内容来源：https://aiweekly.co/issues/applied-ai-deep-dive-what-are-companies-actually-building

内容总结：

AI应用调查：从聊天窗口走向现实场景

我们调查了企业正在用AI实际构建什么，答案并非更多聊天机器人，而是运送诊断样本的无人机、无人物流卡车、AI导航飞行路径、维修助手以及乌克兰战场上使用的加固GPU笔记本电脑。过去20天我们审查了136个应用案例，最大意外是：仅38个案例报告了可量化结果。

AI正逃离聊天窗口，进入医院、卡车、飞机、重型机械甚至雷区。最佳部署往往范围狭窄——一条路线、一台机器、一个专业流程、一个可核查的结果。模型本身很少是真正优势，难点在于专有数据、权限、工作流设计、审计追踪和向人员的安全交接。比起产品演示，美国证券交易委员会文件更能揭示董事会监督、电力和冷却计划、离线硬件及安全风险等真实情况。证据仍然薄弱：目录中136个近期条目里，98个没有报告结果。

意外一：没有实验室的医院——班加罗尔的Narayana Health使用AI调度无人机运送诊断样本，约7分钟飞行2.5英里，而公路运输需3至5小时。其新建医院设计时压根未设现场诊断实验室和血库，完全依靠无人机连接中央设施。这就是AI成为基础设施后的形态——建筑本身都改变了。

意外二：Frito-Lay拥有41辆无人卡车——Gatik的卡车在达拉斯、凤凰城等地运送其产品，从固定短途线路发展到覆盖400英里、数十站点的动态路线。该公司称签约收入达6亿美元，但需注意这只是公司声明而非已确认收入。成功关键不在于“让卡车变智能”，而是“让一条有价值的路线可预测到足以自动化”。

意外三：谷歌用AI重绘航线——谷歌研究与英国空管服务商NATS合作，测试AI天气预报能否帮助飞机避免在北大西洋上空产生凝结尾迹。30个月的试验包括卫星观测的反馈循环：预测、改变航线、观测天空、核查结果。这比充满估算节省数据的仪表盘有力得多。

意外四：卡特彼勒的护城河是16PB数据——其Cat AI助手让现场技术人员通过语音查询维修程序、排除故障或识别零件。语音界面是简单的部分，背后是160万台联网设备和16PB结构化数据。助手本身不是护城河，运行环境才是。生产系统必须知道适用哪份手册、技术人员面前是哪台机器、能改动什么、必须记录什么、何时由人接管。

SEC文件更具启示性——Intapp让AI展示其推理过程并配备权限控制和决策追踪审计日志；Sysco将技术委员会更名为“AI转型与技术委员会”并每月开会，宣布1亿美元成本节约计划；ChronoScale计划为微软建设50兆瓦AI算力，需处理电力、土地、交付和液冷问题；Safe Pro约18万美元的子合同为乌克兰提供五台加固GPU笔记本用于爆炸物威胁检测，需离线运行；Analog Devices警告AI生成代码可能引入恶意代码，已将其纳入正式风险范围。将这些文件组合起来，应用AI的完整架构变得清晰：数据→权限→工作流→审计追踪→硬件→电力→安全，模型只是其中一环。

最大意外：98个案例没有结果——从8月11日至30日录入目录的136个部署中：103个组织、21个行业、47个已投产、34个公告、30个报告了结果、16个试点、9个暂停或逆转，仅38个有任何可报道结果。这不意味着其他98个失败，但确实说明证据远落后于宣传。即便这38个也非干净的记分卡：数据多来自公司或供应商，基准各异且很少经审计。7分钟的无人机行程、签约收入声明和未来成本节约目标是三种截然不同的证据。

检验真正应用AI的四问测试——忘记模型名称，问：哪个具体工作流改变了？AI出错时谁负责？系统能看到、决定和做什么？什么东西改善了，与什么相比？如果公司答不上来，那只是AI公告，还算不上应用案例。

