The AI Revolution
in the Operating Room

U.S. national health spending reached $5.3 trillion in 2024 — 18% of GDP, growing 7.2% year-on-year (CMS NHE Fact Sheet). Hospital expenditures alone hit $1.63 trillion, physician and clinical services $1.11 trillion. Every percentage point of efficiency is measured in tens of billions.

Administrative friction is a real and quantified problem. CAQH estimates the U.S. healthcare industry spends roughly $83 billion annually on administrative transactions between providers and health plans, of which $14.6 billion is prior authorization alone (CAQH Index). Providers absorb 97% of that cost burden — even though payers create the requirement.

This is the backdrop for AI’s accelerating role. The global AI in healthcare market is forecast to grow from $21.66 billion in 2025 to $110.61 billion by 2030 — a CAGR of 38.6% (MarketsandMarkets, 2025). The 2024 baseline was $14.92B.

The single largest documented production deployment of generative AI in healthcare is Kaiser Permanente’s rollout of Abridge, an ambient clinical-documentation tool. As of November 2024, the deployment covers 40 hospitals and more than 600 medical offices across eight states and Washington, D.C., supporting 24,600 physicians and 73,600 nurses. Kaiser reported the tool had been used in over 4 million patient encounters (Kaiser Permanente press release · Fierce Healthcare).

Abridge converts the clinician-patient conversation into a structured clinical note in real time, integrating directly into the EHR. It supports 14+ languages and 50+ specialties. Kaiser’s executive leadership has publicly described it as the largest implementation of ambient listening technology in healthcare to date.

The broader picture: physician adoption of AI tools has roughly doubled in a single year. The American Medical Association’s annual Physician Survey on Augmented Intelligence found physician use rising from 38% in 2023 to 66% in 2024 (AMA, 2025). The most common cited use case is reducing administrative burden, named by 57% of respondents.

The clearest production milestone in AI-driven drug discovery to date belongs to Insilico Medicine. Its lead candidate Rentosertib (ISM001-055) — for idiopathic pulmonary fibrosis — is the first drug where both the biological target and the molecule were discovered using generative AI (EurekAlert).

The numbers: Insilico took the candidate from target identification to a preclinical candidate in under 18 months, at a preclinical cost of approximately $2.6 million — compared to industry norms of 10–15 years and multi-hundred-million-dollar development costs (Insilico via NVIDIA · Insilico). The drug is now in Phase 2 trials in the U.S. and China.

A single Phase-2 readout is not yet proof that AI-discovered drugs work in patients. But it is proof that the discovery half of the pipeline — historically the bottleneck — can be compressed by an order of magnitude. The Phase-2 efficacy and Phase-3 outcomes over the next 24–36 months will determine whether the model translates.

The AMA’s 2024 data is telling: while 66% of physicians now use AI, 68% see “some or definite advantage” — meaningfully more than the 65% who said so in 2023, but still far from the 90%+ adoption-with-conviction that other digital tools have reached. The honest read: doctors are using it because they have to, not yet because they trust it.

Regulatory pressure is rising in parallel. The EU AI Act entered into force on 1 August 2024; the obligations on high-risk AI systems — which includes much of clinical-decision-support AI — apply from 2 August 2026 (for AI embedded in regulated products) and 2 August 2027 for the broader Annex III list (EU AI Act implementation timeline). Healthcare AI sold or used in the EU is squarely in scope.

Strip away the projections and the headline numbers, and the practical picture is simple. The healthcare AI investments paying back today are the ones targeting the boring stuff: documentation, billing, prior authorization, scheduling, and resource forecasting. They work because the failure mode is recoverable — a wrong note can be corrected — and because the value capture is direct.

The hard, high-stakes use cases — autonomous diagnosis, treatment selection, drug discovery to approval — are happening, but on longer timeframes and with much closer regulatory scrutiny. Insilico’s Rentosertib may be the first in its class, but it is one drug in Phase 2. The pattern repeats: real AI in healthcare is incremental, instrumented, and humble.