内容来源：https://www.geekwire.com/2026/the-biggest-unresolved-question-in-ai-right-now-and-more-takeaways-from-madronas-ia40-summit/

内容总结：

本周，微软、亚马逊、Anthropic和Stripe等公司的高管在Madrona举办的IA40峰会上就AI发展方向达成了诸多共识，但一个核心问题始终悬而未决：当AI智能体介入交易中间时，客户关系以及由此产生的数据究竟归谁所有？这一议题在整天的讨论中被反复提及。微软副总裁查尔斯·拉曼纳预测，未来大多数商业软件将由AI助手代表用户使用，软件制造商对定价的掌控力将下降。Town公司CEO让-德尼·格雷兹进一步指出，AI已能像人一样操作浏览器或计算机，无需软件商提供API，“一切都将变成薄应用”。Stripe的AI首席营收官玛雅·何塞巴奇维利表示，AI智能体在Stripe上的商业交易近六周急剧上升，但商家将失去追加销售和展示广告的机会，这种冲突已在亚马逊封锁Meta的Muse助手的争端中实时上演。

在数据归属方面，Anthropic首席技术官拉胡尔·帕蒂尔未直接回答谁拥有AI智能体工作记录的问题，仅表示各公司会利用一切可用数据来改进智能体。AWS副总裁斯瓦米·西瓦苏布拉马尼亚姆介绍了名为AWS Context的服务，用于为企业数据建立关系映射供智能体使用，并预计这类服务将成为未来20年的基础构件。多数企业尚未跟上技术步伐，Anthropic编写的代码量是18个月前的200倍，但几乎无客户达到同等增益。高盛的阿尔查娜·维穆拉帕利指出，瓶颈不是AI，而是人。麦肯锡高级合伙人拉里·哈马莱宁称约40%企业表示AI提升了利润，但仅6%认为增幅显著。

企业在是否依赖单一AI供应商上存在分歧。Anthropic的帕蒂尔认为大企业为保留切换供应商的能力而耗费过多精力，最终只能按最低共同标准构建。Noeri联合CEO卡洛斯·格斯特林则主张智能不应由一两家公司控制，每家企业都应能构建并拥有自己的AI系统。Factory联合创始人埃诺·雷耶斯表示许多企业看不到不向某一AI实验室让渡控制权就能构建未来的路径。亚马逊与多家AI实验室合作，既是Anthropic的主要投资者和合作伙伴，也与OpenAI推出了托管智能体产品，同时销售自有的AgentCore。在信任方面，OpenAI董事会安全委员会主席齐科·科尔特表示控制AI系统的能力必须与能力同步发展，可能意味着要放慢开发速度。AWS的西瓦苏布拉马尼亚姆建议将AI模型与独立规则检查系统配对，以兼顾创造力与合规。

资金高度集中于少数公司。今年IA40上榜企业自成立以来共融资4100亿美元，其中OpenAI、Anthropic和Databricks占92%。Anthropic在截至8月15日的12个月内融资1430亿美元，已保密提交IPO申请，寻求约2万亿美元估值，最早可能于11月中旬上市。亚马逊、Alphabet、微软、Meta和苹果今年资本支出预计增长74%至约7420亿美元。微软内部重度AI用户开发者人均年AI使用支出有望超过100万美元。Madrona闭幕幻灯片指出，对多数使用AI的企业而言回报尚早，但从最早采用者的经验看，回报似乎不可避免。

中文翻译：

本周，来自微软、亚马逊、Anthropic和Stripe等公司的高管们在AI走向何方的问题上达成了许多共识。但有一个大问题悬而未决：当AI智能体介入其中时，谁维系与客户的关系，以及由此产生的数据归谁所有？

这个问题在Madrona的IA40峰会上反复被提及。该峰会是这家风险投资公司每年举办的AI初创企业、投资者和科技高管聚会，周三在西雅图市中心水岸的四季酒店举行。

正如Madrona董事总经理Matt McIlwain在闭幕致辞中所说，谁有权使用人们与AI系统互动所产生的数据这一问题，在当天“一次又一次地被提起”。

GeekWire全程参加了当天的活动，并总结出以下要点，借助与Anthropic的Claude及其他AI工具的大量来回交流，减少我们自身的人类瓶颈（见下文）。

AI智能体正在介入企业与客户之间：微软执行副总裁Charles Lamanna负责Microsoft 365及Copilot背后的平台，他预测大多数商业软件最终将由AI助手代替人来使用，软件制造商设定价格的权力将因此减弱。