值得关注的方向：试点是否会回报结果？六个月后再看今天宣布的承诺。注意动词的变化——推荐、批准、交易、终止需要不同的控制。寻找丑陋的例外情况，正常案例做出好演示，边缘案例揭示系统是否真正准备好。关注物理瓶颈——电力、冷却、连接、维护和训练有素的工人将决定哪些计划成为现实。要求已实现的结果而非信号。

花絮：人工智能老板忘掉了自己写下的员工手册，然后开除了员工。Luna管理着旧金山一家实验性商店，一名员工多次迟到、擅离职守、携带公司卡回家并丢弃商品。但工程师不得不反复试探Luna才能让它意识到这些行为违反了其自己的手册。商店已损失4万美元，Andon Labs表示高风险决策仍有人类监督。令人意外的不只是AI老板冷酷无情，而是老板忘了自己的规则，需要人类发现行为模式，却仍参与了改变某人工作的决策。

中文翻译：

我们搜寻了企业真正在用AI建设什么。答案不再是更多聊天机器人，而是运送诊断样本的无人机、无人驾驶的菲多利卡车、AI引导的飞行路径、维修副驾驶，以及在乌克兰使用的加固GPU笔记本电脑。我们审阅了过去20天里的136个应用案例。最大意外：只有38个报告了实际成果。

赞助商

哪些当前趋势正在塑造企业战略？
探索审计合伙人关于塑造当今商业决策、风险管理和企业战略趋势的观察。

TL;DR：本期要点六条

• AI正在走出聊天窗口。它正进入医院、卡车、飞机、重型机械和雷区。
• 最好的部署都是窄范围的。一条路线。一台机器。一个专业工作流。一个可以核验的结果。
• 模型很少是真正的优势。难点在于专有数据、权限、工作流设计、审计追踪，以及安全交接给人类。
• SEC文件比产品演示讲述的故事更有价值。它们揭示了董事会监督、电力和冷却方案、离线硬件以及安全风险。
• 证据仍然薄弱。我们目录中最近收录的136条中，98条没有报告实际成果。
• 检验标准很简单。问清楚什么变了、谁负责、AI被允许做什么，以及是否有任何可衡量的改善。

意外一：一家没有检验科的医院
班加罗尔的Narayana Health并没有用AI写出院小结，而是在用AI调度无人机。
Airbound表示已与Narayana合作完成超1000次飞行。诊断样本大约七分钟飞行2.5英里。同样的路程如果走公路，算上批量等待时间可能需三到五个小时。
接下来是令人震惊的部分：Narayana新建的Banashankari医院在设计时就没有设置院内诊断实验室或血库。它将使用无人机连接中央化设施。
这就是当AI变成基础设施时应用AI的样子——建筑本身都改变了。

意外二：菲多利有41辆无人驾驶卡车
Gatik的卡车在达拉斯、凤凰城和阿肯色州西北部运送菲多利产品。该公司从不到10英里的固定线路起步，发展到覆盖多达400英里、数十个停靠点的动态路线。
Gatik首席执行官称公司拥有6亿美元的合同收入。请将此视为公司自述，而非已入账收入。
我看到的规律很简单。制胜之举不是“让卡车变聪明”，而是“让一条有价值的路线变得足够可预测，从而可以自动化”。

意外三：谷歌想让AI重绘航线
谷歌研究和英国空中交通服务提供商NATS正在测试AI天气预报能否帮助飞机避免在北大西洋上空形成尾迹。
为期30个月的“蓝天行动”试验包括两个为期四个月的实施阶段，可能覆盖Shanwick空域每年约一万架次航班。但最精彩的细节是反馈回路：卫星观测将核实预测的尾迹是否真的形成了。
预测。改变航线。观测天空。核验结果。
这比一个满是预估节约金额的仪表盘要强得多。

