内容来源：https://aibusiness.com/generative-ai/build-vs-buy-ai-agent-landscape-businesses

内容总结：

生成式AI应用初探：从“自建还是购买”到价值创造

随着生成式AI向智能体AI演进，企业面临的“自建还是购买”决策正变得愈发复杂。这一选择取决于企业规模、应用场景以及战略优先级等多重因素，而非简单的技术路线取舍。

四年前，OpenAI推出ChatGPT之初，企业急于跟进，直接购买OpenAI或Anthropic的大语言模型，或采用Meta的开源模型，曾是稳妥之选。然而，随着生成式AI日渐成熟和智能体AI广泛普及，这一决策的考量维度显著增多。

家得宝高级技术副总裁陈宁宇在拉斯维加斯举行的Ai4 2026大会上表示，客户体验是公司最核心的业务要素之一，“我们绝不会将这部分外包”。尽管如此，家得宝与Anthropic、OpenAI、谷歌和微软等主要AI技术供应商保持合作，并采用其技术。陈宁宇建议，企业在考虑自建与购买时，应更关注正确的应用场景，并对合作伙伴保持开放态度。“我们仍处于早期阶段，变数很多。看看OpenAI，曾经独占鳌头，如今已非如此。”他强调，企业需要建立抽象层以抵御技术快速迭代带来的不确定性。

咨询公司Future Propel首席执行官兼首席AI战略官桑蒂·拉梅什指出，选择自建的企业需审慎评估投入的时间和人力。她建议，若企业的核心竞争力并非技术，则更适合合作或购买；相反，若拥有大量数据、安全需求和定制化解决方案的需求，则应自主构建。“两种路径我都实践过。”她还提到，部分企业可能需要更大灵活性，可先通过测试、学习或小范围试用供应商方案，以便必要时退出合作。

Cardinal Group Companies数据与创新副总裁雷切尔·伊瓦拉则提醒，购买可能带来运营债务。该公司采用混合模式，内部开发了名为Stan的AI助手。伊瓦拉认为，企业需思考“一个协调一致的系统是什么样子”，这不仅关乎技术生态，更涉及组织运作方式。对于数据能带来切实优势的企业，自建或许是更佳选择；尤其当需求规模较小、偏向内部应用且能产生显著内部效益时，自建对缺乏庞大工程团队的非技术型企业更具吸引力。

Smartsheet专业服务赋能总监马库斯·麦凯-弗莱施则强调基础设施的重要性。“真正的问题是，你是否具备合适的自有构建基础设施？”他提出，若采用公民开发者模式由业务人员构建，需明确相应的防护措施。企业还需确保拥有强大的托管基础设施，并能维护系统在故障或变化时的正常运作。

在模型选择层面，人力资源科技公司Phenom首席执行官马赫·巴伊雷迪表示，企业常需根据不同应用场景切换模型，如某些模型适合人力资源，另一些则适合财务。目前许多企业尚未掌握这种行业层面的切换能力，“模型应在运行中灵活切换”。

AI营养应用Just a Bite Better创始人史蒂夫·托伊则选择在应用内自主构建多智能体系统，而非购买现成方案。该应用整合多个智能体协同工作，通过自建的“模型花园”访问各类API，而非依赖单一供应商。托伊认为，购买前必须明确要解决的问题，并理解每个业务都可拆解为不同部分。“购买是为了解决业务中的特定问题。”同时，自建有助于成本控制——他设置了每月各模型Token数或成本上限，以防意外高额账单，“否则很容易一觉醒来面对5万美元的账单”。

总体而言，无论是选择自建还是购买，企业应回归核心问题：明确价值主张、评估数据与基础设施条件，并以灵活开放的姿态应对快速演进的AI技术生态。

中文翻译：

由谷歌云赞助

选择你的首个生成式AI用例

要开始使用生成式AI，首先应聚焦于那些能够改善人类与信息交互体验的领域。

随着生成式AI演进为智能体AI，自行构建还是购买现成方案的决策变得更加复杂，取决于诸多因素，包括企业规模、用例场景以及战略优先级。

面对生成式AI，以及如今的智能体AI，企业和中小型企业长期以来一直面临着一个与它们如何推进技术应用以及希望达到何种成本效益密切相关的重要问题：是自行构建，还是从供应商处购买。

