What 50 clinics across 6 countries told us about AI medical scribe ROI

Every AI-scribe vendor publishes ROI numbers. Most of them come from a single hand-picked pilot — usually a 5-clinician practice in a single country, often run by enthusiasts. The numbers are right, but they're not transportable. Your clinic in Manchester or Toronto or Auckland isn't that pilot.

We did the harder thing. With customer permission and IRB-equivalent approval, we ran a 90-day time-and-motion plus billing-impact study across 50 MedMETs partner clinics. The sample spans Australia (12), the US (11), the UK (10), New Zealand (6), Canada (7), and Singapore (4). Mix of solo, group, and clinic-network practices. Specialties: 78% primary care, 14% allied specialty, 8% mental health.

The headline number nobody contests

Median documentation time per consult dropped from 4.8 minutes to 1.6 minutes. That's 3.2 minutes per encounter — or, at 22 consults per day per clinician, about 70 minutes back per shift. The number is remarkably consistent across countries. Auckland clinicians saved the same time as Atlanta clinicians, regardless of EHR.

Where the maths diverges: revenue impact

Time saved is universal. What you do with it isn't. Three patterns emerge:

• Capitated systems (UK NHS, NZ PHO) — clinicians reinvested time in QOF / Care Plus indicator capture. UK practices saw an average £14,200 annual lift in QOF revenue per GP partner. NZ practices flagged 31% more Care Plus-eligible patients.
• Fee-per-service (US, parts of AU, parts of CA) — clinicians ran an additional 1.4–2.1 consults per day on average. US Medicare-heavy practices specifically increased Annual Wellness Visit (G0438/G0439) throughput by 38%.
• Mixed systems (Singapore Healthier SG, AU MBS, CA provincial) — the revenue lift came primarily from previously-undercoded items: care plan reviews, complex-care premiums, preventive-care line items.

Documentation quality — the gain everyone underestimates

We tracked completeness against each clinic's own note template. Median completeness rose from 71% to 88% — a 17-percentage-point gain. The gain was largest in subjective-history capture, where time pressure historically forced clinicians to skip sections.

Audit-defensibility followed. In US clinics specifically, retrospective E/M 2021 MDM levelling on AI-scribed notes resulted in 4% upcoding (justified by the better-documented decision-making) and 11% downcoding (where the documentation didn't support the previously-billed level — meaning those notes were now safer in an audit, not riskier).

Patient experience — the gain that surprised the cynics

“The doctor actually looked at me. They weren't on the screen. I noticed straight away.”

Patient survey, Wellington practice

Patient-reported "I felt heard" scores rose from a baseline median of 7.3 to 8.5 (on a 10-point scale). Net Promoter Scores across the cohort rose 22 points. The mechanism is obvious in hindsight: when the clinician's eyes are on the patient instead of the keyboard, the room reads differently.

Where ROI was negative or flat

Three of the 50 clinics saw no ROI. All three shared the same pattern: the senior partner refused to use the tool, the junior staff felt cultural pressure to also not use it, and the platform sat unused. Adoption is the bottleneck, not capability. Vendor support during week 1–4 was the single biggest predictor of which clinics hit the median saving and which didn't.

What to do with this data

If your country isn't in the high-revenue-lift cohort, it doesn't mean ROI is absent — it means it's distributed differently. UK clinics get it in QOF capture. NZ clinics get it in capitation-eligible patient identification. US clinics get it in additional billable encounters and audit safety. AU clinics get it in MBS items they were previously leaving unclaimed. Singapore clinics get it in CHAS-stratified workflow.

The base case is the same: 3.2 minutes per consult. Everything else follows from what you do with that time.