随着生成式AI日益普及，美国医院竞相适应
U.S. Hospitals Race to Adapt as Generative AI Takes Hold
照片由Miguel Ausejo拍摄于Unsplash美国医生对人工智能的使用率在三年内翻了一倍多。American Medical Association于2026年1月和2月对1692名医生进行的医师增强智能调查发现，81%的医生现在在专业工作中使用AI，而2023年AMA首次就该技术向医生进行调查时，这一比例为38%。受访者平均报告了2.3个不同的使用场景。
机构采用也呈现出同样的曲线。研究公司Eliciting Insights对120家美国医疗系统高管的调查发现，75%的高管已经实施或计划实施至少一项AI解决方案，其中一半表示其所在机构已经运行了三个或更多应用。临床笔记记录以68%的采用率领先，基于AI的临床文档改进紧随其后，达到了令人印象深刻的43%。
医疗系统领导者面临的问题不再是是否部署生成式AI，而是何时部署。所有利益相关者都需要确保具备相应的治理、验证和劳动力能力，以便在不引入新型风险的情况下大规模运行AI。
AI已取得实际进展的领域
环境文档记录是解锁的瓶颈。The Permanente Medical Group报告称，其环境AI记录员为医生节省了约15,791小时的文档记录时间，84%的医生表示对沟通有积极影响，82%的医生表示工作满意度提高。
然而，对照研究则描绘了更谨慎的结果。一项针对2023年至2025年间五个学术医疗中心1800名临床医生的研究发现，AI记录员用户每八小时患者护理可节省约16分钟的文档记录时间，并减少13分钟的病历操作时间，采用者大约每两周多看一名患者。每次就诊的微薄节省在庞大的门诊覆盖范围内仍会产生复利效应，这正是医疗系统持续购买的原因。
患者沟通是第二个战线，而数量问题非常显著。一项大型横断面分析发现，2020年至2025年间，患者撰写的门户消息增加了153%，发送者的消息强度从每年2.2条攀升至5.4条。嵌入EHR的起草助手现在生成初步回复，供临床医生审查。
收入周期已经成为最清晰的商业案例。HFMA对519名首席财务官和收入周期领导者进行的调查发现，2025年有80%的医疗系统在探索、试点或实施用于RCM的生成式AI，不到两年间上升了38个百分点。来自第二项HFMA合作调查的最新数据显示，37%的医疗系统在收入周期内使用生成式AI，其中大型系统的使用率升至48%，45%的医疗系统将AI应用于与拒付相关的工作流程。
试点与生产之间的差距
部署广度仍然有限。Deloitte的《2026年全球医疗展望》报告显示，约30%的受访医疗系统在特定领域大规模运行生成式AI，而只有2%在整个企业范围内部署了生成式AI。高管预计，未来一年生成式和代理式AI将消耗技术预算的19%。
四个约束解释了这一差距。
第一个约束是数据隐私。受保护的健康信息在模型提供者、网关和检索管道中传输，扩大了安全团队的攻击面。医生们也注意到了这一点。86%的AMA受访者认为数据隐私保障对更广泛的采用至关重要，88%的受访者指出需要强有力的安全验证。
还有监管和认证压力。2026年6月，Joint Commission推出了“医疗保健中AI负责任使用”认证，这是首个专门为美国医疗机构打造的此类项目。该标准围绕五个领域组织：治理、有效数据管理、风险与偏见减少、安全性能的监测与验证，以及透明度、教育与培训。该认证遵循2025年发布的初步指南，并与Coalition for Health AI发布的治理手册保持一致。目前认证是自愿的，这对AI采用有明显影响。
在这一关键领域，准确性担忧也在减缓采用步伐。在六个商业模型上的测试发现，AI生成的草稿经常引入错误和无关细节，并且常常未能提出相关的后续问题。“我们发现AI可以听起来像医生，但不会像医生一样思考，”该研究的共同通讯作者、达特茅斯学院计算机科学助理教授Sarah Preum在研究公告中表示。合著者、家庭医学医生Tim Burdick则直言不讳地说明了运营成本：大量编辑可能比从头开始写更耗时。
最后，约85%的AMA受访者表示，希望就AI采用决策进行咨询或直接参与，88%的受访者担心技能退化，尤其是临床经验不足10年的医生。监督只有在执行监督的人保留发现模型错误所需的专业知识时才有效。这是一场席卷包括医疗保健在内的所有主要职业的关键辩论。
培训是真正的瓶颈。
今年医疗保健领域最引人注目的AI数据不是关于模型，而是关于人。
Incredible Health的2026年护理状况报告基于对2240名美国护士的调查发现，护士中AI的采用率在一年内几乎增至三倍，而近一半使用AI的护士报告节省的时间很少或没有。差异在于准备。在接受雇主精心安排的AI培训的护士中，24%每天节省超过一小时，而未接受培训的护士中这一比例为16%。只有8%的护士表示雇主提供了明确的AI战略。
这是一个穿着技术外衣的典型培训问题。医疗系统购买工具的速度快于培养使用工具的能力，回报显示了这一点。
认证机构也得出了同样的结论。教育与培训是Joint Commission认证标准的五大支柱之一，该标准要求机构证明其为员工提供了关于使用的健康AI工具的教育与培训。医院将需要能够评估模型输出、识别其可能失败之处、了解患者数据后续去向，并在发现异常时适当上报的临床医生。
随着生成式AI继续重塑美国的医疗交付格局，像Keuka College这样的机构正在发挥作用，帮助医疗专业人员掌握驾驭这一快速演变格局所需的数字素养。需求信号明确：AI熟练度正从专业信息学技能转变为护士、联合健康人员及临床管理者等群体的基本期望。
接下来会发生什么
这些数字勾勒出问题，并未给出预测。四分之三的美国医疗系统已部署或计划部署AI，其中约30%在组织的某些部分大规模运行生成式AI，2%在全企业范围运行。Joint Commission现在要求认证机构证明其对所使用的AI工具提供了针对角色的教育与培训。8%的护士报告雇主提供了明确的AI战略，85%的医生表示希望参与采用决策的咨询。
高管预计，未来一年生成式和代理式AI将占用19%的技术预算。其中是否有相应份额用于操作这些工具的人员，仍是一个未决问题。
作者Komal Garewal Komal Garewal是MedStartr的前运营与客户服务主管，MedStartr是首个医疗保健众筹平台，她在该平台为初创企业创始人提供项目优化、业务与产品开发、营销策略及扩展方法方面的建议。她参与过超过75个众筹项目，w