四年前，当OpenAI首次推出ChatGPT、各公司开始尝试生成式AI时，答案似乎很简单。对于那些不想落后于人的企业来说，从OpenAI或Anthropic购买大语言模型，或者尝试使用Meta的开源模型，看起来都是稳妥的选择。

然而，随着生成式AI的成熟和智能体AI的广泛接受，自行构建与购买之间的两难选择变得更加微妙，主要取决于多个因素，例如企业规模、应用场景，以及企业或中小型企业认为的自己的“护城河”——即将其与竞争对手区分开来并保护自身利益的独特价值主张。

对于家得宝这家家居装修巨头而言，客户体验被视为业务最核心的要素之一，该零售商的技术高级副总裁陈宁宇在拉斯维加斯举行的Ai4 2026大会上的一场会议中如此表示。

“我们绝不会把这一部分外包给其他人，”陈说。

然而，家得宝与Anthropic、OpenAI、谷歌和微软等主要AI技术供应商保持着合作关系，并使用它们的技术。

尽管如此，陈建议那些考虑自行构建与购买问题的企业，应更多关注正确的用例，并在选择合作伙伴和供应商时保持灵活。

“我会让合作关系保持非常开放的态度，”他在接受AI Business采访时表示。“我们仍处于早期阶段，很多事情都可能发生。看看OpenAI，它曾经占据主导地位；现在不再是了。”

“你必须有一个抽象层来保护自己免受这些变化的影响，”他补充道。

除此之外，希望自行构建的企业应考虑计划投入多少时间和人力来将生成式或智能体AI技术整合到自己的平台中，咨询公司Future Propel的首席执行官兼首席AI战略官桑蒂·拉梅什表示。

作为企业领导团队AI转型的顾问，拉梅什职业生涯早期曾在好时公司和意大利巧克力公司费列罗等组织主导AI转型工作。

她表示，自行构建与购买的决定还取决于企业的核心竞争力。

“如果你的核心竞争力不是技术公司，那你最好选择合作和购买，”她在采访中说。“如果你有密集的数据、安全性要求、海量数据以及你想要的定制化解决方案，那你就必须自己构建。这两种方式我都做过。”

然而，一些企业和中小型企业可能还需要更大的灵活性，在这种情况下，组织最好先测试、学习，然后购买平台或与供应商结盟，这样如果必要的话，他们可以在试用了测试版之后退出合作关系，她继续说道。

购买可能会导致运营负债，Cardinal Group Companies的数据与创新副总裁雷切尔·伊巴拉表示。该公司是一家物业管理与房地产组织。

虽然Cardinal Group采取了混合方式，拥有一个名为Stan的内部AI代理，但伊巴拉表示，在内部开发和外部寻找之间做选择的企业需要思考“一个完整的系统应该是什么样子。”

“这不只是技术的生态系统，”她在大会的一场演讲中表示。“而是组织运作方式的生态系统。”

伊巴拉说，对于数据能提供切实优势的企业来说，自行构建可能更好。当企业有零散需求时，自行构建也是一个好主意。

“也就是说，你只需要某个相对较小的东西的一部分，而且它更偏向内部使用，能为你带来很多内部优势，”她在采访中说。“我认为这更适合那些不一定有资源去组建庞大工程团队的非技术公司。”

还有基础设施的问题，Smartsheet的专业服务赋能总监马库斯·麦凯-弗莱施表示。

“对我来说真正的问题是，你是否觉得自己拥有合适的基础设施来自己构建？”他在采访中说。“这是一个公民开发项目吗——由业务人员来构建？如果是，你在周围设置了什么样的护栏？”

