超越 AI 抄写员：为什么环境式临床智能是医疗 IT 最大的治理考验
Beyond AI Scribes: Why Ambient Clinical Intelligence Is Health IT’s Greatest Governance Test
Midland Health 数据架构师 Akhila Akula
两年前，环境式 AI 抄写员还只是一个新奇事物。少数几个医疗系统开展了试点。会议演示吸引了大批观众，但没有带来采购订单。那个阶段已经结束。
根据发表于《American Journal of Managed Care》的一项研究，截至 2025 年年中，在使用 Epic 的美国医院中，近三分之二已经部署了环境式 AI 文档工具——共计 1,744 家。在单个医生中，环境式文档记录是增长最快的 AI 用例。Doximity 对 3,100 多名医生开展的 2026 年调查发现，不到一年的时间里，基于语音的文档记录使用率从 20% 跃升至 29%。Grand View Research 的分析师预计，到 2033 年，美国医疗抄写员市场将接近 30 亿美元，高于 2024 年约 4 亿美元的规模。
与大多数需要经历 18 个月采购周期、并伴随医生勉强采用的医疗 IT 部署不同，环境式 AI 的传播部分源于基层需求。医生告诉其他医生，这项技术确实有效。
但速度也会带来自身的问题。那些快速推进环境式文档记录的组织如今发现，这项技术的发展速度已经超过了其治理结构的跟进能力。在得克萨斯州一家多专科门诊机构，Suki 于今年早些时候面向门诊提供者上线。整体反响大多是积极的，但 27 名提供者中有两人拒绝使用该工具，理由是担心患者对话的机密性。人数虽少，但原因很重要。这些提供者并不是对技术持怀疑态度的人。他们是在认真思考，让一个 AI 系统处于诊疗现场究竟意味着什么。
承诺依然成立——大体上如此
两年前，环境式 AI 抄写员还只是一个新奇事物。
其核心理念足够简单。软件聆听医生与患者之间的对话，生成结构化临床记录，然后由提供者进行核验并将其推送至 EHR。
越来越多的证据已经难以忽视。2025 年 10 月发表于 JAMA Network Open 的一项研究涵盖了六个医疗系统的 263 名临床医生，发现使用环境式抄写员仅 30 天后，职业倦怠率就从 51.9% 降至 38.8%，同时认知负荷、下班后文档记录以及对患者的专注程度也得到可测量的改善。
但细节呈现出不同的情况。一项针对五家学术医疗中心开展的 AI 抄写员 ROI 研究——Mass General Brigham、Emory Healthcare、UC-San Francisco、UC-Davis 和 Yale New Haven Health——发现收益更加有限。在每 8 小时的计划诊疗时间内，文档记录时间减少了 16 分钟，在 EHR 中的时间减少了 13 分钟。这些改善具有意义，但与销售演示材料所描述的转型相距甚远。
热情与证据之间的差距值得我们认真思考。ROI 在很大程度上取决于实施质量，也取决于组织是围绕该工具重新设计工作流程，还是只是将其叠加到现有流程之上。得克萨斯州的这家机构在其工作流程中加入了同意步骤：每次就诊开始时弹出提示，让患者在开始录音前选择“是”或“否”。没有同意，就不录音。根据 Black Book Research 的数据，只有 8% 的采用者在第一年实现了正 ROI。大多数组织预计在 24 至 30 个月内获得回报。
从抄写员到 Agent：无人治理的转变
环境式 AI 最初是一种文档工具。它负责聆听和打字。但供应商正朝着更具雄心的方向发展：自主临床 Agent 不仅记录就诊内容，还会据此采取行动。
下一代工具可以聆听一次诊疗，识别护理缺口，例如一名未使用 station 的糖尿病患者；为医生预填订单，供其签署；并起草事先授权申请。所有这些都可以实时完成。
这是一类不同的产品。文档记录属于事务性工作。临床决策支持属于医疗行为。责任暴露会发生变化，监管要求会发生变化，采购评估也应该随之变化。然而，大多数医疗系统仍将其视为自然的功能升级，而不是类别变化。
下一次软件更新前，CIO 应该提出什么问题
FDA 于 2026 年 1 月发布的临床决策支持修订指南划出了一条清晰界线：只有当临床医生能够独立核验底层逻辑时，CDS 软件才可免受医疗器械监管。大多数环境式文档工具都安全地处于这条界线的文档一侧。但随着供应商加入护理缺口检测和订单预填充功能，这条界线很快就会变得模糊。能够起草记录的工具是一回事。能够将处方排入待处理队列的工具则完全是另一回事，而该指南实际上从未使用“AI”一词，这留下了一个行业必须自行应对的空白。
对于医疗 IT 领导者而言，目前最重要的是三个问题。第一：在供应商的路线图中，文档记录止于何处，临床决策支持又从何处开始？如果这些功能即将推出，合规和法务部门现在就需要参与讨论。第二：当工具出错时会发生什么？应要求供应商提供来自真实生产部署的错误率，而不是经过筛选的案例研究。第三：你们是否已经建立治理机制？根据 Black Book Research 的数据，只有 10% 的组织设有正式的 AI 监督委员会。得克萨斯州那两名提供者提出的机密性担忧并不是个别情况，而是一个信号。如果你的框架没有界定何时适合使用环境式 AI、何时不适合，提供者就会自行做出决定，而且彼此不一致。
这对我们意味着什么？
环境式临床智能如今已经成为吸引医生招聘的基本条件。它不再只是一个创新故事，而是一个基础设施故事。但进入主流并不意味着一路上的难题已经得到解决。
那些将环境式 AI 视为又一次软件采购的医疗 IT 领导者，最终得到的将是“每天节省 16 分钟”这一版本的故事。真正从中获得更多价值的组织，会将其视为一项临床运营决策，并在其背后配套治理、培训和工作流程重新设计。技术已经准备就绪。问题在于，采购它的组织是否准备好了。
关于 Akhila Akula
Akhila Akula 是 Midland Health 的数据架构师和高级数据工程师。她负责设计临床和财务数据系统。她撰写有关医疗 IT、AI 治理以及在医疗行业部署临床 AI 的运营现实的文章。
来源：
