"How much has AI actually changed healthcare?" gets answered with whichever number is closest to hand — and the numbers rarely agree, because they rarely measure the same thing. This page is the hub for our statistics cluster: the dozen-or-so figures that carry the most weight, each pinned to a primary source, dated, and labelled with what it really counts. Four questions organize it — who is using AI, what the evidence shows, what the rules now require, and where the money went. As of July 2026.
How to read a healthcare-AI statistic
One habit prevents most confusion: ask what the denominator is. An adoption figure can count clinicians offered a tool, clinicians who used it once, or clinicians who use it for most of their visits — three very different populations behind the same percent sign. A device count depends on whether "AI" means any authorization mentioning the term or a narrower technical definition. Every figure below is written to make its denominator and its date visible, because a stat without those two things is closer to a slogan than a measurement.
Adoption — who is using AI?
The clearest adoption signal comes from the American Medical Association's Augmented Intelligence survey of roughly 1,200 physicians. The share reporting that they use AI in practice rose from 38% in 2023 to 66% in 2024 — a jump of about 78% in a single year 1. Sentiment moved with it: 68% saw a definite or some advantage to AI tools (up from 65%), and the share whose enthusiasm outweighed their concern rose to 35% from 30% 1.
That enthusiasm is conditional. The same physicians named the features they need before they lean harder on these tools — a feedback channel (88%), data-privacy assurances (87%), and EHR integration (84%) — and ranked increased oversight as the top regulatory step that would raise their confidence 1. The appetite is real; the trust is contingent.
How far has AI reached the clinic?
Two very different measures show how far AI has actually reached the clinic.
On the device side, a peer-reviewed taxonomy examined 1,016 FDA authorizations of AI-enabled devices and found that quantitative image analysis remains the most common function, though its relative share is now declining; more than 100 devices use AI to generate data, and none yet rely on large language models 2. The distribution is lopsided by specialty: of the 168 machine-learning-enabled devices the FDA authorized in 2024, 74.4% (125) were radiology, trailed distantly by cardiovascular (6.5%) and neurology (6.0%) 3. The FDA's own device list is the living record here 13; our FDA device tracker breaks the trend down by year and panel.
On the software-in-workflow side, the largest documented rollout is the ambient AI scribe at The Permanente Medical Group. In its first ten weeks, 3,442 physicians used it across 303,266 encounters 5; one year in, the group reported the tool had passed over 2.5 million uses 4. That is the ceiling of what a well-resourced deployment looks like — most systems report pilots. Our AI scribe adoption tracker keeps the running tally.
Evidence — what do the studies show?
Adoption without outcomes is just enthusiasm, so the controlled evidence matters more than the headline counts.
| Intervention | Measured effect | Design | Source |
|---|---|---|---|
| Ambient AI scribe | ~13 fewer EHR min/day; ~16 fewer documentation min/day; only 32% frequent users | ~1,800 clinicians | 6 |
| Ambient AI scribe | Clinician burnout 51.9% → 38.8% after 30 days | Six health systems | 7 |
| AI-assisted colonoscopy | Adenoma detection 33% → 41.4% (RR 1.26) | Meta-analysis, 12 RCTs, 11,340 patients | 8 |
| LLM on MedQA (USMLE-style) | Up to 86.1% accuracy | Benchmark evaluation | 9 |
Two patterns run through the table. First, the documentation-relief story is consistent but modest — real minutes back, not transformed days, and uneven because only about a third of clinicians use the tool for most visits 6. Second, the strongest clinical-outcome signal is where AI acts as a second set of eyes on an image or a video: pooling 12 randomized trials, computer-aided colonoscopy raised the adenoma detection rate from 33% to 41.4% 8. Benchmark scores like the 86.1% on a USMLE-style question set 9 show what a model can do on an exam — not what it does in a clinic, which is a different and harder measurement. Our clinical trial results tracker and LLM benchmark tracker hold the deeper reads.
Regulation — what do the rules now require?
The rulebook filled in quickly across 2024. The EU AI Act (Regulation 2024/1689) entered into force on 1 August 2024; because AI that is a regulated medical device is treated as high-risk, its obligations for those products apply from 2 August 2027, giving manufacturers a defined runway 10. In parallel, the World Health Organization issued dedicated guidance on large multi-modal models in health in January 2024, carrying more than 40 recommendations for governments, developers, and providers 11. Alongside both, the FDA's device list and its evolving change-control expectations remain the operational reference for anything marketed in the US 13. The global regulation tracker follows each milestone.
Investment — where did the money go?
Capital has concentrated hard around the label. US digital-health startups raised $14.2B in 2025 — a 35% rise over 2024 — and companies marketing AI took 54% of those dollars, up from 37% the year before, with larger average rounds than their non-AI peers 12. The concentration is the story as much as the total: a handful of very large raises pulled the annual figure up. Our funding and M&A tracker keeps the series current.
How to read these numbers
Four cautions travel with everything above. Adoption percentages hide their denominators — being offered a tool, trying it once, and depending on it daily are not the same fact. Device counts depend on inclusion criteria and the date the list was pulled, so two credible sources can report different totals for the same agency. Controlled effects are consistently smaller than the ones users describe from memory. And a benchmark score is a laboratory result rather than a clinical one. Treat each figure as a reading with an error bar, never a verdict.
Sources and method
Every figure here is dated and tied to a numbered primary source below — regulator pages, peer-reviewed studies, an official physician survey, and a named funding tracker. Where a headline circulates without a first-party source, we leave it out. We refresh this hub on a 90-day cycle and whenever the FDA list, the AMA survey, a major funding report, or an EU AI Act milestone changes. If you found this through a single stat, the sibling trackers linked throughout carry the depth behind it.