企业应该拥有什么在他们的AI Stack?

思想领袖

企业应该拥有什么在他们的AI Stack?

在 Google 上的首选来源中添加 Unite.AI

AI转型的下一阶段将涉及公司可以从它所做的每个决定中学到什么，以及谁是学习的主人

围绕AI的争论集中于一个主要问题：这些模型能做什么？它们被设定了基准，进行了试验，令人惊奇，这是试验阶段的正确问题。 然而，这不是我们今天所处位置的正确问题。

大赦国际目前正在开展工作，监测变化，准备交付成果，并参与下一阶段的关键决定——代理企业——以推进业务，而不仅仅是回答问题。 这种转变正在加速：IBM的2026年CEO研究发现，CEO预计到2030年AI做出的业务决定份额将几乎翻一番到48 %。 战略问题从“这个模式能做什么？” 变为“我的公司每次做决定都学什么？

如果企业价值一次累积一个决定，那么公司在AI时代最重要的资产就是让组织从每个决定中学习的循环，具体来说就是它背后的背景，导致它的证据，由此产生的结果，以及连接它们的模式。 拥有这个圈子，每个决定下一个更清晰。 错过它，你正在自动化的活动没有积累优势。

几乎所有事情都改变了一个严肃的企业应该如何设计它的AI。

将你的公司投注在一个模型上的失误

其本能一般是通过挑选一个主要的模型提供者来巩固，将其标准化，并在上面建立一切。 不过我认为这是个错误

通常的反对意见是锁定和定价杠杆。 这些是真实的，但都是次要的。 更深层的问题是激励和你的学习走向。 模式提供者从事的是建立世界上最好的通用模式，其做法是通过学习使用。 当你通过一个单一的供应商来引导你公司的专有决定、数据和工作流程时，你将把你拥有的最有区别的材料，即你公司如何看待的记录，交给一个供应商，他的动机是将其概括。 你的优势成为竞争对手也可以购买的商品的投入。

边疆也走得太快了，无法在任何一个模式上长期保持最佳状态。 相反，正确的架构应允许企业将任务引导到最适合该任务的模型或专用的agent，同时拥有agent的牵引和学习循环，以捕捉企业的运作方式并作出决策。 模式可以改变；这种专有基础仍然属于你，并随着时间推移变得更加宝贵。

这种专有基础也可能具有可衡量的战略价值。 IBM的研究还发现，CEO将专有数据和IP纳入定制AI模型，agents预计其2030年收入的13%将来自他们今天不提供的产品和服务。

您无法构建每个 Agent，您不应该尝试

拥有你的绳索环路和路线并不意味着每个agent 运行在它们上面。 领头企业agent——了解你的组织，你的优先事项，你的专有数据——是你的建设，你的拥有。 但是，企业的工作跨越了任何一个团队都无法掌握的领域：销售、人力资源、法律、金融、市场情报等等。 每一个都将会得到最好的专业的agent 在那一件事上是很好的。

任何企业都不会为内部的每个功能建立最优秀的agent. 这不是一个资源竞赛你可以赢得。 获胜模式是一个强大的导师agent组织一个专门的子agents的名册，每个子agents都适合你公司的要求，每个子agent通过干净的接口插入你的控制器。 然而，企业在建造这种建筑方面还很早，2026年德勤公司的一项调查发现，只有15%的组织已扩大结构，跨功能多agent的采用。

因此，企业领导人的真正工作是双重的：构建北星领先的agent及其反馈的环路，然后在专业agents中识别获奖者并将其整合。 拥有权会复合

一个专用的 Agent Versus a spraper

专业的agent必须带来企业无法轻易再制造的东西。 长久的优势是专有数据加上专门制造的绳索，以检索和解释数据。 agent只和它能找到的一样好，而且它如何理解它发现了什么。

这就是在决定在哪里建设和在哪里建立伙伴关系时，律师协会企业领导人应当使用的方法。 如果某一能力取决于独特的内容、域内专门知识和专门建造的基础设施的年代，内部重建这种能力不太可能产生优势。 将这些资源集中到专有数据，决策循环，和独有的管弦系统上。

赢家会明白什么

赢得代理时代的公司将会是那些建筑设计者如何让这些公司随着时间的推移变得特别聪明。 他们将拥有允许他们从每个决定中学习的环路，同时以最好的专业agents环绕，而不是试图单独构建一切。

目标不是拥有AI堆栈的每一块。 它拥有你的优势所在的碎片和其余的伙伴。 连接与智能不同。 在代理企业中，智能是指拥有哪些化合物，并为其他化合物建立伙伴关系。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Thought Leaders

What Should Enterprises Own in Their AI Stack?

