full account export (all notes, templates, documents) is available from the web dashboard at any time. Export is available even during cancellation.
Our Limitations
Honesty beats overselling. Things Scribeable does not do well yet.
Our Incident History
Security incidents, outages, and accuracy regressions will be logged here as they occur — not retroactively summarized.
Fabricated note from an empty transcript (accuracy incident, not a HIPAA-reportable event)
A clinician's dictation was lost on the device before it reached us. The empty generate request was blocked, then a "Generate anyway" override built for short recordings removed the block with nothing in its place, and the draft model composed a complete telehealth encounter from no source. The verification pass asked one question and did not catch it. Contained the same day: both copies of the note and every derived document were held, none had been shared, and nothing reached the patient. What changed: before the draft model is called, the system now measures the actual source content and, if it is essentially empty, does not call the model; the clinician gets a skeleton with clarification questions and the credit back. Held notes no longer feed later notes as history.
No reportable HIPAA breaches
We have not had a HIPAA-reportable breach. We commit to publishing incident summaries on this page when and if they occur, within the HIPAA-mandated notification timeline.
Quarterly transparency report
Starting Q4 2026 we publish an 8–10 page quarterly PDF: uptime, latency, accuracy benchmark updates, incidents, customer-count (anonymized), support response metrics, feature shipping cadence.
Found something wrong? Tell us.
A page like this only works if it is accurate. If you spot a factual error, an out-of-date claim, or a gap, email [email protected]. We will correct within five business days and log the correction in the next quarterly transparency report.
Last reviewed: 2026-10-28 · Next scheduled review: quarterly
Now that you've verified, try it.
Your account is yours — not your employer's.