Graetz 等，《美国医院中的环境式 AI 工具采用及相关因素》，American Journal of Managed Care（2026 年 1 月）https://www.ajmc.com/view/ambient-ai-tool-adoption-in-us-hospitals-and-associated-factors Doximity，《2026 年医学领域 AI 现状报告》https://www.doximity.com/reports/state-of-ai-medicine-report/2026 Grand View Research，美国医疗抄写员市场预测https://www.grandviewresearch.com/industry-analysis/us-ai-medical-scribing-market-report Olson 等，《使用环境式 AI 抄写员减少行政负担和职业倦怠》，JAMA Network Open（2025 年 10 月 2 日）https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2839542五中心 AI 抄写员 ROI 研究：Mass General Brigham、Emory、UCSF、UC-Davis、Yale New Haven Health（2026 年）https://pmc.ncbi.nlm.nih.gov/articles/PMC13044793/Black Book Research，临床文档记录中的 AI 调查，2025 年第 1 至第 3 季度https://blackbookmarketresearch.com/uploads/pdf/399U19-TBB-Global-Healthcare-IT-Q2-2025-Update.pdf FDA《临床决策支持软件指南》，发布于 2026 年 1 月 6 日（2026 年 1 月 29 日重新发布）https://www.fda.gov/media/191560/download
Akhila Akula, Data Architect at Midland Health Two years ago, ambient AI scribes were a mere curiosity. A handful of health systems ran the pilots. Conference demos drew crowds but not purchase orders. That phase is over.
Nearly two-thirds of U.S hospitals running Epic had deployed an ambient AI documentation tool by mid-2025 – 1,744 of them – according to a study published in the American Journal of Managed Care. Among the individual physicians, ambient documentation is the fastest-growing AI use case. Doximity’s 2026 survey of more than 3,100 physicians found voice-based documentation use jumped from 20 to 29 percent in under a year. Analysts at Grand View Research project the U.S. medical scribe market will approach $3 billion by 2033, up from roughly $400 million in 2024.
Unlike most health IT rollouts, which drag through an 18-month procurement cycle and grudging physician adoption, ambient AI spread partly through grassroots demand. Doctors told other doctors it worked.
But speed creates its own problems. Organizations that moved fast on ambient documentation are now discovering that the technology is evolving faster than their governance structures can keep up. At one multispecialty ambulatory practice in Texas, Suki went live for outpatient providers earlier this year. The response was mostly positive, but two of the 27 providers declined the tool, citing concerns about the confidentiality of their patient conversations. It is a small number, but the reason matters. These were not providers skeptical of technology. They were providers thinking carefully about what it means to have an AI system in the room.
The Promise Still Holds – Mostly
Two years ago, ambient AI scribes were a mere curiosity.
The core pitch is simple enough. Software listens to doctor-patient conversations, generates a structured clinical note and providers verify and push it into the EHR.
