AI medical scribe development · United States
We build AI medical scribes for health tech products and health systems: ambient capture, transcription, grounded note drafts, clinician review and EHR write-back, with consent and audit built in from the first visit. You own the code, the prompts and the data path, whether it ships as your own product, a white label scribe or a scribe API inside your platform.
Record only after consent is captured, with the audio kept for the time your policy allows and no longer.
Transcribe, attribute speakers and draft a note that cites the transcript, in your specialty's template.
The clinician edits and signs, and the note goes to the EHR through the write-back path each EHR supports.
The short answer
Building an AI medical scribe means more than calling a speech model. A production scribe needs consent before recording, transcription that handles real clinic audio, note drafts grounded in what was said, a review and sign-off step, versioned prompts, PHI controls and write-back to the EHR. We build that whole system for health tech companies that want their own or a white label scribe, and for health systems that need a scribe for a specialty the off-the-shelf products do not serve well.
Scribe components
These are the parts we build and the detail that usually decides whether a scribe passes a health system's review.
| Component | What it does | What can go wrong | What we design for |
|---|---|---|---|
| Consent and capture | Records the visit only after consent is captured and stored with the visit | Recording before consent, or no record of who agreed | A consent step that blocks capture, and a stored record per visit |
| Transcription and speakers | Turns audio into text and attributes clinician and patient speech | Accents, noise, cross-talk and medication names | Model choice tested on your own audio, custom vocabulary and confidence flags |
| Grounded drafting | Drafts the note from the transcript in your template | Invented findings, wrong laterality, missed negatives | Drafts that cite the transcript, section rules and checks for unsupported statements |
| Review and sign-off | Lets the clinician edit, accept and sign | Notes filed as final with no clinician action | A note lifecycle where only a recorded clinician action makes a note final |
| Prompt governance | Controls how prompts change over time | A silent prompt edit changes every note | Versioned prompts with preview, publish, rollback and an audit fingerprint on every draft |
| Scribe API | Lets your product or your customers call the scribe | Unclear data ownership and tenant mixing | Per-tenant keys, scoped access, rate limits and a clear contract for audio and note retention |
| EHR write-back | Files the signed note in the chart | Duplicates, wrong encounter, wrong note type | The write-back layer on our AI scribe EHR integration page |
Speech and language models change every few months. We keep them behind one interface so you can switch providers without rebuilding the product.
What we build
Pick the workstreams your product needs. Each one ships with a specification, tests on real clinic audio and an owner.
Mobile, desktop or in-room capture with consent captured first, offline buffering for weak networks and audio retention set by your policy.
Consent record · capture apps · retention rules
HIPAA compliant AI scribe: what it takes to build one →Speech-to-text tested on your own de-identified audio, speaker attribution, medical vocabulary and confidence flags that tell the clinician where to look.
ASR · diarization · custom vocabulary
Note templates per specialty and visit type, drafts that cite the transcript and checks that hold back statements the conversation does not support.
SOAP · HPI · assessment and plan · therapy notes
AI medical scribe comparison →Prompts as versioned, reviewed releases with preview, publish and rollback, and the version recorded on every draft so any note can be explained later.
Prompt versions · audit fingerprint · rollback
Prompt governance for clinical AI →Your brand on the scribe, or an API your product and customers call, with tenant isolation, scoped keys and usage records.
REST API · webhooks · tenant isolation
Signed notes filed in Epic, Oracle Health and athenahealth through the path each supports, on the right encounter, exactly once.
FHIR DocumentReference · HL7 MDM · athenaOne APIs
AI scribe EHR integration →Buyer decisions
Most teams should buy unless the scribe is part of their product or their specialty is poorly served. This is how the options compare.
| Option | Best fit | Strengths | What to plan for |
|---|---|---|---|
| Buy an off-the-shelf scribe | Practices and health systems that need notes now | Fast to start, vendor carries the model work | Fit for your specialty, EHR write-back depth and per-clinician fees |
| Use the EHR vendor's scribe | Health systems on Epic, Oracle Health or athenahealth | Lives in the chart, one vendor | Specialty templates and features still on the vendor roadmap |
| White label a scribe | Health tech companies adding a scribe to their product quickly | Your brand, less model work | Dependence on the white label vendor's roadmap and data terms |
| Build on a speech and model platform | Products where the scribe is a core feature | Full control of templates, data path and UX | Consent, governance, write-back and support are yours to run |
| Build with us | Teams that want ownership without building a scribe team first | You own the code and prompts; we bring the production pieces | A product owner on your side and clinical reviewers for templates |
If an off-the-shelf scribe fits your specialty and EHR, buy it. We say so on the scoping call.
Integration boundaries
A scribe crosses clinical, legal and vendor lines. We write the split down before the build starts.
Owns templates, what a good note looks like and the review of draft quality. We implement the rules exactly as signed off.
Own the consent policy, state recording rules and AI disclosure. Some states require every party to agree before recording, and Texas requires disclosure of AI use in care. We build the consent step and its record.
