The Two Hours Nobody Bills For
A physician's day has a visible half and an invisible one. The visible half is eight hours of patients — the work everyone pictures when they think of a doctor. The invisible half comes after, when the clinic has emptied and the notes still aren't written. Two to four hours of typing up what was said, structuring it, coding it, and clicking it into the electronic health record. The profession has a name for it: "pyjama time," because that's where it happens.
That invisible half is one of the largest single contributors to physician burnout, and it produces no clinical value on its own. The patient was already seen. The decisions were already made. The two hours are pure transcription and structuring — exactly the kind of work language models are good at, and exactly the kind of work a busy clinician should not be doing at 10pm.
An AI medical scribe sits inside the visit and drafts the note as the conversation happens. The doctor talks to the patient the way they always have; the scribe listens, transcribes, understands, and produces a structured draft — SOAP notes, suggested diagnoses, suggested ICD codes — ready for the clinician to review, correct, and sign. It is the single highest-impact workflow in healthcare AI precisely because it hands time back to the most expensive and most stretched person in the building.
This guide walks through how a scribe actually works, the accuracy problem at its centre, and the compliance and integration decisions that separate a demo from something a clinic will trust with real patients.
Ambient vs Dictation: Two Very Different Products
The first design decision shapes everything else: does the clinician talk to the scribe, or does the scribe listen to the room?
Dictation scribing is the older model. The clinician speaks deliberately to the system after the visit — "Patient presents with a three-day history of..." — and the tool transcribes and formats it. It's accurate because the input is clean and intentional, but it doesn't remove the documentation burden so much as move it. The doctor is still narrating the whole note, just out loud instead of by keyboard.
Ambient scribing is the model that actually changes the day. The system listens to the natural conversation between clinician and patient — the questions, the answers, the small talk, the examination findings said aloud — and reconstructs a structured clinical note from it. The clinician does nothing extra. They have the consultation they'd have anyway, and a draft note is waiting when it ends.
Ambient is far more valuable and far harder to build. Real consultations are messy: people interrupt each other, wander off topic, mention symptoms and then contradict them, and switch between the clinically relevant and the entirely irrelevant. Turning that into a clean SOAP note is a genuine language-understanding problem, not a transcription one — which is why the two-model architecture below exists.
How It Works, Step by Step
Under the hood, a scribe is a pipeline. Each stage has a clear job, and each has a preferred set of tools.
1. Capture and speech-to-text. Audio from the consultation is transcribed into text. This is where accuracy is won or lost, so the transcription model matters. Production systems typically use a strong medical-aware speech engine — Whisper or Deepgram are common choices — tuned for clinical vocabulary, accents, and the crosstalk of a real room. Drug names, dosages, and anatomy are where general transcription models stumble, so this stage often carries a medical vocabulary layer on top.
2. Understanding the conversation. The raw transcript is not a note — it's a wall of dialogue. A general-purpose LLM reads it and works out what actually happened: who said what, which statements are the patient's history, which are the clinician's findings, what was examined, what was decided. This is the comprehension layer that ambient scribing lives or dies on.
3. Structuring into SOAP. A medical-tuned model organises the understood content into the standard clinical format — Subjective, Objective, Assessment, Plan. The subjective section captures the patient's account; the objective captures examination findings and vitals; the assessment holds the clinician's reasoning; the plan holds next steps. Splitting the work between a general LLM for comprehension and a medical-tuned model for clinical structuring tends to produce cleaner, safer output than asking a single model to do both.
4. Suggesting diagnoses and ICD codes. From the assessment, the system proposes likely diagnoses and maps them to ICD-10 codes, and it can suggest the procedure codes that documentation supports. This is where a scribe stops being a note-taker and starts feeding the revenue cycle — more on that below.
5. Review, sign, and write back. The draft is presented to the clinician, who edits and signs it. Only then does it flow into the EHR — Epic, Cerner, or whatever the clinic runs — via a standards-based integration built on FHIR. The clinician's signature is the gate. Nothing reaches the chart unreviewed.
The Accuracy Problem, and Why the Clinician Always Signs
Here is the uncomfortable truth at the centre of every medical scribe: language models can hallucinate. They can invent a symptom that was never mentioned, attach a medication to the wrong patient statement, or confidently produce a plausible-but-wrong dosage. In most software that's an annoyance. In a clinical note it's a patient-safety issue and a legal one.
This is why a responsible scribe is never autonomous. The AI drafts; a licensed clinician reviews, corrects, and signs. That is not a limitation to be engineered away in a future version — it is the design. The value of the scribe is that it removes the first-draft burden, the mechanical work of turning a conversation into structured text. It does not remove, and must not remove, the professional judgement that decides what is true and goes in the record.
Good systems make that review fast and honest rather than pretending it isn't needed. Practical patterns that help:
- Grounding, not guessing. The note should be built from what was actually said. Retrieval and citation techniques — the same medical LLM and RAG discipline used across clinical AI — let the draft point back to the transcript, so a clinician can see why a line is there rather than trusting it blind.
