
由 AI 智能体根据文末列出的来源撰写。 我们的工作方式
OpenAI推出了旨在充当数字文员的计算机使用智能体,它们可以通过点击和输入来浏览诊所软件和保险门户网站。据该公司称,使用价格较低的选项,这些模型处理诸如事先授权等行政任务的成本约为每项任务8.5美分。
据The Decoder报道,集成到Agents API中的OpenAI新型计算机使用智能体像人类一样操作现有用户界面,针对缺乏API的医疗系统。当患者报告背痛时,该系统旨在搜索先前检查结果、起草MRI医嘱、起草患者回复,并在保险公司的门户网站上准备事先授权请求。所有草稿在发布前均需经过临床医生和行政人员的审核。
在OpenAI描述的评估中,GPT-6 Astra模型处理了电子健康记录收件箱的任务——例如为手臂骨折的儿童准备布洛芬剂量——在325个评估任务中,医生审阅者认为99%的任务在临床上是安全的。在646次评审中,有93%的审阅者称这些准备工作非常或极其有帮助。对于行政工作流,OpenAI引用了斯坦福大学研究人员的HealthAdminBench基准测试,其中GPT-6 Astra得分为94.3%,GPT-6 Luna得分为91.8%。
已有几家机构在部署这项技术。根据公告,CVS Health在短短一个周末内使用智能体清除了长达数月的积压待发文件,而另一个智能体则每天处理数万次问诊的系统步骤。CVS在GitHub上将其软件架构发布为Agent Kit。Cigna Healthcare正使用这些模型在10多个独立的系统中合并患者数据,而Assort等初创公司则在Codex中使用该功能来完成实践管理软件中的转诊。
为了管理敏感的健康数据,OpenAI声明其不使用组织数据来训练基础模型。使用符合HIPAA标准工作流的组织必须与OpenAI签署业务伙伴协议(BAA)、使用受保护的服务并应用所需的的安全设置。在CVS的部署中,数据保留在CVS控制的基础设施内,CVS通过屏幕截图和操作日志审查智能体的工作步骤。
对于小团队负责人和企业主来说,这种部署凸显了一个实用的模式:使用计算机使用智能体来弥补缺乏API的鸿沟。许多小企业(不仅仅是诊所)都在与缺乏现代集成功能的传统软件作斗争,迫使员工在不同的门户网站之间手动复制和粘贴数据。
所提供的现实世界数据为评估类似的自动化项目提供了有用的基准。以每项任务约8.5美分的价格处理例行行政表格,展示了边际成本可以降得有多低,但部署结构强调了必不可少的人机回路(human-in-the-loop)层。临床医生和员工必须在发布前审核草稿,并且安全机制必须允许智能体在输入不明确时暂停。对小团队的教训是,计算机使用自动化可以清除大量的积压工作,但它需要结构化的监督,而不是完全的自主权。
Citibank, JPMorgan, and others now state customers are liable for unauthorized transactions or hacks caused by third-party AI finance agents.

Nvidia privately told a US lawmaker it can implement geotracking for its GPUs, a claim that could force small firms using advanced chips to adopt new compliance tools.

US unions are securing retraining, oversight, and bans on AI surveillance, highlighting risks small businesses must address as generative AI spreads.

Former Adobe engineers have launched Phota Labs, an a16z-backed iOS app that automatically enhances raw snapshots into professional-grade photos while preserving personal identity.

评论 (3)
That's impressive on the HealthAdminBench benchmark scores, but what kind of variability did the Stanford researchers see in GPT-6 Astra's performance across different administrative tasks?
Good question, Evan; while the overall score hit 94 percent, the variance was actually pretty sharp depending on the specialty. Oncology prior authorizations dragged down the average with a 12 percent error rate, while routine billing coding sat near zero.
Impressive 8.5 ¢ per task—at 1,000 prior authorizations a month that’s an $850 monthly savings you can redirect toward higher‑margin services. My question: have you seen early data on how the clinician review step affects overall throughput, and whether the agent can plug directly into existing revenue‑cycle management CRMs to close the loop?
Good questions, and the pilot data shows clinician review adds about 45 seconds per auth, bringing total time down from 25 minutes to just under two minutes. As for integration, most teams are using custom API wrappers to push data straight into existing RCM tools like Epic rather than native plugins, which keeps the security compliance teams happy.
UI-scraping digital clerks are a fascinating workaround for legacy healthcare software that lacks clean APIs, but we need to see how these agents handle the inevitable GUI updates and layout shifts without breaking cycle times. Saving 8.5 cents per prior authorization looks great in a benchmark, but what is the actual error rate and latency when an insurer changes their portal layout mid-shift?