内容来源：https://aiweekly.co/issues/ai-got-too-expensive-so-companies-are-moving-to-cheaper

内容总结：

过去三周，企业在AI应用上呈现出三类显著动向：一是通过更换模型来控制成本，二是大规模部署已进入日常使用阶段，三是越来越多场景对AI施加限制或直接撤除。

成本压力倒逼模型切换

Harvey是一家向律所销售AI工具的公司，其产品最初构建于OpenAI和Anthropic的模型之上，按token计费。据彭博社报道，随着token使用量增长20倍，Harvey的毛利率从约50%降至负50%。今年8月，Harvey推出基于Moonshot开放权重模型Kimi K3的自研模型，毛利率重新转正。彭博社指出，Abridge和Ramp等初创公司也在走同样的路。开放权重模型允许企业下载并在自有服务器上运行，付费对象从按token计价的公司变为算力成本。

据《金融时报》报道，开放模型目前承担了AT&T约40%的AI工作负载。大型实验室的专有模型仍占据企业AI支出的大部分，但变化在于，大买家开始将模型视为价格不合适时即可替换的组件。

卖方定价策略也在调整。据The Information报道，Anthropic在客户用完合同用量后，会终止通常约15%的企业折扣，客户需重新谈判或按标价付费。对于在企业中负责AI运营的人来说，实际的经验是衡量“每完成一项任务的成本”。掌握这一数字的团队才有谈判空间，或在质量允许时将工作迁移至更便宜的模型。

AI已进入日常使用的场景

据路透社报道，麦当劳在美国近1.4万家餐厅使用机器学习为菜单项建议“最优价格”，系统分析每日数百万笔交易。在弗雷斯诺，同一公司直营的两家门店相距两英里，巨无霸售价分别为5.69美元和6.89美元。部分加盟商向路透社表示，该工具拉大了邻近门店的价格差距，也有人称受到使用压力。麦当劳将该工具描述为建议性质，而非强制要求。

OpenEvidence允许医生提出临床问题并从医学研究中获取答案。据Business Insider报道，目前约40%的美国医生在使用该工具。

据《纽约时报》报道，DraftKings在2023年建立了一个模型，根据客户在收到免费投注后预计亏损的金额进行评分。高管将AI驱动的促销归功于2025年体育博彩利润率约13%的提升。前员工向《纽约时报》表示，用于标记问题赌博的类似工作被搁置。DraftKings称其促销针对的是持续活跃的用户。

据彭博社报道，Devin编程智能体的开发商Cognition年化收入有望达到10亿美元，高于5月的4.92亿美元。客户包括英伟达、花旗集团和梅赛德斯-奔驰。

受到限制或撤回的案例

沃尔玛禁止门店使用AI制作的标识。Business Insider获取的一份沃尔玛备忘录显示，“门店不应展示由AI工具制作的标识”。门店现须通过公司批准的渠道订购标识。

三家公司在数据保留问题上限制了Anthropic的Fable模型。据The Information报道，在Anthropic开始保留30天使用日志后，英伟达、Palantir和博思艾伦限制了对该模型的使用。博思艾伦禁止员工将其用于专有网络安全软件工作。

洛杉矶联合学区暂时在发给学生的设备上屏蔽了生成式AI，教师仍保留访问权限。

荷兰眼镜连锁店Hans Anders在荷兰和比利时停售Ray-Ban Meta智能眼镜，理由是围绕智能眼镜的公众和政治讨论。

开源项目System76禁止任何LLM生成的内容出现在其Pop!_OS项目（包括COSMIC桌面）的issue和pull request中，规则覆盖代码、评论和描述。

英国健康观察组织的一项审查发现，一款AIscribe将检验结果中的“null demyelination”错误转换为意思相反的结果，原因是它漏掉了“null”一词。

监管文件中的AI披露

企业也在向美国证券交易委员会提交的文件中描述AI。近期四个例子：

• Paychex推出AI招聘工具。其季度业绩公告了WISE Hire这一“智能体招聘解决方案”，并称“早期采用者结果令人信服”，但未提供具体数字。Paychex服务约84万客户。
• Guidewire表示客户可能用AI自行构建软件。这家保险软件制造商在年报中警告，客户可能利用AI构建“可能减少或消除对我们解决方案需求”的工具。
• AMC将AI电影列为业务风险。这家院线连锁将“观众对使用AI技术制作的电影的接受度”列入风险因素。
• People Inc将AI授权计为收入。这家出版商将其内容“用于大语言模型”列为授权版税来源之一，与Apple News+协议并列。

