2026年重大构想：企业编排层

2026年，多人模式将正式上线。

2026 is when multiplayer mode comes into gear.

如果你有一群代理自主工作，难道不会存在大规模的多代理故障连锁反应的风险吗？

If you have a bunch of agents autonomously working, isn't there potential for a huge, you know, multi agent cascade of failures?

关于AI帮助自动化工作并降低成本，有很多说法。

There's a lot of narrative around AI helping automate work and reducing cost.

但我认为，在AI真正强化商业模式、推动收入增长的情况下，客户采用这项技术的潜力是无限的。

But I think in instances where AI is actually reinforcing the business model in driving revenue, there's really no limit to the amount that customers may want to adopt that technology.

真正竞争的不是AI。

It's not AI that's the competition.

而是你的竞争对手在使用AI。

It's your competitors using AI.

每年，我们都会退一步问一个简单的问题。

Every year, we step back and ask a simple question.

接下来，开发者们会关注什么？

What will builders focus on next?

我们2026年的重大构想汇集了投资团队认为将塑造来年科技趋势的主题。

Our twenty twenty six big ideas bring together the themes our investing teams believe will shape the coming year in tech.

本集围绕一个核心构想展开。

This episode is built around one big idea.

人工智能正逐渐成为企业内部的协调层，而非一系列独立工具的集合。

AI is becoming an orchestration layer inside the enterprise, not a collection of standalone tools.

一个能够跨部门和软件进行规划、分析与执行工作的协同代理系统。

A coordinated system of agents that can plan, analyze, and execute work across departments and software.

你将听到四种观点，探讨当人工智能开始运行工作流程时发生的变化、组织如何提取上下文、为何遗留系统替换加速、多玩家人工智能在实际中的形态，以及这些系统为何具有商业护城河。

You'll hear four perspectives on what changes when AI starts running the workflow, how organizations extract context, why legacy replacement accelerates, what multiplayer AI looks like in practice, and what makes these systems commercially defensible.

要理解这一转变，我们先从企业整体视角出发。

To understand the shift, we start with the enterprise wide view.

西玛·安布认为，从实验阶段转向协调的多代理系统，将迫使企业从文档、流程和人员头脑中提取隐性知识，将其转化为可用的操作上下文。

SEMA AMBL argues that the move from experimentation to coordinated multi agent systems will force organizations to extract tacit knowledge from documents, processes, and people's heads, turning it into usable operational context.

以下是西玛的分享。

Here's Seema.

你好。

Hi.

我是西玛·安布尔，我们应用投资团队的合伙人。

I'm Seema Amble, a partner on our apps investing team.

我的2026年核心观点是，人工智能将创造一个新的协调层和新的岗位，尤其是在财富五百强企业中。

My 2026 big idea is that AI will create a new orchestration layer and new roles, particularly in the Fortune five hundred.

到2026年，企业将从孤立的人工智能工具进一步转向多智能体系统，这些系统需要像协调的数字团队一样运作。

In 2026, enterprises will shift further from isolated AI tools to multi agent systems that'll need to behave like coordinated digital teams.

当智能体开始管理复杂的相互依赖的工作流程，例如共同规划、分析和执行时，组织将需要重新思考工作的结构以及上下文在这些系统中的流动方式。

As agents start to manage complex interdependent workflows, like planning, analyzing, and executing together, organizations will need to rethink how work is structured and how context flows across these systems.

财富五百强企业将最深刻地感受到这一转变。

The Fortune five hundred will feel this shift most acutely.

它们拥有最丰富的孤岛数据、组织知识和运营复杂性，其中许多知识仍存在于员工的头脑中。

They sit on the deepest reservoirs of siloed data, institutional knowledge, and operational complexity, much of which sits in people's brains.

要将这些上下文从人们的头脑中提取出来，需要结合收集文档和观察人类行为。

To get this context out of people's brains, it's some combination of collecting documentation and watching human actions.

文档是什么？

What's the documentation?

可能是入职视频、书面说明，或者已经写好的完整文档。

It could be onboarding videos, written instructions, full documentation that's been written up.

