AI scribes for clinical notes: the real math of time, cost and risk
What documentation actually costs a cash-pay practice, how AI note drafting works, and why paying per visit through your own OpenAI key beats a flat $99 add-on.
August 07, 2026 · 7 min read
Ask any clinician what they would change about their work and the answer is rarely the medicine. It is the typing. Study after study puts clinical documentation at 1 to 2 hours per day, and a good part of it happens after dinner. The industry even has a name for it: pajama time.
This post does three things. First, it puts a dollar figure on that time, because "documentation is a burden" is an emotion and emotions do not survive a budget meeting. Second, it explains how an AI scribe actually works, in plain terms, including the part where a human still signs every note. Third, it walks through the pricing model most vendors use, a flat $35 to $99 per provider per month, and compares it with the alternative: running the AI on your own OpenAI API key and paying cents per visit.
What documentation actually costs you
Take a clinician in a cash-pay practice who charges $150 per session. A 45 to 50 minute visit at that rate values an hour of clinical time somewhere around $180 to $200. Now the documentation math:
| Item | Conservative figure |
|---|---|
| Documentation time per day | 1 to 2 hours (use 1.5) |
| Working days per month | 22 |
| Hours per month spent on notes | ~33 |
| Value of that time at even $100/hour | ~$3,300 per month |
| Value at $180/hour | ~$5,900 per month |
You will not recover all of it. Some documentation is unavoidable and some of the recovered time turns into rest, not revenue. But even if an AI scribe gives you back half, that is 16 hours a month. Four more patients a week, or four evenings that belong to your family again. An empty slot in a cash-pay practice costs $150 to $250, so the same math that makes reducing no-shows worthwhile makes documentation time worth attacking.
There is also the cost that never shows up in a spreadsheet: a note written at 10 pm from memory is a worse note than one drafted from a recording of the actual conversation. Your malpractice carrier cares about that more than you might think.
How the recording and drafting actually work
The mechanics are simpler than the marketing suggests.
- You ask the patient for consent and start recording. This step is not optional. Several states, including California, Florida and Washington, require all parties to consent to a recording. Others require only one party. The safe rule for a clinic: always ask, always note the consent in the chart, and check your own state's rule with your board or attorney, because this varies by state.
- The audio is transcribed. A speech-to-text model turns the conversation into raw text, including the parts where the patient wanders into their neighbor's dog and back.
- A language model drafts the note. It receives the transcript plus a template for your specialty (SOAP, intake, follow-up, procedure note) and produces a structured draft: subjective, objective, assessment, plan, in the format you chart with.
- You review, correct and sign. The draft is a draft. The clinician who signs the note owns the note. More on this below, because this is where vendors get quiet and where you should get loud.
Total added work per visit for the clinician: pressing record, and 1 to 3 minutes of review instead of 8 to 12 minutes of writing. That is the entire trick. There is no magic, just a very good typist who was listening the whole time.
The pricing detail nobody explains: whose API key is it?
Here is the part of the market worth understanding before you buy anything.
Most practice software sells AI scribing as an add-on, typically $35 to $99 per provider per month, flat, and standalone scribe tools commonly land at the top of that range or above. The vendor pays the AI provider wholesale, marks it up, and charges you retail whether you saw 400 patients that month or you were on vacation.
The alternative model, the one we chose to build, is that the software connects to your own OpenAI account with your own API key. You pay OpenAI directly for exactly the audio you transcribe and the notes you draft. No markup, no flat fee, no paying for August while you are at the beach.
What does that cost per visit? At the API prices published as of this writing, and being deliberately conservative:
| Piece | 15-minute visit | 30-minute visit |
|---|---|---|
| Transcription (speech to text) | ~$0.05 to $0.09 | ~$0.09 to $0.18 |
| Note drafting (language model) | ~$0.01 to $0.04 | ~$0.02 to $0.05 |
| Total per visit | ~$0.06 to $0.13 | ~$0.11 to $0.23 |
Now scale it:
- A solo therapist seeing 25 patients a week, about 100 visits a month of 45 to 50 minutes, lands somewhere around $15 to $25 a month. Against a $99 add-on, that is roughly a 75 to 85 percent saving.
- A busy clinic doing 440 short visits a month pays roughly $26 to $57 a month, total, not per provider.
- A three-provider group on a $99-per-provider add-on pays $297 a month, about $3,560 a year. The same group on its own API key will usually stay under $100 a month.
Prices from AI providers have been falling, not rising, and with your own key every price cut lands in your pocket instead of the vendor's margin. The honest caveat: you have to create an OpenAI account and generate a key, which takes about ten minutes and feels technical the first time. After that, it is a line item on a card statement like any other.
One more honest caveat: if you want a scribe with a dedicated medical-only model, human quality review of every note, or specialty fine-tuning under contract, a premium standalone scribe may serve you better and the $99 or more can be worth it. High-volume, high-complexity documentation is their whole business. For most cash-pay practices, it is overkill.
Limits, and why the human review is not a formality
An AI scribe is an excellent first draft and a terrible final authority. Three limits you should hold any vendor to admitting:
It can make things up. Language models occasionally state something the patient never said, usually plausible, which is what makes it dangerous. Review is not skimming. Read the assessment and plan word by word before signing. The clinician who signs is legally responsible for the content, full stop.
It does not examine anyone. The objective section of your note comes from what you said out loud during the visit. Clinicians who use scribes well learn to narrate: "blood pressure 128 over 82, lungs clear." If you do not say it, it is not in the transcript, and the model should not invent it.
It is documentation support, not a diagnostic device. Some platforms, ours included, also offer an AI clinical assistant for things like differential lists or medication interaction checks. Useful as a second set of eyes, never as the decider. If a tool claims to diagnose, ask who carries the liability. The silence will be informative.
On privacy: the audio and transcript are protected health information. Whoever processes them must be covered by your HIPAA business associate agreement, and API traffic should be encrypted in transit and at rest. Ask the vendor to put in writing whether recordings are retained, for how long, and whether anything is used to train models. "We don't know" is an answer, and it is a disqualifying one.
What to look for when you evaluate one
- Template control. The draft should follow your note structure per specialty, not a generic SOAP blob you spend ten minutes reformatting. If your charting is built on custom forms, the scribe should write into that structure, which is exactly how we approached it in our own charting.
- Where the draft lands. Inside the chart, editable, attached to the visit. If the scribe lives in a separate app and you copy-paste, you have added a step, not removed one.
- Cost model in writing. Flat fee per provider, or metered on your own key. Do the math above with your real visit volume before signing anything.
- A consent workflow. A checkbox or script for recording consent, stored with the chart.
- An exit. Your notes are yours. Confirm you can export everything the day you leave.
The scribe is also only half of the AI story in a small practice. The other half sits at the front desk, where missed calls and repetitive questions burn money in a different way. We ran those numbers separately in what AI can actually save at the clinic front desk.
The bottom line
Documentation costs a typical clinician $3,000 to $6,000 a month in time. An AI scribe recovers a real share of that for either a flat $35 to $99 per provider per month, or for cents per visit if the software runs on your own API key. The second model is cheaper for almost every cash-pay practice, and it is transparent: you can read your own OpenAI invoice and see exactly what a month of notes cost.
Whatever tool you pick, keep the two rules that do not change: the patient consents to the recording, and the clinician reads what they sign.
DrinCloud includes an AI scribe that records the visit and drafts the note inside your own chart, running on your own OpenAI key at cents per visit. See it on the AI scribe page or start a 15-day free trial, sample data loaded, no credit card.
See it in your own practice
Fifteen days free, sample data already loaded, no credit card.