The real cost of a free AI scribe — calculation record
Analysis date: 24 October 2026 (Australia/Melbourne)

Input: accuracy-over-inference-matched-runs.csv in this directory, sourced on
18 October 2026. See accuracy-over-inference-method.txt for selection rules.
Eight matched consultations, five runs each, 40 observations per product.
This is a secondary analysis, not a new capture or a verified free-tier test.
Review time is modelled, not observed clinician time. Captures and grading
limitations: https://scribestandard.com/methodology/
Current export checked 24 October 2026:
https://scribestandard.com/assets/scribestandard-2026-09-22.json

Reproduce table using Python standard library, from this directory:
import csv, collections
from decimal import Decimal, ROUND_HALF_UP
groups = collections.defaultdict(list)
with open('accuracy-over-inference-matched-runs.csv') as source:
    for row in csv.DictReader(source):
        groups[row['scribe']].append(row)
def mean(rows, key):
    return sum(Decimal(r[key]) for r in rows) / len(rows)
baseline = mean(groups['Hanah'], 'review_time_s')
for name in ['Hanah', 'Heidi', 'Lyrebird', 'Freed', 'CliniScripts', 'PatientNotes']:
    rows = groups[name]
    assert len(rows) == 40
    errors = mean(rows, 'note_errors_weiner')
    review = mean(rows, 'review_time_s')
    needs_fix = sum(r['ready_to_finalise'].lower() != 'true' for r in rows)
    print(name, errors.quantize(Decimal('.01'), rounding=ROUND_HALF_UP),
          review.quantize(Decimal('1'), rounding=ROUND_HALF_UP),
          (review-baseline).quantize(Decimal('1'), rounding=ROUND_HALF_UP),
          needs_fix, '/', len(rows))

Heidi minus Hanah errors: 7.925 - 1.275 = 6.65.
Heidi minus Hanah modelled review seconds: 154.72725 - 83.934 = 70.79325.
At 20 notes/day: 70.79325 * 20 / 60 = 23.59775 minutes/day.
Drafts needing correction: Hanah 28/40, Heidi 40/40.
Absolute difference: 100% - 70% = 30 percentage points.
Relative increase: (100% / 70% - 1) * 100 = 42.857...%.
This is descriptive draft-error incidence, not post-review clinical risk.

Worked scenario, not a measured free-scribe result:
One eight-word missing fact: 5 seconds recall + 8/40*60 typing = 17 seconds.
Two extra missing facts: 34 seconds; three: 51 seconds, reading held equal.
20 notes/day: 680 to 1020 seconds = 11m20s to 17m.
400 notes/month: 13600 to 20400 seconds = 3h46m40s to 5h40m.
At illustrative A$100/hour: A$377.78 to A$566.67/month, rounded A$378–A$567.
A$100 monthly subscription / A$100 hourly value * 3600 / 400 notes = 9s/note.