内容来源：https://aibusiness.com/generative-ai/how-enterprises-catch-up-with-rapid-pace-of-ai-advances

内容总结：

企业应用生成式AI：从概念验证到核心流程重塑

近日在拉斯维加斯举行的Ai4 2026大会上，Dataiku人工智能与平台高级副总裁Jed Dougherty就企业在应用生成式智能体AI过程中面临的挑战与机遇接受了采访。随着OpenClaw开源个人智能体现象的出现、Anthropic发布领域自适应Claude Cowork及强大的Mythos模型，以及英伟达CEO黄仁勋呼吁企业拥抱“OpenClaw战略”，企业界对生成式AI技术的态度正从怀疑转向积极尝试。

人类监督仍不可或缺

针对企业是否应完全接受AI智能体自主性的问题，Dougherty明确表示：“人类指导仍然必要。目前取得最大成效的企业，依然是在人机协同模式下运行的。”他认为，每家企业都需要制定策略，明确哪些流程可以完全自动化，哪些环节仍需人工介入。识别组织内部哪些政策和流程的子集可以被自动化，是当前的重要课题。

前沿部署工程师的关键作用

Dougherty强调，Dataiku在组建前沿部署工程师团队方面投入了大量时间和资金。“这些工程师的价值在于，他们能帮助客户跨越那些看似简单、但对初次使用者来说却构成重大障碍的技术门槛。”他举例说，在帮助客户构建网站时，许多日常操作中的小技术问题会极大拖慢部署进度，而有经验的工程师能有效解决这些瓶颈。

专家支持仍是必要投入

对于智能体AI是否需要专家支持的问题，Dougherty认为，AI是全新的技术领域。四年前，人们对机器学习自助服务已经相当熟悉，但面对全新的智能体能力，厂商需要花时间帮助用户“翻过最初的山丘”。他预测，一两年后随着更多人接触并掌握这项技术，对专家支持的需求将会下降，但目前阶段，构建结构化的智能体来复制业务流程仍然是相对复杂的工作，专家介入依然很有价值。

核心流程是最佳应用场景

Dougherty指出，企业选择生成式AI应用场景的前提是深刻了解自身业务。“任何企业的最佳应用场景都是其核心流程。”他以太理赔为例，从理赔受理、调查、赔付决策到审计和法律通知，保险公司若能清晰描述这些现有流程，就能明确哪些环节适合用智能体AI来复制或增强。

平台选择：灵活性与可替换性优先

面对众多供应商，Dougherty建议企业应具备灵活选择底层工具的能力。“有时Snowflake是最佳数据库，有时则不是；有时OpenAI是最优选择，但涉及敏感数据或成本控制时，本地部署模型可能更合适。”他强调，一个集中、中立的编排层，让企业能够在安全、受控、合规的前提下自由组合和切换底层工具，是构建AI系统的基石。

成本飙升：2027年或现“账单冲击”

Dougherty将当前AI使用成本比作云计算初期的“价格惊愕”——企业将所有业务切换到AWS后，第二年便收到巨额账单。“2027年这种情况将在生成式AI领域重演。”他坦言，自己高强度工作时一天消耗1000美元的token并不罕见。控制成本的关键在于：不应让单一智能体承担所有任务，而应将其拆分为尽可能多的小步骤；同时密切关注市场动态，及时调整底层模型选择，这能在定价上带来显著差异。

中国开源模型：性价比优势明显

对于阿里巴巴、DeepSeek等中国供应商提供的开源模型，Dougherty表示赞赏：“它们非常出色。我们看到大量第三方厂商以极具竞争力的价格提供这些中国模型和其他开源模型。”他认为，企业完全可以通过引入第三方开源模型来节省大量成本，且其能力已足够满足需求。

Dougherty最后总结道，随着智能体在企业组织中不断普及，管理将变得至关重要。“对话式AI和利用AI开发应用已经成为趋势，所有人都会参与其中。我们需要在实践中摸索出什么有效、什么无效。”

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互动体验的领域。

要在不断演进的AI技术中取得成功，企业需要深刻理解自身的业务流程，并在所使用的模型和智能体中保持灵活性。

过去几年，生成式和智能体AI在企业中得到了广泛应用。但现在，正如2022年OpenAI发布Chat-GPT开启AI热潮之初那样，技术的飞速发展似乎并未与采用率同步。当前对AI使用成本的关注，意味着企业必须认真审视自身应用生成式AI的方式，确保从中获得价值。

然而，近期随着企业逐渐淡化对生成式和智能体AI技术能否帮助它们的怀疑态度，一种转变已经开始。近来，面对OpenClaw开源个人智能体现象、Anthropic发布的可适应领域的Claude Cowork和强大的Mythos模型，以及英伟达CEO黄仁勋呼吁企业拥抱“OpenClaw战略”，企业感受到了实施该技术的迫切压力。

在本月早些时候于拉斯维加斯举行的Ai4 2026大会上，企业AI和机器学习平台供应商Dataiku的AI与平台高级副总裁Jed Dougherty接受了采访，讨论了企业在智能体自主性和选择合适模型或智能体方面面临的一些障碍。对于Dougherty来说，无论企业选择哪种AI品牌，都必须加以管理，使其为组织服务。