Town的CEO Jean-Denis Greze走得更远。Town是一家为工作场景打造AI助手的初创公司。他说，自大约7月以来，AI已经能够以合理成本几乎像人一样操作浏览器或计算机。这意味着AI助手可以像人一样通过应用的用户界面完成操作，无需软件制造商提供API，他说。

“一切都将变成薄应用，因为AI使用计算机的能力越强，应用作为软件单元的重要性就越低，”Greze说，这与Lamanna当天上午用来形容人们只短暂打开一下的软件所用的术语如出一辙。

同样的转变正在零售业发生。Stripe负责AI业务的首席营收官Maia Josebachvili表示，Stripe上由AI智能体完成的交易在大约八九个月里基本持平，然后在过去六周急剧上升。

她说，许多商家靠让购物者在结账时追加订单和向他们展示广告来赚钱，而当智能体代为购买时，他们就失去了这些机会。Josebachvili说，长远来看这行不通。

这种冲突正在两家科技巨头之间一场备受瞩目的争端中实时上演。亚马逊上月阻止了Meta的Muse助手在其网站上购物。

问题的数据层面在Anthropic首席技术官Rahul Patil参与的一场讨论中被提及。主持人Raphaëlle d'Ornano问及AI智能体工作记录（包括其错误和纠正）归谁所有，是否属于客户。她说她从未得到过明确答案。

Patil没有直接回答。他说，每一家提供智能体运行软件的公司都会努力改进它们，并“会利用一切可用的数据来让它变得更好”。

“如果智能体确实在进步，这对整个生态系统是好事，”他说。

亚马逊云科技负责代理式AI和新兴技术的副总裁Swami Sivasubramanian表示，让AI智能体发挥作用最被低估的环节之一，是为它们提供完成任务所需的正确企业数据。

AWS在6月发布了一项为此目的服务的产品，名为AWS Context，它为企业数据中的关系建立映射，供智能体使用。Sivasubramanian说，该服务旨在跨平台和跨云服务商运行，他预计这类服务将成为未来20年的基础构建模块，就像二十年前的云存储和数据库一样。

大多数公司没有跟上这项技术的步伐：Patil说，Anthropic现在的代码编写量是18个月前的约200倍。他说，一些客户的产出翻了两倍或三倍，但几乎没有哪家能取得Anthropic那样的增幅。

“瓶颈实际上不是AI。瓶颈是人，”高盛AI产品管理全球主管Archana Vemulapalli说。她说，公司的角色和流程都是在AI出现之前建立的。

Sivasubramanian说，亚马逊内部团队向他展示了他们在两三周内构建出的令人印象深刻的AI智能体。他说，当他问这些智能体何时可以推出时，几乎每个团队都告诉他，还需要解决安全、身份认证和监控问题。

麦肯锡高级合伙人Lari Hamalainen说，约40%的公司表示AI增加了它们的利润，但只有6%称增幅显著。

McIlwain说，尽管AI的能力越来越强，要弥合“目标、能力与我们所有人所期望实现的东西”之间的差距，仍有大量工作要做。

企业在是否依赖单一AI供应商的问题上存在分歧：Anthropic的Patil特别指出，大公司正在花费过多精力来保留在AI供应商之间切换的能力。他说，为了保持这一选项，它们最终按照最低共同标准来构建，将自己限制在每个AI模型都具备的功能上。这种努力还让它们的工程师脱离了使自身公司与众不同的工作。

AI初创公司Noeri的联合CEO Carlos Guestrin在同一场讨论中提出了另一种观点。他说，智能不应由拥有一两个模型的公司来控制，每家公司都应该能够构建并拥有自己的AI系统。

Guestrin是斯坦福大学计算机科学教授，曾任华盛顿大学教授，他创立了西雅图机器学习初创公司Turi，该公司于2016年被苹果收购。

Factory的联合创始人兼首席技术官Eno Reyes在随后的一场讨论中说，许多企业看不到在不将控制权让渡给某一家AI实验室的情况下构建未来的途径。Factory是一家打造AI编程工具的初创公司。前GitHub CEO、现领导初创公司Entire的Thomas Dohmke说，开发者总是想要选择权。