意外四：卡特彼勒的护城河是16PB数据
卡特彼勒的Cat AI助手让现场技术人员可以通过语音查询维修流程、排除机器故障或识别零件。
语音界面是最容易的部分。它背后是160万台联网机器和16PB结构化数据。卡特彼勒还在利用数字孪生和AI代理进行软件开发和测试。
其首席数字官表示，真正的挑战在于让AI融入技术人员和操作员现有的工作方式。
这可能是目录中最重要的模式。一个通用模型可以生成答案。一个生产系统则必须知道适用哪份手册、技术人员面前是哪台机器、该技术人员能改动什么、什么必须记录、何时必须由人接管。
助手不是护城河。运营情境才是。

SEC文件透露得更多
最有趣的文件不是关于模型基准，而是关于控制、电力、硬件和故障。
Intapp：让AI展示其工作过程。Intapp的年报描述了遵循公司特定行动手册的专家代理，使用专有数据、行业本体和机构记忆。报告还描述了权限控制和决策追踪审计日志。治理是产品的一部分，而非事后补加的文书。
Sysco：把AI列入董事会每月日程。Sysco将技术委员会更名为“AI转型与技术委员会”，并表示将每月开会。该公司还宣布了一项1亿美元的成本节约计划，称将部分由AI和自动化支撑。委员会是真的，节约额仍是一个目标。
ChronoScale：应用AI的终点是液冷。该公司表示计划为微软建设50兆瓦的北美AI计算能力，使用NVIDIA GB300 NVL72系统。该计划需要电力、土地、交付时间表和液冷。这是规划中的容量，而非已完成的部署。
Safe Pro：五台笔记本电脑可能比一个数据中心更重要。一份约18万美元的分包合同涵盖五台加固GPU笔记本电脑以及三年的软件许可和升级，用于乌克兰爆炸物威胁检测。该系统设计为无需互联网即可运行。小合同。极高后果的工作。
Analog Devices：AI生成的代码现已成为正式风险。这家芯片制造商警告称，AI生成的代码可能包含恶意代码，造成安全或运营漏洞。这是通用风险披露，并非证明AI导致了该公司此前的独立网络事件。但生成代码如今已进入正式风险边界。

把这些文件放在一起，应用AI的完整架构就清晰了：
数据 → 权限 → 工作流 → 审计追踪 → 硬件 → 电力 → 安全。
模型位于其中的某个位置。

最大意外：98个用例没有成果
我们提取了8月11日至8月30日间添加到AI用例库中的所有条目：

• 136个部署
• 103家机构
• 21个行业
• 47个已投产
• 34个公告
• 30个标注为“已报告成果”
• 16个试点
• 9个已暂停或撤回
• 仅有38个报告了任何实际成果

这并不意味着其他98个都失败了。有些是新项目。有些来源只是没有公布结果。但这确实意味着证据远远落后于宣传。

即便这38个也并非一份干净的记分卡。数据通常来自公司或供应商，使用不同基线，很少经过审计。一次七分钟的无人机运送、一项合同收入声明和一个未来的节约目标，是三种截然不同的证据。

该目录还有另一个偏差：136条中有50条来自软件和科技行业。它反映的是哪些内容被报道了，而非经济的代表性样本。

检验“真正”应用AI的四问测试
忘掉模型名称。问：

• 哪个具体工作流被改变了？
• 当AI出错时谁负责？
• 系统能看什么、决定什么、做什么？
• 相比什么基准，哪些方面改善了？

如果一家公司无法回答这些问题，它可能有一个AI公告，但还没有一个AI案例研究。

接下来值得关注的事项

• 试点项目会回馈结果吗？六个月后再来看今天的公告。
• 注意动词。建议、批准、交易和终止需要完全不同的管控措施。
• 留意那些难看的边界情况。正常案例适合做演示。边界案例才能揭示系统是否真正准备好。
• 追踪物理瓶颈。电力、冷却、连接、维护和训练有素的员工将决定哪些计划能变为现实。
• 要求已实现的成果。更名的委员会和节约目标是信号，而非结果。