- Flagging uncertainty. Where the model is unsure — an inaudible dosage, an ambiguous symptom — it should surface the uncertainty rather than smoothing over it with a confident guess.
- Easy correction. The review interface should make editing faster than rewriting. If checking the AI's note takes as long as writing one, the product has failed.
The same care that governs any AI agent handling sensitive data and taking actions applies here, with a clinician's licence on the line. Assistive, reviewed, and signed — that framing is not marketing caution. It is what keeps the tool safe, keeps it clear of being regulated as an autonomous diagnostic device, and keeps clinicians willing to use it.
PHI, BAAs, and What Happens When Audio Hits a Model API
A medical scribe handles some of the most sensitive data there is: a recording of a patient describing their health, and a transcript of it. That is protected health information (PHI) from the moment the microphone opens, and HIPAA governs every step it takes.
The decision that matters most is what happens when that audio or text is sent to a model. If transcription or structuring runs against a third-party API, PHI is leaving your environment and entering someone else's — which is permitted, but only under a signed Business Associate Agreement (BAA) with that provider, on an enterprise arrangement that excludes your data from training. Consumer API terms do not carry those commitments. Sending a patient's consultation through a standard, un-agreed API is the kind of mistake that turns a helpful tool into a reportable breach.
The architectural options, roughly in order of control:
- Enterprise APIs under a BAA. The most common path. You use a strong hosted model, but under terms that make it a compliant business associate — data processing agreement in place, no training on your data, audit trail available.
- In-region, de-identified processing. Minimise or strip identifiers before data reaches the model where the workflow allows it, reducing what's exposed even under a BAA.
- Self-hosted or private models. For clinics that cannot let PHI leave their environment at all, running open medical models inside a controlled boundary keeps the data in-house — at higher engineering cost.
Whichever path, the non-negotiables are the same: encryption in transit and at rest, audit logging of every access, strict access control, and a clear, documented data flow that a compliance officer can actually read. HIPAA architecture is not a review you bolt on before launch. It is a set of decisions made before the first line of code.
From Note Quality to Getting Paid: The Billing Connection
A scribe looks like a documentation tool, but its real business case reaches into the revenue cycle. Clinical documentation is the raw material for coding, and coding is what determines whether — and how much — a clinic gets paid.
When notes are thin, rushed, or written at 11pm, the downstream coding is worse: diagnoses that should have been captured get missed, the specificity that supports a higher-value code isn't documented, and claims get denied or underpaid for lack of supporting evidence. A scribe that produces complete, structured notes with correctly suggested ICD-10 and procedure codes doesn't just save time — it improves the accuracy and completeness of the record that billing depends on.
Here's the difference in concrete terms.
Before (manual, end-of-day notes): A clinician sees a patient with poorly controlled type 2 diabetes and an early diabetic complication. Rushing the note that evening, they record "diabetes, follow up" and move on. The coder, working from a vague note, submits a low-specificity code. The complication goes undocumented and unbilled, and if the payer queries it, there's nothing in the record to support anything richer.
After (AI scribe, reviewed and signed): The scribe captures the full conversation — the specific complication mentioned, the medication change, the follow-up plan — and drafts a structured note that surfaces the appropriate specific ICD-10 codes for the clinician to confirm. The signed note supports accurate, defensible coding, the complication is captured, and the claim reflects the care actually delivered.
The revenue effect is real, but it should be framed honestly. Illustrative improvements in documentation completeness and captured revenue vary widely by specialty, payer mix, and how disciplined the existing process already is. There is no guaranteed number, and any vendor quoting one should be treated with suspicion. The reliable claim is directional: better documentation supports better coding, and better coding reduces leakage.
What Good Looks Like
A medical scribe worth trusting has these properties:
Ambient, not effortful. The clinician has a normal consultation. The scribe does the work in the background, not by demanding structured dictation.
Accurate on clinical language. Drug names, dosages, and anatomy are transcribed correctly, because the transcription layer is medical-aware, not a general consumer engine.
Grounded and reviewable. Every line in the draft traces back to something said in the visit. The clinician can see why, and correct fast.
Clinician-signed, always. Nothing enters the chart without a licensed human reviewing and signing. The AI never diagnoses autonomously.
HIPAA-architected from day one. PHI is encrypted, access is logged, third-party model calls run under a BAA, and the data flow is documented.
EHR-integrated properly. The signed note writes back into Epic, Cerner, or the clinic's system through a real FHIR integration — not a copy-paste that clinicians quietly abandon.
Questions to Ask Before You Build or Buy
"Where does the patient audio go, and under what agreement?" You want a specific answer: which model, which provider, and whether a signed BAA and no-training terms are in place. "It's secure" is not an answer.