鉴于OpenClaw的成功，企业现在是否应完全接受AI智能体的自主性，还是仍然需要更多的人工参与？

Jed Dougherty：仍然需要人工指导。人们在使用智能体时看到的最大收益和成功，仍然来自那些保留人工参与的企业。每家公司都需要一个策略，来识别哪些内容可以完全过渡到全自主系统。识别组织内部哪些政策和流程可以被自动化，或哪些子集可以自动化，这一点很重要。

FDE（即前部署工程师）在帮助企业采用AI技术方面发挥什么作用？

Dougherty：我们投入了大量时间和资金来招聘并组建一支强大的前部署工程团队，协助我们的客户。这非常重要。例如，我在帮助一个客户时，我们试图搭建一些相对简单的网站。我没有预料到那些我因为每天做而完全视为理所当然的小技术障碍，会让人寸步难行，或极大地拖慢网站首次部署的时间点。拥有一位懂得如何处理简单基础设施事务的前部署工程师仍然很有价值。

让智能体AI发挥作用需要多少专家？您是否认为，在生成式AI广泛应用之初，供应商试图将其包装成不需要专家的工具，这种宣传是错误的？

Dougherty：我不这么认为。AI是一项全新的技术。四年前，人们对机器学习的自助服务已经相当熟练，因为他们已经做了十年。现在，随着新的智能体能力四处扩展，供应商需要花点时间帮助人们跨过最初的这道坎。一两年后，随着更多人接触并习惯了这项技术，我认为就不需要同样程度的现场协助了。但眼下，当我试图向一个从未接触过智能体或从未构建过智能体的人描述时，中间有一位专家确实很有帮助。拥有结构化智能体来复制业务流程的很大一部分，这种业务转型能力相对复杂。

最能推动企业转型的智能体AI最佳用例是什么？

Dougherty：企业必须先了解自己，才能确定最佳用例。任何企业中智能体AI的最佳用例就是其核心流程。

以保险为例。一份保险索赔进来，对索赔进行一些处理，需要对索赔进行一些调查，然后决定是否赔付该保险索赔。可能还有一些审计，也许他们会告知法务团队。因此，保险业务的核心是决定如何以及是否赔付保险索赔，以及是否再次与该客户做生意。所以，一家保险公司现在能越清晰地描述没有生成式AI时他们如何做这件事，就越能理解保险索赔赔付流程中的哪些子集可以被智能体AI复制或增强。

不过，对企业来说，有时它们知道自己的业务价值，但选项太多。它们该如何选择合作的供应商？

在机器学习以及现在的AI领域，最优的工作方式是为正确的任务选择正确的子工具，并且能够轻松更换或调整。有时Snowflake是最好的数据库，有时则不是。有时OpenAI是最好的选择，而对于非常敏感的数据或为了控制成本，本地部署的模型有时是最佳选择。

拥有一个编排层，拥有一个平台，让你能够以安全、受控、合规的方式混搭这些底层个体工具，这是基础。

你需要一个集中式的、厂商中立的编排层，让你能轻松为合适的工作选择合适的工具。

企业如何应对不断上升的AI使用成本？

我把它比作云计算时代初期人们感受到的“账单冲击”——他们把一切都迁移到AWS上，第二年就收到一张巨额账单。

到2027年，随着更多人开始使用生成式AI，这种情况将会发生。

对我来说，当我在某件事上努力工作时，一天烧掉1000美元的token并不难，这并不便宜。

我确实认为会面临一些抵制，但这也极其有价值。如果在合适的人手中，你能构建的东西——我可以在一天或两个月内完成。这是无可替代的。

这就是为什么非常重要的是不要让一个智能体执行所有任务。你要尽可能将其分解成更多小步骤。在成本管理方面，另一件要注意的事情是对市场变化保持开放。因此，及时了解市场上最好的基础模型有哪些、价格点在哪里，确实能在管理定价方面产生巨大差异。能够在它们之间切换至关重要。