亚马逊与不止一家AI实验室合作。它是Anthropic的重要投资者和合作伙伴，亚马逊商业与企业发展副总裁Dan Grossman在一场投资者讨论中指出，亚马逊当周与OpenAI合作推出了一款托管智能体产品。

亚马逊还销售自己的AI智能体运行产品AgentCore，该产品与Anthropic的Managed Agents存在重叠。

企业尚不信任AI做重要决策：卡内基梅隆大学教授、OpenAI董事会安全与安保委员会主席Zico Kolter说，控制AI系统的能力必须与其能力同步发展，这可能意味着要以低于可能速度的节奏来开发它们。

Sivasubramanian说，高管们担心AI智能体是否会始终遵守公司的规章制度。他说，公司可以通过将AI模型与独立的系统配对来保留其创造力，这些系统会对照公司规则检查模型的工作并捕捉违规行为。

资金集中在少数几家公司：根据Madrona的数据，今年IA40榜单上的公司自成立以来已融资4100亿美元。OpenAI、Anthropic和Databricks占了其中的92%。

Anthropic在截至8月15日的12个月中，其1610亿美元总融资额中有1430亿是在此期间筹集的，该数字包括债务。据路透社本周报道，该公司已保密提交IPO申请，寻求约2万亿美元的估值。彭博社报道称，它最早可能在11月中旬上市。