"How does the clinician review and sign, and how fast is it?" If reviewing the AI's note takes as long as writing one, the time saving evaporates. Ask to see the review flow.
"How does it handle the messy parts — accents, crosstalk, drug names?" These are where scribes fail. A confident answer names the transcription approach and the medical vocabulary handling.
"How does the note get into our EHR?" A real FHIR-based write-back to your specific system, or a fragile export? This is usually the riskiest integration and the one to prove early.
"What happens when it's unsure?" Good systems flag uncertainty for the clinician. Ones that always sound confident are the dangerous ones.
What It Costs and How Long It Takes
Medical scribes are commonly sold as a subscription rather than a one-off build — often in the region of $300–$800 per physician per month, which is why the model scales so well across a group practice or a health system with thousands of clinicians. For a buyer, that price has to be weighed against the hours it returns and the documentation quality it improves; for a builder, it's a recurring-revenue product rather than a closed project.
Building one to production standard is a multi-month engagement, not a weekend prototype, because the hard parts are real. The transcription pipeline, the two-model comprehension-and-structuring layer, the review interface, the HIPAA architecture, and the EHR write-back each take genuine work — and the EHR integration in particular tends to be the long pole. A demo that transcribes a scripted conversation is quick. A scribe a 12-physician primary care group will trust with every patient, every day, is a serious build.
The honest caveat on time savings: the "two hours back per day" figure is illustrative and varies enormously by specialty, patient complexity, and how the clinic works today. A high-volume primary care setting sees a different return from a specialty clinic with long, complex consultations. Treat any single headline number as directional, and measure the real one in your own setting.
Related guides
- Healthcare AI development: the workflows worth automating
- FHIR, HL7 and EHR integration
- Medical LLMs, RAG and human-in-the-loop
- Healthcare voice AI receptionist
- AI agents for healthcare: reducing admin burden
- AI agent security: what to know before you build
- Our AI development services
We Build Scribes Clinicians Actually Sign Off On
A medical scribe only earns its keep if clinicians trust it enough to use it every day — which means it has to be accurate on clinical language, fast to review, honest about uncertainty, and safe with PHI from the first architecture decision. We build documentation workflows that are HIPAA-aware from day one, integrate properly with the EHR your clinicians already use, and keep a licensed human reviewing and signing every note.
Whether you want to automate documentation for a single practice or build toward a scribe product with recurring revenue, we're happy to scope the right starting point for your setting.
Talk to us about your platform — no commitment, just a conversation.
Frequently Asked Questions
What is an AI medical scribe?
It's software that listens to a clinical consultation, transcribes it, understands the conversation, and drafts a structured clinical note — typically in SOAP format — with suggested diagnoses and ICD codes. The clinician then reviews, corrects, and signs the note before it enters the electronic health record. The goal is to remove the two to four hours a day many physicians spend writing notes, by handing them a clinician-ready first draft instead of a blank page.
Is an AI scribe accurate enough to trust?
It's accurate enough to draft, not to decide. Language models can mis-transcribe or hallucinate, which is exactly why a licensed clinician always reviews and signs the note. The scribe removes the mechanical work of turning a conversation into structured text; the clinician retains the judgement about what is clinically true. A well-built system grounds every line in what was actually said, flags where it's uncertain, and makes correction fast — so review is quick without being a rubber stamp.
Does an AI scribe diagnose the patient?
No. It can suggest likely diagnoses and matching ICD codes from the documented conversation, but it does not diagnose autonomously, and it shouldn't be built to. The clinician makes the diagnosis and signs the note. Keeping the AI in an assistive, human-reviewed role is both better clinical practice and what keeps the tool clear of being regulated as an autonomous diagnostic device.
Is sending patient audio to an AI model HIPAA-compliant?
It can be, but only with the right agreements. Patient audio and transcripts are protected health information, so any third-party model that processes them must be under a signed Business Associate Agreement, on enterprise terms that exclude your data from training. Consumer API plans do not carry those commitments. Alternatives include de-identifying data before it reaches the model, or self-hosting a private model so PHI never leaves your environment. The data flow has to be designed with HIPAA in mind from the start — it can't be added afterwards.
How does the note get into our EHR?
Through a standards-based integration, ideally FHIR, that writes the signed note back into your EHR — Epic, Cerner, Athenahealth, or whatever you run. This write-back is usually the riskiest and most involved part of the build, because each EHR has its own APIs, certification, and quirks. A scribe that only produces a note you then copy and paste by hand hasn't finished the job; clinicians quietly abandon tools that add clicks rather than removing them.
How much time does an AI scribe actually save?
Vendors often cite figures like "two hours a day," and while that's a realistic ceiling in high-documentation settings, the honest answer is that it varies. A high-volume primary care clinic sees a very different return from a specialty practice with long, complex consultations. The savings depend on your specialty, patient mix, and how you document today. Treat any single headline number as illustrative and directional, and measure the real one in your own setting rather than trusting a marketing figure.