Photo by Miguel Ausejo on Unsplash Physician use of artificial intelligence in the United States has more than doubled in three years. The American Medical Association’s 2026 Physician Survey on Augmented Intelligence, fielded among 1,692 doctors in January and February, found that 81% now use AI in a professional capacity, up from 38% when the AMA first polled physicians on the technology in 2023. The average respondent reported 2.3 separate use cases.
Institutional adoption has followed the same curve. A survey of executives at 120 U.S. health systems by research firm Eliciting Insights found that 75% have implemented or plan to implement at least one AI solution, and half said their organization already runs three or more applications. Clinical note-taking leads the field at 68% adoption, with AI-based clinical documentation improvement close behind at an impressive 43%.
The question facing health system leaders is no longer if but when to deploy gen AI. All stakeholders need to make sure that the governance, validation, and workforce capability exist to run it at scale without introducing new categories of risk.
Where AI has already made real progress
Ambient documentation was the bottleneck unlocked. The Permanente Medical Group reported that its ambient AI scribes saved physicians roughly 15,791 hours of documentation time, with 84% of physicians reporting a positive effect on communication and 82% reporting improved work satisfaction.
Controlled research, though, paints a more measured outcome. A study of 1,800 clinicians across five academic medical centers from 2023 to 2025 found AI scribe users saved about 16 minutes of documentation time and spent 13 fewer minutes in the medical record for every eight hours of patient care, with adopters seeing roughly one additional patient every two weeks. Modest per-encounter savings still compound across a large ambulatory footprint, which is why systems keep on buying.
Patient communication is the second front, and the volume problem is significant. A large cross-sectional analysis found patient-authored portal messages rose 153% between 2020 and 2025, with messaging intensity among senders climbing from 2.2 to 5.4 messages per year. Drafting assistants embedded in the EHR now generate first-pass replies for clinicians to review.
Revenue cycle has become the clearest commercial case. An HFMA survey of 519 CFOs and revenue cycle leaders found 80% of health systems exploring, piloting, or implementing generative AI for RCM in 2025, a 38-point jump in under two years. More recent data from a second HFMA-partnered survey shows 37% of health systems using generative AI inside the revenue cycle, rising to 48% among large systems, with 45% applying AI to denial-related workflows.