企业还需要确保拥有强大的托管基础设施，并且能够在AI代理或系统出现故障或发生变化时进行维护。

此外，构建AI代理或系统需要应对为每个用例选择最佳AI模型的诸多细微差别。

Phenom是一家HR技术和应用AI公司，构建软件帮助组织招聘候选人和发展技能。

该供应商使用多种模型，包括来自法国、北美和中国供应商的开源模型。据首席执行官马赫·巴伊雷迪称，许多构建AI代理或包含AI代理的系统的组织，有时需要根据合适的用例切换模型。例如，有些模型适合HR应用，其他的适合财务应用。

公司们“并不理解如何在行业层面——而非公司层面——实现切换的细微差别，”巴伊雷迪说。“你的模型会在运行中实时切换。”

史蒂夫·托伊是AI营养应用Just a Bite Better的创始人兼缔造者，他选择在应用内自行构建AI代理，而不是购买系统来部署它们。

该应用使用多个代理协同工作，为消费者提供营养概况。

“我们不使用任何单一供应商或单一模型，”托伊在接受采访时说。“我们构建了自己的模型花园，如果你愿意这么说的话——其实就是一段代码，用来访问所有这些服务的API。”

对托伊来说，选择自行构建而非购买或使用谷歌、AWS或微软的模型花园，是出于他的技术知识和他特定的业务应用场景。

“我不需要处理大量资金的往来，”他说，并补充道消费者通过Apple、Google或Stripe支付应用费用。此外，消费者可以删除任何个人身份信息，而且他们的信息经过编码处理，不会指向特定的个人。

“我们在很多方面风险都很低，而IBM和PayPal必须关心不同的事情，”托伊说。

他补充说，对于仍在探索购买还是构建的企业来说，最好先弄清楚自己需要购买什么，尤其是在一个充斥着AI供应商、产品和服务的市场中。

“要弄清楚如何购买，你需要知道自己试图解决什么问题，并且记住每个企业都可以分解为各个部分，”他说。“当你购买时，你是为了解决问题而买，为了解决你业务的一部分而买。”