那来自阿里巴巴、DeepSeek和其他中国供应商的开放模型呢？企业应如何应对这一趋势？

它们很棒。我们看到很多第三方供应商在提供这些第三方中国模型或其他开源模型，以非常有竞争力的价格为用户提供服务。

你完全可以把第三方开源模型引入进来，省下很多钱。而且，它们在很大程度上取决于能力。

随着智能体在组织中扩散，管理它们将变得更加关键。对话式AI和用AI开发应用程序的趋势会持续下去。每个人都会构建这些东西。我们需要弄清楚什么有效、什么无效。

编者注：本次采访为清晰和简洁起见进行了编辑。

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.
To be successful using constantly evolving AI technology, enterprises need a deep understanding of their business processes and flexibility in the models and agents they use.
Generative and agentic AI has seen wide use by enterprises over the last few years. But now, as at the start of the AI boom with OpenAI’s release of Chat-GPT in 2022, it appears as if the rapid pace of technology development has not matched adoption. The current focus on the cost of using AI means that enterprises have to take a hard look at the ways they’re applying generative AI and make sure they’re gaining value in how they apply it.
However, recently a shift has started as enterprises have toned down their skepticism about how generative and agentic AI technology could help them. Businesses of late have felt a pressing need to implement the technology in light of the OpenClaw open source personal agent phenomenon, Anthropic’s release of the domain-adaptable Claude Cowork and powerful Mythos models and Nvidia CEO Jensen Huang's call to enterprises to embrace an “OpenClaw strategy.”
In this interview from the Ai4 2026 conference in Las Vegas earlier this month, Jed Dougherty, senior vice president of AI and platform at enterprise AI and machine learning platform vendor Dataiku, discusses some of the obstacles enterprises face with agentic autonomy and choosing the right models or agents. For Dougherty, no matter the brand of AI an enterprise chooses, it must manage it so it works for its organization.
Should enterprises now fully accept the autonomy of AI Agents, given the success of OpenClaw, or is there still a need for more human-in-the-loop?
Jed Dougherty: There’s still a need for human guidance. The biggest gains and successes that people see while using agents are still at enterprises that still have humans in the loop. Every company needs a strategy for identifying what can be fully transitioned to fully autonomous systems. Identifying which policies and processes within your organization can be autonomized, or which subsets could be, is important.
What role do FDEs, or forward-deployed engineers, play in getting enterprises to adopt AI technology?
Dougherty: We’ve invested a lot of time and money in hiring and putting together a strong, forward-deployed engineering team that we help our clients with. It's very important. For example, I was helping a client. We were trying to build out some relatively simple websites. I did not anticipate the number of small technical hurdles I take totally for granted because I do them every day that stop people in their tracks, or that just drastically slow down the first point at which you could deploy a website. Having a forward-deployed engineer who understands how to do the simple infrastructure stuff is still valuable.
How much need is there for experts to make agentic AI work? And would you say vendors were wrong in their messaging at the start of wide use of generative AI by trying to make it look like a tool that doesn’t require experts?
Dougherty: I don't think so. AI is a brand-new technology. Four years ago, people were pretty good at self-service with machine learning because they'd been doing it for 10 years. Now, with new agent capabilities spreading everywhere, vendors need to spend a little time getting people over this initial hump. In a year or two, as more people have touched the technology and gotten used to it, I don't think you'll need quite the same level of hands-on assistance. But right now, when I'm trying to describe to somebody who's never touched an agent before or built an agent before, it’s just helpful to have an expert in between. The business transformational capabilities of having structured agents that replicate large portions of your business pipelines is relatively complicated.
What is the best use case for agentic AI that will transform enterprises?
Dougherty: A business needs to know itself before it can identify the best use case. The best use case for agentic AI in any business is its core process.
Let's take insurance, for example. An insurance claim comes in, and some processing is done on that claim. Some investigation must be done into that claim. A decision is made on whether to pay out that insurance claim. There's probably some auditing. Maybe they will tell the legal team. So, the core part of the insurance business is deciding how and whether to pay insurance claims, and whether to do business with that customer again. So, the better an insurance company can describe, right now, how they do that without generative AI, the better they'll be able to understand which subsets of that insurance claim payoff process can then be replicated or augmented by agentic AI.
For enterprises, though, sometimes they know their business value, but there are so many options. How can they choose which vendor to partner with?
Optimal work with data in machine learning and now in AI means choosing the right sub-tool for the right task and being able to swap or tweak easily. Sometimes Snowflake is the best database, and sometimes it's not. Sometimes OpenAI is the best option, and sometimes an on-prem model is the best option for very sensitive data or to watch costs.
Having an orchestration layer, having a platform that allows you to mix and match these underlying individual tools in a safe, controlled, governed way, that's the foundation.
You want a centralized, agnostic orchestration layer that lets you easily choose the right tool for the right job.
How can enterprises deal with the rising cost of using AI?
I liken it to the cloud sticker shock people felt at the beginning of the cloud age, when they swapped everything over to AWS and then got a massive bill the next year.
That's going to happen come 2027 as more people start using generative AI.
It's not difficult for me to burn $1,000 in tokens a day when I'm working hard on something, and that's not cheap.
I do think there's going to be some pushback to that, but it's also incredibly valuable. If in the right person's hands, the things you can build, I can build in a day or two months. There's no substitute for that.
That's why it's very important not to have a single agent performing all the tasks. You want it broken down into as many small steps as possible. The other thing to keep track of in cost management is being open to shifts in the market. So, keeping on top of the market as far as what the best foundation models are and what the price points are really does make a drastic difference, potentially in managing your pricing. Being able to switch between them is critical.
What about the open models from Alibaba, DeepSeek and other Chinese providers? How should enterprises respond to that trend?
They're great. We're seeing a lot of third-party [vendors] that are providing these third-party Chinese models or the other open source models, serving them up for people at a very competitive price point.
You absolutely can slide in third-party open source models and save yourself a lot of money. Plus, they're essentially dependent upon capability.
As agents proliferate across your organization, managing them will become much more critical. Conversational AI and using AI to develop applications are here to stay. Everybody's going to be building this stuff. We need to figure out what works and what doesn’t.
Editor’s note: This interview has been edited for clarity and conciseness.