而观察人类行为，则是真正地观察人类在浏览器中如何点击、他们采取的行动、打的电话等等，然后将这些信息整合为共享的上下文。

And then the watching human actions is literally watching how humans are clicking around in their browsers, the actions they take, the phone calls they make, etcetera, and then piecing this together as shared context.

这些代理之间需要解决什么问题？

What needs to be solved across these agents?

就是在代理之间提供反馈，并最终能够确定，在这种情况下，谁是优质客户，我们的投资回报率如何，以及我们是如何花费金钱或时间的。

It's providing the feedback across the agents and being able to ultimately determine, in this case, who is a good customer and are we getting the ROI, how we're spending our dollars or our time.

更具体地说，客服需要能够说：这个客户不好，销售团队。

To put that even more concretely, customer support needs to be able to say, this is a bad customer, sales.

你们应该减少对客户A的优先级投入，转而关注客户画像B。

You should spend less time prioritizing customer a and go for a customer profile B.

但目前，如果我们观察一下，销售代理是独立运作的，客服代理也是独立运作的，而且他们可能更多地被以效率指标来衡量，而不是从整体上考虑什么最有利于企业。

But right now, if we looked at it, the sales agent is operating autonomously, the support agent is operating autonomously, and they're probably, if anything, being measured more on efficiency metrics versus holistically looking at what's best for the business.

由此自然会产生一个问题：如果你有一群自主工作的代理，会不会出现大规模的多代理级联故障？

One of the natural questions that comes out of this is, well, if you have a bunch of agents autonomously working, isn't there a potential for a huge, you know, multi agent cascade of failures?

是的，这是有可能的。

Yes, it's possible.

但请记住，我们并不是一夜之间就转向这种模式的。

But remember, we're changing to this overnight.

在任何组织中，都可能存在多人级联故障的情况。

There could be, you know, multi human cascading failures in any organization.

我认为代理也应被类似地对待。

I think agents have to be treated similarly.

如果从这个角度来看，有两个检查机制。

If you think about it this way, there are two checks.

一是仍然会在各个节点保留人工干预。

One is there still will be humans in the loop at various points.

这将是一个检查点。

That will be one check.

那么，人类和代理最终分别会做什么？

And what will the human do and eventually the agent?

会有一套审计流程和评估机制。

There will be a set of audit procedures and evals.

再说一遍，回到这些可量化的指标，好吧。

You know, again, go back to these quantifiable metrics and say, okay.

我们的销售代理表现非常出色。

Our sales agent is doing really well.

我们成功签下了大量客户。

We're closing a lot of customers.

我们的谈判代理带来了很好的价格，但所有客户都在流失。

Our negotiation agent is bringing in great pricing, but all our customers are churning.

如果我们把这些指标相互对比，发现某个指标相对于其他指标过高，并且我们有这些可量化的数据，就可以回过头来调整任何代理的目标函数。

If we measure all those against each other and say, we see that one is too high relative to the others and we have these quantifiable metrics, we can go back and change the objective function for any of the agents.

我认为每个代理都会有自己独立的评估函数，并且像人类一样拥有关键绩效指标。

I think every agent will have its own eval function, and it will have KPIs just like humans are measured against right now.

必须有一种逻辑，能够表达‘如果a，那么b’。

There will have to be logic that's saying if a, then b.

最终，正如组织会朝着一系列整体的组织KPI努力一样，智能体也将如此运作。

Ultimately, right, just as organizations work towards some set of overall organizational KPIs, that's how agents will work too.

在这个问题背景下，与财富500强企业合作存在巨大的机会。

There's a huge opportunity specifically working with Fortune 500 in the context of this problem.

如今，我们看到财富500强公司对人工智能非常感兴趣，但更多还停留在实验阶段，而非深度实施人工智能。

Today, we've seen Fortune 500 companies be very interested in AI, but it's been, I'd say, more on the experimentation side than deeply implementing AI.

但我认为这种情况即将改变。

But I think that's about to change.