根据Madrona展示的PitchBook数据，亚马逊、Alphabet、微软、Meta和苹果的资本支出预计今年将增长74%，达到约7420亿美元。

Lamanna说，微软软件开发者中AI使用量最大的一批人，每人预计每年将在AI使用上花费超过100万美元，即使算上公司的内部折扣。

Madrona的闭幕幻灯片指出，对大多数使用AI的公司来说，AI的回报仍处于早期阶段，但从最早采用者看到的情况来看，回报似乎不可避免。

英文来源：

Executives from companies such as Microsoft, Amazon, Anthropic and Stripe agreed on a lot this week about where AI is headed. But one big question was left unresolved: who keeps the relationship with the customer, and the data that comes from it, once an AI agent is in the middle?
That question was a recurring theme at Madrona’s IA40 Summit, the venture capital firm’s annual gathering of AI startups, investors and tech executives, held Wednesday at the Four Seasons Hotel on the downtown Seattle waterfront.
As Madrona Managing Director Matt McIlwain said in his closing remarks, the question of who gets to use the data that comes from people’s engagement with AI systems came up “over and over and over again” during the day.
GeekWire attended the entire day and came up with these takeaways, reducing our own human bottleneck (see below) with lots of back-and-forth with Anthropic’s Claude and other AI tools.
AI agents are coming between companies and their customers: Charles Lamanna, the Microsoft executive vice president who oversees Microsoft 365 and the platform behind Copilot, predicted most business software will end up being used by AI assistants on a person’s behalf, leaving software makers with less power to set prices.
Jean-Denis Greze, CEO of Town, a startup that makes an AI assistant for work, went further. Since about July, he said, AI has been able to operate a browser or a computer almost as well as a person, at a reasonable cost. That means an AI assistant can work through an app’s user interface the way a person does, without needing an API from the software maker, he said.
“Everything is going to become a thin app because the better the AI gets at using the computer, the less the app matters as a unit of software,” Greze said, echoing the term Lamanna had used that morning for software that people open only briefly.
The same shift is under way in retail. Maia Josebachvili, Stripe’s chief revenue officer for AI, said commerce by AI agents on Stripe was roughly flat for eight or nine months and then rose sharply in the past six weeks.
Many merchants make their money by getting shoppers to add to their orders at checkout and by showing them ads, she said, and they lose those chances when an agent does the buying. That won’t work over the long term, Josebachvili said.
That conflict is playing out in real time in a high-profile dispute between two tech giants. Amazon blocked Meta’s Muse assistant from shopping on its site last month.
The data side of the question came up in a session with Anthropic Chief Technology Officer Rahul Patil. Moderator Raphaëlle d’Ornano asked who owns the record of an AI agent’s work, including its mistakes and corrections, and whether it belongs to the customer. She said she has never gotten a clear answer.
Patil didn’t answer directly. He said each company that supplies the software for running agents will work to improve them, and “will use every data that’s available to them to make it better.”
“It’s good for the ecosystem if the agents actually improve,” he said.
Swami Sivasubramanian, Amazon Web Services’ vice president of Agentic AI and Emerging Technologies, said one of the least appreciated parts of making AI agents work is getting them the right company data for the task.
AWS in June announced a service for that purpose, called AWS Context, which maps the relationships in a company’s data for agents to use. Sivasubramanian said it’s meant to work across platforms and cloud providers, and he expects this kind of service to be a basic building block for the next 20 years, as cloud storage and databases were two decades ago.
Most companies aren’t keeping up with the technology: Patil said Anthropic writes about 200 times as much code as it did 18 months ago. Some customers have doubled or tripled their output, he said, but almost none have seen gains like Anthropic’s.
“The bottleneck is actually not AI. The bottleneck is human,” said Archana Vemulapalli, global head of AI product management at Goldman Sachs. A company’s roles and processes were built before AI, she said.
Sivasubramanian said teams inside Amazon were showing him impressive AI agents they had built in two or three weeks. When he asked when the agents could be rolled out, he said, almost every team told him it still had to work out security, identity and monitoring.
McKinsey senior partner Lari Hamalainen said about 40% of companies say AI has increased their profits, but only 6% describe the increase as substantial.
Even as AI becomes more and more capable, there’s still lots of work to do to close the gap between “the goals, the capabilities, and what we all are aspiring to do,” McIlwain said.
Companies are split on depending on a single AI provider: Anthropic’s Patil, notably, argued that large companies are spending too much effort preserving the ability to switch between AI providers. To keep that option open, he said, they end up building for the least common denominator, limiting themselves to the features every AI model shares. That effort also takes their engineers away from the work that sets their own company apart.
Carlos Guestrin, co-CEO of the AI startup Noeri, offered an alternative take on the same panel. Intelligence shouldn’t be controlled by one or two companies that own the models, he said, and every company should be able to build and own its own AI systems.
Guestrin is a Stanford computer science professor and former University of Washington professor who founded the Seattle machine learning startup Turi, which Apple acquired in 2016.
Eno Reyes, co-founder and chief technology officer of Factory, a startup that makes AI coding tools, said on a later panel that many businesses don’t see a way to build their future without ceding control to one AI lab. Thomas Dohmke, the former GitHub CEO who now leads the startup Entire, said developers always want choice.
Amazon works with more than one AI lab. It is a major Anthropic investor and partner, and Dan Grossman, Amazon’s vice president of business and corporate development, pointed out on an investor panel that Amazon had launched a managed agents product with OpenAI that week.
Amazon also sells its own product for running AI agents, AgentCore, which overlaps with Anthropic’s Managed Agents.
Companies don’t yet trust AI with important decisions: Zico Kolter, the Carnegie Mellon professor who chairs the safety and security committee of OpenAI’s board, said the ability to control AI systems has to keep pace with their capabilities, which might mean developing them more slowly than is possible.
Sivasubramanian said executives worry about whether an AI agent will always follow their company’s rules and regulations. Companies can keep the creativity of AI models, he said, by pairing them with separate systems that check a model’s work against the company’s rules and catch violations.
The money is concentrated in a few companies: The companies on this year’s IA40 list have raised $410 billion since they were founded, according to Madrona’s data. OpenAI, Anthropic and Databricks account for 92% of it.
Anthropic raised $143 billion of its $161 billion total in the 12 months ending Aug. 15, a figure that includes debt. The company has filed confidentially for an initial public offering and is seeking a valuation of about $2 trillion, Reuters reported this week. Bloomberg reported that it could go public as early as mid-November.
Capital spending by Amazon, Alphabet, Microsoft, Meta and Apple is projected to rise 74% this year to about $742 billion, according to PitchBook figures Madrona presented.
Lamanna said Microsoft’s heaviest AI users among its software developers are each on track to spend more than $1 million a year on AI usage, even with the company’s internal discount.
Madrona’s closing slide said returns on AI are still early for most of the companies using it but look inevitable based on what the earliest adopters are seeing.