Own their models and their BAA terms. We pick providers that sign a BAA for PHI and keep them behind one interface.
Signs every note. The scribe drafts; it never finalizes a note on its own.
Capture, consent record, transcription pipeline, drafting, review, prompt governance, API and write-back. Defects in this layer are ours to fix.
Audio, transcripts and notes are PHI. We sign a BAA, encrypt in transit and at rest, keep PHI out of logs and work inside our ISO 27001:2022 certified management system.
Deliverables
Everything lands in your repository and your runbook.
Capture apps, pipeline, drafting, review UI and API, with configuration per tenant kept out of the code.
Specialty templates and versioned prompts with clinical sign-off and a release process.
De-identified test visits with reviewer scores, so you can see draft quality before and after each change.
The consent flow, its stored record and the retention rules for audio, transcripts and drafts.
Data flow, BAA chain, access controls and audit logging, ready for a health system's security questionnaire.
Write-back for your first EHR and a support runbook for the live product.
Delivery phases
Plan for 12 to 16 weeks to a first specialty live with real clinicians. Clinical review of templates and the first EHR write-back set most of that date.
Weeks 1 to 2
Specialties, templates, consent policy, providers and the first EHR agreed
Weeks 2 to 8
Capture, transcription, drafting, review and prompt governance built and scored on test visits
Weeks 7 to 13
Write-back built and tested, pilot clinicians live, draft quality reviewed weekly
Weeks 12 to 16
Monitoring live, runbook and evaluation set handed over, next specialty planned
Cost drivers
You get a fixed scope and price after one scoping call. These are the factors that move it.
Each specialty needs its own templates, test visits and clinical review.
Mobile, desktop and in-room capture each add build and testing.
Hosted provider APIs are quicker to start; self-hosted models add infrastructure and tuning.
Each EHR adds its own write path and sandbox testing.
Multi-tenant branding, keys and usage records add platform work.
Monitoring, evaluation reruns and on-call support are priced per month.
Proof
Client names are withheld. The work below is delivered and described without identifying the client.
Consent before recording, bounded transcription, grounded drafts, clinician sign-off and versioned prompts, built into a live behavioral health EHR.
Read the build guide →How we publish, preview, roll back and fingerprint prompts so every draft can be traced to the prompt version that wrote it.
Read the guide →A voice platform that turns clinical conversations into structured notes for clinicians.
Read the case study →We are ISO 27001:2022 certified and sign Business Associate Agreements for work that touches patient data.
FAQ
A production AI medical scribe includes consent capture, recording, transcription with speaker attribution, specialty templates, note drafts grounded in the transcript, clinician review and sign-off, prompt version control, PHI controls, an API if others will call it, and write-back to the EHR.
Buy one if an off-the-shelf scribe fits your specialty and EHR. Build when the scribe is part of your product, when you need a specialty or workflow the main scribes do not serve, or when you need control of the data path. We say which applies on the scoping call.
A white label AI scribe is a scribe built so another company can ship it under its own brand. It needs tenant isolation, per-customer branding and templates, scoped API keys and clear terms for who holds the audio, transcripts and notes.
Yes. We build a scribe API your product or your customers call to start a session, send audio, get a draft and file a signed note, with per-tenant keys, rate limits, webhooks and usage records.
We test candidate providers on your own de-identified audio and pick the ones that perform best and sign a BAA for PHI. We keep them behind one interface so you can switch providers later without rebuilding the product.
Drafts cite the transcript, section rules check for statements the conversation does not support, the clinician reviews every note, and an evaluation set of test visits is scored before and after each prompt or model change.
HIPAA compliance comes from how the scribe is built and run: a BAA with every vendor that touches PHI, encryption, access controls, audit logs, minimal retention and PHI kept out of logs. There is no HIPAA certification for software; our HIPAA compliant AI scribe guide covers the controls.
Usually, and the rules vary by state. Some states require every party to agree before recording, and Texas requires providers to disclose AI use in care. Your legal team sets the policy, and we build the consent step and keep its record with the visit.
Yes. Epic accepts plain text notes through FHIR on an existing encounter, Oracle Health accepts final notes through FHIR, and athenahealth takes notes through its athenaOne APIs. Our AI scribe EHR integration page covers each path.
Plan for 12 to 16 weeks to a first specialty live with real clinicians. Clinical review of templates and the first EHR write-back set most of that date.
Cost depends on the specialties, capture surfaces, model choices, EHR write-back, white label or API needs and the support level. We give a fixed scope and price after one scoping call.
You do. The code, templates, prompts, evaluation set and runbook land in your repository, and we hand over everything needed for your team to run the scribe without us.
Next step
Tell us the specialties, who will use the scribe and which EHR it must write to. An engineer replies with a build, buy or white label recommendation and a fixed scope.
Start with your workflow
Share the specialties, the users, the EHRs and your target date. Use de-identified examples only. Read by an engineer, answered within one business day.