Add Unite.AI to your preferred sources on Google

The next phase of AI transformation will be about what companies can learn from every decision it makes, as well as who owns that learning

Conversations surrounding AI have been focused on one main question: what can these models do? They were benchmarked, piloted, and marveled at, and it was the right question for the experimentation phase. However, it’s not the right question for where we stand today.

AI is now doing work, monitoring changes, preparing deliverables, and participating in key decisions in this next phase — the agentic enterprise — to move business forward rather than just answering questions. That shift is accelerating: IBM’s 2026 CEO Study found that CEOs expect the share of operational decisions made by AI to nearly double to 48% by 2030. Once AI is doing the work rather than merely describing it, the strategic question changes from “What can this model do?” to “What does my company learn every time it makes a decision, and who owns that learning?”

If enterprise value accrues one decision at a time, then a company’s most important asset in the age of AI is the loop that allows the organization to learn from every decision, specifically the context behind it, the evidence that led to it, the outcome that resulted from it, and the pattern that connects them. Own that loop, and every decision makes the next one sharper. Miss it, and you are automating activity without accumulating advantage.

By simply reframing it, almost everything changes about how a serious enterprise should architect its AI.

The Mistake of Betting Your Firm on One Model

The instinct has generally been to consolidate by picking a leading model provider, standardizing on it, and building everything on top of it. However, I think that’s a mistake, although not for the reason people usually give.

The usual objection is lock-in and pricing leverage. Those are real, but they are secondary. The deeper problem is about incentives and where your learning goes. Model providers are in the business of building the best general-purpose models in the world, and the way they do that is by learning from usage. When you route your firm’s proprietary decisions, data, and workflows through a single provider, you are handing the most differentiating material you have — the record of how your company thinks — to a vendor whose incentive is to generalize it. Your edge becomes an input to a commodity your competitors can also buy.

The frontier is also moving too fast to bet on any one model staying best for long. The right architecture should instead allow enterprises to route tasks to the model or specialized agent best suited for that task, while owning the agent harness and learning loop that capture how the business works and makes decisions. Models can change; that proprietary foundation remains yours and grows more valuable over time.

That proprietary foundation may also have measurable strategic value. IBM’s study also found that CEOs incorporate proprietary data and IP into custom AI models and agents expect 13% more of their 2030 revenue to come from products and services they do not offer today.

You Can’t Build Every Agent, and You Shouldn’t Try

Owning your harness, loop, and routing does not mean building every agent that runs on top of them. The lead enterprise agent — the one that understands your organization, your priorities, your proprietary data — is yours to build and yours to own. But the work of the business spans domains no single team can master: sales, HR, legal, finance, market intelligence, and more. Each of those will be served best by a specialized agent that is excellent at that one thing.

No enterprise is going to build a best-in-class agent for every function internally. It is not a resourcing race you can win. The winning pattern is a strong lead agent orchestrating a roster of specialized sub-agents, each one customizable to your firm’s requirements, each plugging into your harness through a clean interface. Yet enterprises are still early in building that kind of architecture, with a 2026 Deloitte survey finding that only 15% of organizations have scaled orchestrated, cross-functional multi-agent adoption.

So, the real work for enterprise leaders is twofold: build the north-star lead agent and the loop it feeds and then identify the winners among specialized agents and integrate them. Ownership where it compounds; partnership where it doesn’t.

A Specialized Agent Versus a Wrapper

A specialized agent has to bring something the enterprise cannot easily recreate. The durable advantage is proprietary data plus a purpose-built harness to retrieve and reason over it. An agent is only as good as what it can find and how well it understands what it found.

That is the bar enterprise leaders should use when deciding where to build and where to partner. If a capability depends on unique content, domain expertise, and years of purpose-built infrastructure, rebuilding it internally is unlikely to create an advantage. Focus those resources instead on the proprietary data, decision loop, and orchestration layer that are uniquely yours.

What the Winners Will Understand

The firms that win the agentic era will be those that architect around what makes them uniquely smarter over time. They will own the loop that allows them to learn from every decision, while surrounding it with the best specialized agents rather than trying to build everything alone.

The goal is not to own every piece of the AI stack. It is to own the pieces where your advantage compounds and partner for the rest. Connected is not the same as intelligent. In the agentic enterprise, intelligent means owning what compounds and partnering for the rest.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.