The real cost of a free AI scribe
Key takeaways
- Transcribing consultations and drafting notes creates a recurring AI processing cost for Hanah.
- Hanah’s approach is to offer a trial of the paid scribe using the same model as the paid plan.
- Two or three extra corrections can add 34–51 seconds per note in the worked example below.
- In our matched benchmark subset, Heidi needed an estimated 71 seconds more review and correction per note than Hanah. This compares tested configurations, not verified free tiers.
- Draft error counts measure correction burden; they do not establish the risk of an error surviving clinical review.
We don't want to sacrifice output quality by using cheaper LLMs, that's why Hanah offers a generous 30 day trial and no cheap model option.
Every note has a cost
Every note Hanah drafts runs through two AI models. One transcribes the consult as it happens, and a large language model then drafts the note from that transcript and your template. We pay for both on every consult, and the cost grows with every patient you see.
We pick our AI vendors by rigorously testing them so we can publish honest benchmarks, stand behind them and know we add value and minimise harm.
Unfortunately this comes with additional compute costs to us. Most software costs almost nothing extra per user, which is why a calendar or a form builder can give a tier away for free. An AI scribe can't do that. A clinician seeing a full caseload runs every one of those consults through both models, and every one of them has a bill attached.
The subscription is only part of the cost
We don't want a permanent free AI tier to dictate which models we can afford to use. We choose models by testing the documentation they produce, and charge for the ongoing processing. A free service can be subsidised in different ways; its price alone does not tell you which model it uses or how accurate it is.
What matters to your practice is the total cost: the subscription plus the time you spend reading and correcting drafts. A missing detail needs adding back. An invented statement needs checking and removing. A wrong detail may need both. Even a plausible error that takes seconds to fix can be difficult to spot.
What two or three extra fixes cost
Our benchmark review-time model estimates the work as:
Review seconds = reading seconds + recall and typing seconds for missing or partial facts + checking and removal seconds for unsupported claims.
The assumptions are reading at 190 words a minute, five seconds to recall each missing or partial fact, typing at 40 words a minute, and 12 seconds to check and remove each unsupported claim. A partial fact needs half the fact retyped. A wrong fact can need removal as well as replacement. These are modelling assumptions, not stopwatch measurements of clinicians.
For a worked example, take a missing fact that needs eight words to restore. Typing takes 8 ÷ 40 × 60 = 12 seconds. Add five seconds to recall it and that is 17 seconds per fix. Holding reading time equal:
| Extra missing facts per note | Extra correction time per note | Extra time at 20 notes a day |
|---|---|---|
| 2 | 34 seconds | 11 minutes 20 seconds |
| 3 | 51 seconds | 17 minutes |
At 400 notes a month, those differences become 3 hours 47 minutes to 5 hours 40 minutes. If you value that time at an illustrative A$100 an hour, it is A$378–A$567 a month in clinician time. This is a worked scenario, not a measured result for a particular free scribe or a promise of extra revenue.
The calculation for your practice is: extra seconds per note × notes per month ÷ 3,600 × value of an hour. For comparison, Hanah's monthly Practice subscription is A$100 per practitioner. At the illustrative volume and hourly value above, a difference of just nine seconds per note equals that subscription cost.
What our matched benchmark actually found
We used the matched observations already published with our accuracy article, sourced on 18 October 2026. There are eight consultations, each with five complete runs for each of six products: 40 drafts per product. The table averages each product's 40 observations, so every product is compared on the same consultations.
| Tested scribe | Scored errors per note | Estimated reading and correction per note | Extra time versus Hanah |
|---|---|---|---|
| Hanah | 1.28 | 84 seconds | — |
| Heidi | 7.93 | 155 seconds | 71 seconds |
| Lyrebird | 10.00 | 192 seconds | 108 seconds |
| Freed | 7.18 | 222 seconds | 138 seconds |
| CliniScripts | 8.75 | 206 seconds | 122 seconds |
| PatientNotes | 15.93 | 270 seconds | 186 seconds |
For example, Heidi averaged 6.65 more scored errors per note and 71 seconds more estimated review and correction time than Hanah. At 20 notes a day, that estimated difference is about 24 minutes a day. Time differences use the unrounded means before rounding to seconds. Errors include missing facts, partial facts and unsupported claims; an error count is not a count of keystrokes or necessarily a separate editing action.
These are vendor-run, AI-graded benchmark findings, without independent clinician validation. Review time is modelled and excludes generation wait and formatting. The captures do not establish a verified free-tier comparison, so these figures describe the tested configurations, not every free AI scribe or a free-versus-paid model experiment. See the source export and disclosures, matched-subset method and reproducible calculations for this article.
More draft errors means more to catch
In that matched sample, 28 of 40 Hanah drafts (70%) needed at least one correction, compared with 40 of 40 Heidi drafts (100%). That is 30 percentage points higher, or 43% higher relative to Hanah, for a draft needing correction in this sample. It does not mean an error was 43% more likely to reach the patient's record.
To quantify errors surviving review, we would need to measure which errors clinicians missed. This benchmark does not measure that. We can quantify how much there is to catch and the estimated work to correct it; we cannot turn those counts into a clinical risk percentage. Every draft still needs review, including Hanah's.
A full month on the same model
Instead of a free tier, the Practice plan comes with a one-month free trial and no card needed. The trial runs on the same model paying clinicians use, on your consults. A month is long enough to see how much time Hanah gives back to you before you decide whether it's worth paying for.
If you have any questions please don't hesitate to reach out to us at hello@hanah.health.
Frequently asked questions
Why does Hanah charge for the AI scribe?
Hanah pays for transcription and note generation as consultations are processed. Its paid scribe plan funds those ongoing costs.
Does the trial use the same model as the paid plan?
The approach described in this article uses the same model for trial and paid users. Check current pricing for the available plan and trial terms.
What should I look for during the trial?
Use your consultation types and templates. Check captured details, unsupported statements, review effort and how the draft reaches your clinical record.
Sources cited in this article
- publish honest benchmarks — hanah.health
- benchmark review-time model — scribestandard.com
- source export and disclosures — scribestandard.com
Update note: Removed the caveat that described the Hanah notes as an 11 October rerun. Benchmark figures and source dates are unchanged.