等等，什么？

50万张人脸被扫描，零逮捕。英国交通警察在伦敦火车站测试实时面部识别摄像头，扫描超过50万张人脸，产生一次警报，为误匹配，未导致任何逮捕。

中文翻译：

本期《应用AI》观察过去三周各组织实际如何使用AI。三种模式尤为突出：企业为控制成本而更换模型、大型部署已进入日常使用、以及越来越多的场景对AI加以限制或撤除。

赞助商

零数据留存就该是零。但在大多数AI方案中，它意味着30天。

大多数AI聊天工具会把你团队分享的每一条提示和文件至少保留30天，而且往往存在共享服务器上。Faradex为每个项目提供独立的AI服务器。不存储、不训练。每月120美元起。

AI账单成了管理难题

Harvey的利润率转为负值，于是它换了模型

Harvey向律所销售AI工具。这些工具建立在OpenAI和Anthropic的模型之上，而这两家按模型读写的每一段文本收费。这些文本单位被称为token。

据彭博社报道，随着Harvey的token用量增长20倍，其毛利率到6月从约50%跌至负50%。8月，Harvey推出了自研模型，基于月之暗面的开放权重模型Kimi K3构建，毛利率随之回正。彭博社还点名Abridge和Ramp等初创公司也在走同样的路。

开放权重模型是企业可以下载并在自有服务器上运行的模型。它为算力付费，而不是为每一个token向另一家公司付费。

开放模型如今承担了AT&T AI工作的一大部分

据《金融时报》报道，开放模型目前处理AT&T约40%的AI工作负载。

大型实验室的专有模型仍然拿走了企业在AI上支出的大部分资金。变化在于，大买家现在把模型视为一种可以在价格不合适时随时替换的东西。

Anthropic在客户用满额度后终止折扣

定价在卖方一侧也在变化。据The Information报道，一旦客户用完合同约定的用量，Anthropic就会终止其企业折扣——通常约为15%。客户随后必须重新谈判，或按标价付费。

对任何在工作中运行AI的人来说，实际的教训是衡量每完成一项任务的成本。知道这个数字的团队可以谈判，或在质量允许时将工作转移到更便宜的模型上。

AI如今在日常使用中的场景

麦当劳用算法定价

据路透社报道，麦当劳使用机器学习为全美近1.4万家餐厅的菜单项建议“最优价格”。该系统分析每日数百万笔交易。

在弗雷斯诺，一家公司直营店的大麦克卖5.69美元，而两英里外的另一家卖6.89美元。一些加盟商告诉路透社，该工具拉大了附近门店之间的价格差距，还有人表示他们被施压使用它。麦当劳称该工具只是建议，而非强制。

约40%美国医生在用的医学搜索工具

OpenEvidence让医生提出临床问题，并从医学研究中获取答案。据Business Insider报道，目前约40%的美国医生在使用它。

DraftKings用AI定向投放促销

据《纽约时报》报道，DraftKings在2023年构建了一个模型，根据客户在获得免费投注后预计会输掉多少钱来给他们打分。高管将AI驱动的促销归功于2025年体育博彩利润率提升约13%。

前员工告诉《纽约时报》，用于标记问题赌博的类似工作却被搁置。DraftKings表示，其促销针对的是持续活跃的用户。

编程智能体进入大公司

据彭博社报道，Devin编程智能体的开发商Cognition有望实现10亿美元的年化营收，高于5月的4.92亿美元。客户包括英伟达、花旗集团和梅赛德斯-奔驰。

哪些被限制或撤回

沃尔玛禁止AI制作的店内标识。Business Insider报道的一份沃尔玛备忘录称：“门店不得展示由AI工具制作的标识。”门店现在通过经批准的公司渠道订购标识。

三家公司因数据留存问题限制使用Anthropic的Fable。据The Information报道，在Anthropic开始保留30天使用日志后，英伟达、Palantir和博思艾伦限制了对该模型的使用。博思艾伦禁止员工将其用于专有网络安全软件工作。