The evidence has gotten harder to ignore. A study published in JAMA Network Open in October 2025, covering 263 clinicians across six health systems, found burnout dropped from 51.9 to 38.8 percent after just 30 days of ambient scribe use alongside measurable improvements in cognitive load, after-hours documentation and focused attention on patients.
The fine print tells a different story. A study of AI scribe ROI across five academic medical centers Mass General Brigham, Emory Healthcare, UC-San Francisco, UC-Davis and Yale New Haven Health found more modest gains. 16 fewer minutes on documentation and 13 fewer minutes in the EHR per eight hours of scheduled care. Meaningful, but a long way from the transformation in the sales deck.
The gap between enthusiasm and evidence is worth sitting on. ROI depends heavily on implementation quality and whether organizations redesign workflows around the tool or simply drop it on top of existing ones. The Texas practice built a consent step into their workflow, a prompt at the start of each encounter that lets the patient choose yes or no before recording starts. No consent, no recording. Only 8 percent of adopters reached positive ROI in year one, according to Black Book Research. Most expect returns within 24 to 30 months.
From Scribe to Agent: The Shift Nobody Is Governing
Ambient AI started as a documentation tool. It listened and typed. But vendors are building towards something more ambitious: autonomous clinical agents that do not just document the visit but act on it.
The next generation can listen to an encounter, identify a care gap, such as a diabetic patient not on station, pre-populate orders for physicians to sign and draft a prior authorization request. All in real time.
That is a different kind of product. Documentation is clerical work. Clinical decision support is medicine. The liability exposure changes, the regulatory requirements change and the procurement conversion should change too. Yet most health systems are treating this as a natural feature upgrade rather than a category change.
What CIOs Should be Asking Before the Next Software Update
The FDA’s January 2026 revised guidance on clinical decision support draws a clear line: CDS software escapes medical device regulation only if the clinician can independently verify the underlying logic. Most ambient documentation tools sit safely on the documentation side of that line. But as vendors add care gap detection and order pre-population, that line blurs fast. A tool that drafts a note is one thing. A tool that queues a prescription is something else entirely and the guidance never actually uses the word “AI”, leaving a gap the industry will have to navigate on its own.
For health IT leaders, three questions matter most right now. First: where does documentation end and clinical decision support begin in your vendor’s roadmap? If those features are coming, compliance and legal need to be in the conversation now. Second: what happens when the tool is wrong? Push for error rates from real production deployments, not curated case studies. Third: do you have governance in place? Only 10 percent of organizations have formal AI oversight boards, per Black Book Research. The confidentiality concern raised by those two providers in Texas is not an edge case; it is a signal. If your framework does not define what ambient AI is appropriate and when it is not, providers will decide for themselves, inconsistently.
Where this leaves us?
Ambient clinical intelligence is now table stake for physician recruitment. It is no longer an innovation story; it is an infrastructure story. But going mainstream does not mean the hard questions got answered along the way.
Health IT
leaders who treat ambient AI as just another software purchase will get the 16-minute-a-day version of the story. The organizations getting more out of it are treating it as a clinical operations decision, with governance, training and workflow redesign behind it. The technology is ready. The question is whether the organizations buying it are.
About Akhila Akula Akhila Akula is a Data Architect and Senior Data Engineer at Midland Health. She designs clinical and financial data systems. She writes about health IT, AI governance and the operational realities of deploying clinical AI in the healthcare industry.
Sources:
Graetz et al., “Ambient AI Tool Adoption in US Hospitals and Associated Factors,” American Journal of Managed Care (Jan 2026) https://www.ajmc.com/view/ambient-ai-tool-adoption-in-us-hospitals-and-associated-factors Doximity, 2026 State of AI in Medicine Report https://www.doximity.com/reports/state-of-ai-medicine-report/2026 Grand View Research, U.S. medical scribe market forecast https://www.grandviewresearch.com/industry-analysis/us-ai-medical-scribing-market-report Olson et al., “Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout,” JAMA Network Open (Oct 2, 2025) https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2839542 Five-center AI scribe ROI study: Mass General Brigham, Emory, UCSF, UC-Davis, Yale New Haven Health (2026) https://pmc.ncbi.nlm.nih.gov/articles/PMC13044793/ Black Book Research, AI in Clinical Documentation surveys, Q1-Q3 2025 https://blackbookmarketresearch.com/uploads/pdf/399U19-TBB-Global-Healthcare-IT-Q2-2025-Update.pdf FDA Clinical Decision Support Software Guidance, issued Jan 6, 2026 (re-issued Jan 29, 2026) https://www.fda.gov/media/191560/download Reader Interactions