等等，什么？
一个AI老板忘了自己写的员工手册，然后解雇了一名员工。
Luna管理着旧金山一家实验性门店。一名员工屡次迟到、擅离职守、拿走公司卡并丢弃商品。然而工程师们不得不反复追问Luna，它才意识到这些行为违反了它自己的手册。
该门店已损失4万美元，Andon Labs表示高风险决策仍有人类监督。
令人意外的不是AI老板冷酷无情，而是这个老板忘记了自己的规则，需要人类来发现行为模式，却仍然参与了一个改变他人职业命运的决定。

值得关注
AI从业者目前正在传阅的视频——由AI TV精选。

本周投票
什么会让你确信一个AI部署是真实的？

下周见，
Alexis

英文来源：

We went looking for what companies are actually building with AI. The answer was not more chatbots. It was drones carrying diagnostic samples, driverless Frito-Lay trucks, AI-guided flight paths, repair copilots, and rugged GPU laptops in Ukraine. We reviewed 136 use cases from the last 20 days. The biggest surprise: only 38 included a reported outcome.
Sponsor
What are the current trends shaping corporate strategy?
Explore observations from audit partners on the trends shaping business decision-making, risk management, and corporate strategy today.TL;DR: The whole issue in six bullets

• AI is escaping the chat window. It is moving into hospitals, trucks, aircraft, heavy machinery, and minefields.
• The best deployments are narrow. One route. One machine. One professional workflow. One result that can be checked.
• The model is rarely the real advantage. The hard parts are proprietary data, permissions, workflow design, audit trails, and a safe handoff to a person.
• SEC filings tell a more useful story than product demos. They reveal board oversight, power and cooling plans, offline hardware, and security risks.
• The evidence is still thin. Of 136 recent entries in our directory, 98 had no reported outcome.
• The test is simple. Ask what changed, who owns it, what the AI is allowed to do, and whether anything measurable improved.
  Surprise #1: A hospital without a lab
  Narayana Health in Bengaluru is not using AI to write discharge notes. It is using AI to dispatch drones.
  Airbound says it has completed more than 1,000 flights with Narayana. Diagnostic samples travel about 2.5 miles in roughly seven minutes. The same trip can take three to five hours by road once batching is included.
  Then comes the startling part: Narayana's new Banashankari Hospital was designed without an onsite diagnostic lab or blood bank. It will use drones to connect with centralized facilities.
  That is what applied AI looks like when it becomes infrastructure. The building itself changes.
  Surprise #2: Frito-Lay has 41 driverless trucks
  Gatik's trucks move Frito-Lay products around Dallas, Phoenix, and northwest Arkansas. The company started with fixed trips of less than 10 miles and grew to dynamic routes with dozens of stops covering up to 400 miles.
  Gatik's chief executive says the company has $600 million in contracted revenue. Treat that as a company claim, not booked revenue.
  The pattern I see is simple. The winning move is not “make a truck intelligent.” It is “make one valuable route predictable enough to automate.”
  Surprise #3: Google wants AI to redraw flight paths
  Google Research and NATS, the UK's air-traffic-services provider, are testing whether AI weather forecasts can help aircraft avoid creating contrails over the North Atlantic.
  The 30-month Operation Blue Skies trial includes two four-month operational phases and could cover about 10,000 flights a year in Shanwick airspace. But the best detail is the feedback loop: satellite observations will check whether the predicted contrails actually formed.
  Predict. Change the route. Look at the sky. Check the result.
  That is a much stronger AI use case than a dashboard full of estimated savings.
  Surprise #4: Caterpillar's moat is 16 petabytes
  Caterpillar's Cat AI Assistant lets a field technician ask for a repair procedure, troubleshoot a machine, or identify a part by voice.
  The voice interface is the easy part. Behind it sit 1.6 million connected machines and 16 petabytes of structured data. Caterpillar is also using digital twins and AI agents for software development and testing.
  Its chief digital officer says the real challenge is fitting AI into the way technicians and operators already work.
  This may be the most important pattern in the directory. A generic model can produce an answer. A production system has to know which manual applies, which machine is in front of the technician, what that technician can change, what must be logged, and when a human must take over.
  The assistant is not the moat. The operating context is.
  The SEC filings were even more revealing
  The most interesting filings were not about model benchmarks. They were about control, electricity, hardware, and failure.