洛杉矶学区在学生设备上屏蔽AI。洛杉矶联合学区临时在发给学生的学校设备上屏蔽生成式AI。教师仍可访问。

荷兰一家眼镜连锁店停售摄像眼镜。Hans Anders在荷兰和比利时停售雷朋Meta眼镜，理由是围绕智能眼镜的公众和政治争论。

一个开源项目拒绝AI编写的代码。System76现在禁止任何LLM生成的内容出现在其Pop!_OS项目（包括其COSMIC桌面）的issue和pull request中。该规则涵盖代码、评论和描述。

英国NHS的AI转录工具漏掉一个词，导致诊断结果反转。Healthwatch England的一项审查发现，一个转录工具把检验结果“null demyelination”（无脱髓鞘）变成了暗示相反含义的内容，因为它漏掉了“null”一词。

来自申报文件

企业也在提交给美国证券交易委员会的文件中描述AI。以下是最近的四个例子：

• Paychex推出AI招聘工具。其季度业绩公告了WISE Hire，一个“代理式招聘解决方案”，并称“早期采用者结果令人信服”，但未给出数字。Paychex服务约84万家客户。
• Guidewire称客户可能用AI自建软件。这家保险软件制造商的年报警告，客户可能利用AI构建“可能减少或消除对我们解决方案需求”的工具。
• AMC将AI电影列为业务风险。这家影院连锁将“观众对使用AI技术制作的电影的接受度”列为其风险之一。
• People Inc将AI授权计为收入。这家出版商将“在大语言模型中使用”其内容列为授权版税来源之一，与Apple News+协议并列。

等等，什么？

扫描50万张人脸，零逮捕。英国交通警察在伦敦火车站测试实时人脸识别摄像头。摄像头扫描了超过50万张人脸。它们产生了一次警报，还是误匹配，最终没有逮捕任何人。

探索更多案例

我们的AI用例库收录了600多个此类有据可查的部署案例，可按行业、职能和状态筛选。我们的EDGAR AI雷达追踪企业向SEC披露的AI相关信息。

值得一看

AI从业者此刻正在转发的视频——在AI TV上精选呈现。

| Black Hat USA 2026 | 爆炸性新闻：OpenAI与Hugging Face事件 Black Hat | |
| 我如何忽悠大科技的AI定价算法省下8793美元 Chris the Producer |