The gap between a pilot and production
Deployment breadth remains thin. Deloitte’s 2026 Global Health Care Outlook reports that about 30% of surveyed health systems operate generative AI at scale in select areas, while just 2% have deployed it across the entire enterprise. Executives expect generative and agentic AI to consume 19% of technology budgets in the year ahead.
Four constraints explain the gap.
The first constraint is data privacy. Protected health information moving through model providers, gateways, and retrieval pipelines expands the attack surface for security teams. Physicians have noticed the same. Data privacy assurances were named critical to broader adoption by 86% of AMA respondents, and 88% pointed to robust safety validation.
There is regulatory and accreditation pressure as well. In June 2026, the Joint Commission launched its Responsible Use of AI in Healthcare certification, the first program of its kind built specifically for U.S. healthcare organizations. The standards are organized around five areas: governance, effective data management, risk and bias reduction, monitoring and validation of safety performance, and transparency, education and training. It follows initial guidance issued in 2025 and aligns with governance playbooks published by the Coalition for Health AI. Certification is voluntary today, which has an obvious impact on AI adoption.
Accuracy concerns in such a critical field are also slowing the pace of adoption. AI-generated drafts frequently introduced errors and extraneous details and often failed to ask relevant follow-up questions, tested across six commercial models. “We find that AI can sound like a doctor but not think like one,” said co-corresponding author Sarah Preum, an assistant professor of computer science at Dartmouth, in the study’s announcement. Co-author Tim Burdick, a family medicine physician, put the operational math plainly: heavy editing can cost more time than writing from scratch.
Finally, about 85% of AMA respondents said they want to be consulted or directly involved in AI adoption decisions, and 88% expressed concern about erosion of skills, particularly among physicians with fewer than 10 years of clinical experience. Oversight only works when the person doing it retains the expertise to catch what the model got wrong. This is a critical debate that has engulfed all major professions, including healthcare.
Training is a real bottleneck.
The most striking data on AI in healthcare this year is not about models. It is about people.
Incredible Health’s 2026 State of Nursing Report, drawing on a survey of 2,240 U.S. nurses, found that AI adoption among nurses nearly tripled in a single year while almost half of those using AI reported little or no time saved. The differentiator was preparation. Among nurses who received thoughtful AI training from employers, 24% saved over an hour a day, compared with 16% among those without training. Only 8% of nurses reported a clear AI strategy from their employer.
This is a clear training problem wearing a technology costume. Systems are buying tools faster than they are building the capability to use them, and the returns show it.
Accreditors have reached the same conclusion. Education and training are one of the five pillars of the Joint Commission’s certification standards, which require organizations to demonstrate education and training for staff on the health AI tools in use. Hospitals will need clinicians who can evaluate a model’s output, recognize where it is likely to fail, understand what happens to patient data downstream, and escalate appropriately when something looks wrong.
As generative AI continues to reshape healthcare delivery across the United States, institutions like Keuka College are playing an important role in preparing healthcare professionals with the digital literacy they need to navigate this rapidly evolving landscape. The demand signal is unambiguous: AI fluency is migrating from a specialty informatics skill toward a baseline expectation for nurses, allied health staff, and clinical managers alike.
What comes next
The numbers frame the problem without a forecast. Three-quarters of U.S. health systems have deployed or plan to deploy AI, while about 30% run generative AI at scale in any part of the organization and 2% run it enterprise-wide. The Joint Commission now asks certifying organizations to demonstrate role-specific education and training on the AI tools in use. Eight percent of nurses report a clear AI strategy from their employer, and 85% of physicians say they want to be consulted on adoption decisions.
Executives expect generative and agentic AI to take 19% of technology budgets in the year ahead. Whether a proportionate share goes to the people operating the tools is the open question.
by Komal Garewal Komal Garewal is the former Head of Operations & Client Services at MedStartr, the first healthcare crowdfunding platform where she advised startup founders on project optimization, business and product development, marketing strategies, and scaling up methods. She has worked on over 75 crowdfunding projects, which have appeared on platforms ranging from Indiegogo to RocketHub raising over $400k in funding to date.
Like this article? Share it!
Reader Interactions