Healthcare adopts AI faster

For years, healthcare was supposed to be the industry that moved last. Regulations were thick. Data systems were messy. Software rollouts dragged on for years. Then AI arrived, and the pattern broke. Menlo Ventures found that healthcare organizations are now deploying AI at 2.2 times the rate of the broader economy. Twenty-two percent have already implemented domain-specific AI tools. In 2023, that figure was barely visible. In 2025, healthcare AI spending reached $1.4 billion, nearly triple the year before. (menlovc.com) That speed matters because it tells you this is no longer a story about demos. The biggest traction is showing up where the economics are brutal and the workflows are repetitive enough for software to matter. NVIDIA’s 2026 healthcare survey found that medical imaging is one of the most established use cases, while drug discovery is a leading use case in pharma and biotech. Executives reported not just experimentation but return on investment, with most saying AI is helping raise revenue or cut costs. (blogs.nvidia.com) Imaging is where the change becomes easiest to see. In March, a large multicenter study in *Nature Cancer* tested Google’s mammography AI on more than 115,000 NHS screening exams and then prospectively deployed it at 12 sites. Used as a simulated second reader, the system increased cancer detection, caught a quarter of interval cancers, and cut reading time by 32%. It did especially well on first-time screens, with far fewer recalls and higher detection. That is why breast screening has become the sharp edge of the argument over medical AI. (nature.com) The argument is not really about whether the software can do something useful. It is about where to place it in the workflow. The same *Nature Cancer* paper says prospective deployment exposed a distribution shift that required recalibration. In other words, a model that looks excellent in retrospective testing can still need tuning once it meets the mess of actual clinical practice. That is why the practical position emerging from the field is augmentation, not replacement. The software reads. The radiologist still governs. (nature.com) Once you move outside the reading room, the center of gravity shifts from diagnosis to paperwork. Healthcare does not only run on scans and lab values. It runs on phones, scheduling, eligibility checks, follow-ups, coding, and forms. Menlo’s report points to administrative overhead and clinician burnout as core reasons providers are moving fast. NVIDIA’s survey says the most visible impact over the next year is likely to come from logistics and administrative streamlining. That prediction already looks less like forecasting than description. (menlovc.com) A new crop of startups is building exactly for that layer. Companies such as EliseAI pitch voice systems that answer calls, schedule visits, and handle follow-ups around the clock. Others are selling AI receptionists that verify insurance, route messages, and write directly into practice workflows. These are not glamorous tasks, but they are where clinics lose time, money, and patients. If AI can reliably pick up the phone at 5:17 p.m. and book the appointment that a human front desk misses, adoption does not need a philosophical defense. It has a cash-flow one. (eliseai.com) That also explains why healthcare-specific systems are gaining ground over generic chatbots. Electronic health records are not just another text corpus. They are structured, irregular, incomplete, and full of institution-specific logic. Recent benchmark work from Stanford and collaborators built a virtual EHR environment precisely because ordinary chatbot-style testing misses the hard part: acting inside clinical software. The point is not that general models are useless. It is that the winning products are increasingly the ones wrapped in healthcare-specific guardrails, interfaces, and compliance controls. (hai.stanford.edu) Even the biggest model vendors are moving in that direction. OpenAI launched a healthcare offering in January aimed at HIPAA-aligned deployment, evidence retrieval, and clinical and administrative workflows, and said major health systems were already rolling it out. The market is converging on the same lesson from both ends. Raw model intelligence matters. But in healthcare, the decisive advantage comes from fitting that intelligence into the rules, records, and routines of the place. The software has to survive contact with the EHR. (openai.com)