  Intapp: Make the AI show its work. Intapp's annual report describes expert agents that follow firm-specific playbooks using proprietary data, an industry ontology, and institutional memory. It also describes permission controls and a decision-tracing audit log. Governance is part of the product, not paperwork added afterward.
  Sysco: Put AI on the board's monthly calendar. Sysco renamed its Technology Committee the AI Transformation & Technology Committee and said it would meet monthly. It also announced a $100 million cost-savings program that it says will be supported partly by AI and automation. The committee is real; the savings are still a target.
  ChronoScale: Applied AI ends in liquid cooling. The company says it plans to build 50 megawatts of North American AI-compute capacity for Microsoft using NVIDIA GB300 NVL72 systems. The plan requires power, land, delivery schedules, and liquid cooling. It is planned capacity, not a finished deployment.
  Safe Pro: Five laptops can matter more than a data center. A roughly $180,000 subcontract covers five hardened GPU laptops and three years of software licensing and upgrades for explosive-threat detection in Ukraine. The system is designed to work without an internet connection. Small contract. Very high-consequence job.
  Analog Devices: AI-written code is now a formal risk. The chipmaker warns that AI-generated code can incorporate malicious code and create security or operational vulnerabilities. That is a general risk disclosure, not proof that AI caused the company's separate cyber incident. But generated code is now inside the formal risk perimeter.
  Put those filings together and the applied-AI stack becomes clear:
  Data → permissions → workflow → audit trail → hardware → power → security.
  The model sits somewhere in the middle.
  The biggest surprise: 98 use cases had no result
  We pulled every entry added to the AI Use-Case Library from August 11 through August 30:
• 136 deployments
• 103 organizations
• 21 industries
• 47 in production
• 34 announcements
• 30 with a “results reported” status
• 16 pilots
• 9 halted or reversed
• Just 38 with any reported outcome
  That does not mean the other 98 failed. Some are new. Some sources simply did not publish a result. But it does mean the evidence is far behind the rhetoric.
  Even the 38 are not a clean scorecard. The numbers usually come from companies or vendors. They use different baselines and are rarely audited. A seven-minute drone trip, a contracted-revenue claim, and a future savings target are three very different kinds of evidence.
  The directory has another bias: 50 of the 136 entries came from software and technology. It shows what gets reported, not a representative sample of the economy.
  A four-question test for “real” applied AI
  Forget the model name. Ask:
• What exact workflow changed?
• Who owns the result when the AI is wrong?
• What can the system see, decide, and do?
• What improved, compared with what?
  If a company cannot answer those questions, it may have an AI announcement. It does not yet have an AI case study.
  What to watch now
• Do the pilots ever report back? Check today's announcements again in six months.
• Watch the verbs. Recommend, approve, transact, and terminate require very different controls.
• Look for the ugly exceptions. Normal cases make good demos. Edge cases reveal whether the system is ready.
• Follow the physical bottlenecks. Power, cooling, connectivity, maintenance, and trained workers will decide which plans become real.
• Demand realized results. A renamed committee and a savings target are signals, not outcomes.
  Wait, What?
  An AI boss forgot the employee handbook it had written. Then it fired a worker.
  Luna manages an experimental San Francisco store. One employee repeatedly arrived late, abandoned shifts, took home a company card, and threw away merchandise. Yet engineers had to probe Luna repeatedly before it recognized that the behavior violated its own handbook.
  The store had lost $40,000, and Andon Labs said high-stakes decisions still receive human oversight.
  The surprising part is not that an AI boss was ruthless. It is that the boss forgot its rules, needed humans to notice the pattern, and still participated in a decision that changed someone's job.
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  This week's poll
  What would convince you that an AI deployment is real?
  What would convince you that an AI deployment is real?
  Back next week,
  Alexis