不过，托伊补充说，自行构建有助于控制成本。对于Just a Bite Better，他实施了一个系统，为每个模型每月可使用的令牌数量或成本设定了硬性上限。

“这就是防止灾难的方法，这非常重要，因为真的很容易早上一醒来就发现一张5万美元的账单，”托伊说。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
As generative AI evolves into agentic AI, the build-or-buy decision becomes more complex and depends on numerous factors, including business size, use cases, and strategic priorities.
With generative AI, and now agentic AI, enterprises and small and mid-sized businesses have long faced an important question that correlates with how they move forward with technology and also how cost-effective they aim to be: building in-house or buying from a vendor.
Four years ago, when OpenAI first introduced ChatGPT and companies began experimenting with generative AI, the answer appeared easy. For those who did not want to be left behind, buying large language models from OpenAI or Anthropic or trying to use open source models from Meta looked like a safe bet.
However, with the maturity of generative AI and the wide acceptance of agentic AI, the build versus buy dilemma is more nuanced, depending mainly on multiple factors such as the business size, application and also what an enterprise or SMB considers to be its “moat,” or defining value proposition that separates and protects it from the competition.
For Home Depot, the home improvement giant sees CX as one of the most important core factors of the business, said Ningyu Chen, the retailer’s senior vice president of technology, during a session at the Ai4 2026 conference in Las Vegas.
“We would never outsource that piece to others,” Chen said.
However, Home Depot maintains partnerships with key AI technology vendors such as Anthropic, OpenAI, Google and Microsoft -- and uses their technology.
Despite this, Chen advises businesses considering the build-versus-buy question to focus more on the right use case and be flexible about whom they partner with and buy from.
“I will leave the partnership very open,” he said in an interview with AI Business. “We are still early, and there are a lot of things that can happen. Look at OpenAI, it was dominant; now it is not anymore.”
“You have to have an abstraction layer to prevent yourself from these changes,” he added.
Beyond that, businesses looking to build should consider the time and people they plan to invest in building generative or agentic AI technology into their platform, said Santhi Ramesh, CEO and chief AI strategy officer of Future Propel, an advisory firm.
As an adviser to enterprise leadership teams for AI transformation, Ramesh spent her early career leading AI transformation at organizations such as The Hershey Company and Ferrero, the Italian chocolate company.
She said the build-versus-buy decision also depends on a business's core competency.
“If your core competency is not a technology company, you are better off partnering and buying,” she said in an interview. “If you have intense data, security, massive amounts of data and a custom solution that you want, then you must build on your own. I have done both.”
However, some enterprises and SMBs may also need more flexibility, in which case it is better for an organization to test, learn and buy a platform or ally with a vendor so that, if necessary, they can get out of the partnership, after, for example, trying it in beta, she continued.
Buying can lead to operational debt, said Rachel Ibarra, vice president of data and innovation at Cardinal Group Companies, a property management and real estate organization.
While Cardinal Group takes a hybrid approach, with an in-house AI agent called Stan Ibarra said that enterprises choosing between developing in-house and looking outside the company will need to consider “what a cohesive system looks like.”
“It’s not just the ecosystem of technology,” she said during a talk at the conference. “It’s the ecosystem of how the organization works.”
For businesses whose data provides a tangible advantage, it may be better to build, Ibarra said. It could also be a good idea for organizations to build when they have a fractional need.
“So, you only need a part of something that’s relatively small [and] where it’s more internal facing and gives you a lot of internal advantage,” she said in an interview. “That’s what I would say are better built for nontechnical companies that don’t necessarily have the resources to go and hire a huge engineering team.”
There is also the issue of infrastructure, said Markus McKay-Fleisch, professional services enablement director at Smartsheet.
“The real question for me is, do you feel like you have the right infrastructure to build yourself?” he said during an interview. “Is it a citizen development program where people in business are building, and if so, what are the guardrails you put around that?”
Businesses also need to ensure they have strong hosting infrastructure and can maintain the AI agent or system if it breaks or changes.
Moreover, building an AI agent or system requires navigating the nuances of choosing the best AI model for each use case.
Phenom is an HR technology and applied AI company that builds software to help organizations hire candidates and develop skills.
The vendor uses various models, including open models from vendors based in France, North America and China. According to CEO Mahe Bayireddi, many organizations building AI agents or systems with AI agents sometimes need to switch models for appropriate use case. For example, some models are good for HR applications, and others for finance applications.
Companies are “not understanding the nuances of how to be able to switch, not at your company level, but at an industry level,” Bayireddi said. “Your models will switch on the fly.”
Steve Toy, founder and creator of Just a Bite Better, an AI nutrition app, has chosen to build AI agents within his app rather than buy a system to implement them.
The app uses multiple agents working together to provide consumers with an overview of their nutrition.
“We don’t use any one provider or any one model,” Toy said in an interview. “We built our own model garden, if you will, which is merely just here is the code that accesses the APIs of all these things.”
For Toy, the decision to build rather than buy or use a model garden from Google, AWS or Microsoft is driven by his technical knowledge and his specific business application.
“I’m not dealing with passing lots of money back and forth,” he said, adding that consumers pay for the app through Apple, Google or Stripe. Moreover, consumers can delete any personally identifiable information, and their information is encoded so it does not specify a specific person.
“We’re low stakes in many respects, whereas IBM and PayPal have to care about different things,” Toy said.
He added that for businesses still exploring whether to buy or build, it is best to know what they need to buy, especially in a market saturated with AI vendors, products and services.
“To figure out how to buy, you need to know what problem you are trying to solve and remember that every business is decomposed into pieces,” he said. “When you’re buying, you’re buying to solve a problem, a part of your business.”
Building, though, can help with cost, Toy added. For Just a Bite Better, he has implemented a system that sets a hard cap on the number of tokens or the cost each model can use monthly.
“That’s how you can prevent disaster, and that’s really important because it’s really easy to wake up in the morning and find a $50,000 bill,” Toy said.