本周投票

什么会让你信任自己公司内部的AI推广？

上周，165位读者投票：

你的大学应该把最坚定的AI红线划在哪里？

什么会让你信任自己公司内部的AI推广？

下次见，

Alexis

英文来源：

This Applied AI edition looks at what organizations actually did with AI over the past three weeks. Three patterns stand out: companies switching models to control cost, large deployments that are now in daily use, and a growing list of places where AI was restricted or removed.
Sponsor
Zero data retention should mean zero. In most AI plans, it means 30 days.
Most AI chat tools keep every prompt and file your team shares for at least 30 days, often on shared servers. Faradex gives each project its own isolated AI server. Nothing is stored or trained on. From $120 a month.
The AI bill became a management problem
Harvey's margins went negative, so it switched models
Harvey sells AI tools to law firms. It built them on models from OpenAI and Anthropic, which charge for every piece of text the model reads and writes. These pieces are called tokens.
Bloomberg reported that Harvey's gross margins fell from about 50% to minus 50% by June as its token usage grew 20 times. In August Harvey launched its own model, built on Moonshot's open-weight Kimi K3, and its margins turned positive again. Bloomberg named Abridge and Ramp among the other startups moving the same way.
An open-weight model is one a company can download and run on its own servers. It pays for computing power instead of paying another company for every token.
Open models now carry a large share of AT&T's AI work
The Financial Times reported that open models now handle 40% of AT&T's AI workloads.
Proprietary models from the large labs still take most of the money companies spend on AI. The change is that big buyers now treat the model as something they can swap when the price is wrong.
Anthropic ends discounts when customers hit their limit
Pricing is moving on the seller's side too. The Information reported that Anthropic ends its enterprise discounts, usually around 15%, once a customer uses up its contracted volume. The customer then has to renegotiate or pay list price.
For anyone running AI at work, the practical lesson is to measure cost per completed task. Teams that know that number can negotiate, or move work to a cheaper model when quality allows.
Where AI is now in daily use
McDonald's sets prices with an algorithm
Reuters reported that McDonald's uses machine learning to suggest an "optimal price" for menu items at nearly 14,000 US restaurants. The system analyzes millions of daily transactions.
In Fresno, one company-run store sells a Big Mac for $5.69 and another two miles away charges $6.89. Some franchisees told Reuters the tool has widened price gaps between nearby stores, and some said they were pressured to use it. McDonald's describes the tool as a recommendation, not a mandate.
A medical search tool used by roughly 40% of US doctors
OpenEvidence lets doctors ask clinical questions and get answers drawn from medical research. Business Insider reported that roughly 40% of US physicians now use it.
DraftKings used AI to target its promotions
The New York Times reported that DraftKings built a model in 2023 that scores customers by how much they are expected to lose after receiving a free bet. Executives credited AI-driven promotions with a roughly 13% improvement in sportsbook margins in 2025.
Former employees told the Times that similar work to flag problem gambling was stalled. DraftKings says its promotions target sustained, engaged users.
Coding agents reach large companies
Bloomberg reported that Cognition, maker of the Devin coding agent, is on track for $1 billion in annualized revenue, up from $492 million in May. Customers include Nvidia, Citigroup and Mercedes-Benz.
What got restricted or pulled back
Walmart banned AI-made store signs. A Walmart memo reported by Business Insider says "Stores should not display signing created by AI tools." Stores now order signs through approved corporate channels.
Three companies limited Anthropic's Fable over data retention. The Information reported that Nvidia, Palantir and Booz Allen restricted their use of the model after Anthropic began keeping usage logs for 30 days. Booz Allen barred employees from using it for proprietary cybersecurity software work.
Los Angeles schools blocked AI on student devices. LAUSD temporarily blocked generative AI on students' school-issued devices. Teachers kept access.
A Dutch optician chain stopped selling camera glasses. Hans Anders stopped selling Ray-Ban Meta glasses in the Netherlands and Belgium, citing the public and political debate over smart glasses.
An open-source project refused AI-written code. System76 now bars content generated by any LLM from issues and pull requests on its Pop!_OS projects, which include its COSMIC desktop. The rule covers code, comments and descriptions.
NHS AI scribes dropped a word that reversed a diagnosis. A Healthwatch England review found that a scribe turned a test result reading "null demyelination" into one suggesting the opposite, because it dropped the word "null."
From the filings
Companies also describe AI in documents they file with the US Securities and Exchange Commission. Four recent examples:

• Paychex launched an AI recruiting tool. Its quarterly results announce WISE Hire, an "agentic recruiting solution", and cite "compelling early adopter results" without giving numbers. Paychex serves about 840,000 customers.
• Guidewire says customers may build their own software with AI. The insurance software maker's annual report warns that customers could use AI to build tools "that could reduce or eliminate the need for our solutions."
• AMC lists AI films as a business risk. The cinema chain names "audience acceptance of movies made utilizing AI technology" among its risks.
• People Inc counts AI licensing as revenue. The publisher lists use of its content "in large-language models" as a source of licensing royalties, next to its Apple News+ deal.
  Wait, What?
  500,000 faces scanned, zero arrests. British Transport Police tested live facial recognition cameras at London railway stations. The cameras scanned more than 500,000 faces. They produced one alert, which was a false match, and led to no arrests.
  Explore more examples
  Our AI Use-Case Library holds more than 600 sourced deployments like these, which you can filter by industry, job function and status. Our EDGAR AI Radar tracks what companies tell the SEC about AI.
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  | Black Hat USA 2026 | The 'Breaking' News: The OpenAI–Hugging Face Incident Black Hat | |
  | How I Tricked Big Tech’s AI Pricing Algorithms to Save $8,793 Chris the Producer |
  This week's poll
  What would make you trust an AI rollout at your own company?
  Last week, 165 of you voted:
  Where should your university draw its firmest AI line?
  What would make you trust an AI rollout at your own company?
  Until next time,
  Alexis