TLDR: 2025 brought a clear-eyed picture of medical artificial intelligence (AI). A large meta-analysis found that generative AI matches non-specialist doctors on diagnostic accuracy while specialists stay ahead, an industry adoption index showed most health systems experimenting while about a third reach production, and the medtech sector published a policy roadmap for regulators. The capability is real and bounded, and the value now lies in closing the gap from pilot to production with sound governance.
The evidence sets a realistic bar
A research team at Osaka Metropolitan University pooled 83 studies in a meta-analysis published in npj Digital Medicine and found that generative AI reached an average diagnostic accuracy of 52.1%. The latest models performed on par with non-specialist doctors, while medical specialists held a 15.8% edge.
That result reframes the conversation around a useful middle ground. Generative AI works as a capable generalist, which suits triage, medical education, and settings with limited clinician access. Specialists keep their advantage on complex cases, so the practical design pairs AI breadth with expert oversight, keeping a specialist in the loop.
Adoption reveals the real bottleneck
Capability is one thing, and deployment is another. The Healthcare AI Adoption Index from Bessemer Venture Partners, AWS, and Bain surveyed more than 400 healthcare buyers. Around 70% are pursuing generative AI, and about a third of pilots reach production.
The gap between experiment and production is the story. Security, data readiness, integration cost, and in-house expertise hold projects back. The index also found that established vendors drive most adoption, with startups behind a smaller share, which signals that trust and integration carry as much weight as raw model performance.
Policy races to keep pace
Regulation moved in parallel. AdvaMed, the medtech industry association, released an AI Policy Roadmap to guide Congress and federal agencies. It spans five areas: privacy and data access, the United States Food and Drug Administration (FDA) regulatory framework, reimbursement and coverage, AI assurance labs, and generative AI, and it notes that the FDA has authorized more than 1,000 AI-enabled devices over 25 years.
The roadmap matters because reimbursement and clear rules decide whether a validated tool reaches patients at scale. A model can prove itself in a study and still wait years for routine, paid use. Industry now asks regulators to update the framework so that proven AI earns coverage while patient data stays protected.
What it means for builders in regulated industries
Three signals point the same way. The capability is genuine and bounded, the bottleneck sits in production downstream of the model, and policy is the gate to scale. The advantage goes to teams that treat governance as a product feature: clinical validation, human oversight, security, transparent data handling, and a reimbursement strategy built in from the start.
That work, where powerful technology meets the demands of a regulated industry, is the terrain Kainjoo operates in every day. The organisations that pair real evidence with real trust will convert promising pilots into everyday care, and they will own the share of value that the production gap leaves on the table today.
References
- A systematic review and meta-analysis of diagnostic performance between generative AI and physicians (npj Digital Medicine): https://www.nature.com/articles/s41746-025-01543-z
- The Healthcare AI Adoption Index (Bessemer Venture Partners, AWS, Bain): https://www.bvp.com/atlas/the-healthcare-ai-adoption-index
- AdvaMed releases AI Policy Roadmap to guide Congress and federal agencies: https://www.advamed.org/industry-updates/news/advamed-releases-ai-policy-roadmap-to-guide-congress-and-federal-agencies/