对于财富500强企业来说，这尤其有趣，因为随着这些组织的发展，它们在人员和流程之间积累了大量孤立的信息。

It's most interesting for the Fortune five hundred because they have all of this siloed context across people and processes as these organizations have gotten built.

许多财富500强企业是通过收购成长起来的。

A lot of Fortune five hundred have grown through acquisition.

它们拥有不同的地域分布。

They have different geographies.

每个地区都有不同的软件系统。

Each of these geographies have different software systems.

他们有不同的人员。

They have different people.

他们的运作方式也不同。

They operate differently.

这在今天意味着什么？

And what does that mean today?

这些公司都运作得非常缓慢且官僚。

These companies all operate very slowly and bureaucratically.

实施新软件需要数年时间。

Implementing new software takes years.

任何改变都需要很长时间。

Anything that to change, you know, takes forever.

现在，如果你能创建这样一个上下文层，把人们脑海中的信息提取出来并构建上下文层，那么想象一下，部署新的ERP系统或新的采购代理就会快得多。

Now if you're able to create this context layer where you're able to take things out of people's heads and create a context layer, you know, imagine, like, putting in a new ERP or a new procurement agent becomes much faster.

然后，这些代理可以以一种远比亚洲团队和欧洲团队需要频繁开会、两人持续沟通以完成跨越多个地区的合同更快的方式协同工作。

And then you can actually have these agents work with each other in a way that's much faster than the Asia team and the Europe team needing to set a bunch of meetings and two people needing to continuously talk to each other about closing a contract that spans multiple geographies.

我最兴奋的是这种将人们头脑中的信息提取出来，从而突然释放代理真正潜力的能力。

What I'm most excited about is this ability to pull things out of people's heads and then suddenly, you know, unlock the real power of agents.

我认为财富五百强企业拥有最分散和孤立的数据。

And I think the Fortune five hundred has the most siloed and distributed data.

我认为这为实现更顺畅的运营带来了巨大机会。

And I think there can be a lot of opportunity for smoother operations.

SEMA为我们提供了运营模式：协调、上下文提取和数字团队。

SEMA gives us the operating model: orchestration, context extraction, and digital teams.

现在，让我们看看这一模式变得不可避免的最清晰的行业。

Now let's look at the clearest industry where this becomes unavoidable.

安吉拉·斯特兰奇专注于金融服务和保险领域，在这些领域，统一的数据和并行化的工作流程使取代遗留核心系统、提升速度、利润率和规模成为可能。

Angela Strange focuses on financial services and insurance, where unified data and parallelized workflows make possible to replace legacy cores and unlock speed, margin, and scale.

以下是安吉拉。

Here's Angela.

我是安吉拉·斯特兰奇，AI应用基金的普通合伙人。

I'm Angela Strange, a general partner on the AI Applications Fund.

我对2026年的重大预测是，金融服务和保险行业将迎来一个重大转折点。

And my big idea for 2026 is there will be a dramatic turning point coming to financial services and insurance.

或者说，不取代遗留系统的风险将超过变革本身的风险。

Or finally, the risk of not replacing legacy systems will exceed the risk of change.

这已经正在发生。

It's already happening.

大型机构将让现有合同到期，并采用他们新一代的AI原生竞争对手。

Major institutions will let standing contracts lapse and implement their newer AI native competitors.

为什么？

Why?

下一代基础设施不仅仅是添加AI。

The next generation of infrastructure doesn't just add AI.

它们将来自遗留核心系统、外部系统和非结构化数据的数据统一到一个新的记录系统中，使金融机构不仅能实现规模化，还能充分释放AI的潜力。

They unify the data from legacy cores, from external systems, from unstructured data into a new system of record, enabling FIs not only to scale, but to take full advantage of AI.

当这种情况发生时，对于客户和构建者来说，将出现三个重大变化。

When this happens, there are three major changes that are important for both customers and builders.

第一，工作流程终于将实现并行化。

One, workflows will finally become parallelized.

不再需要在不同屏幕间来回切换、复制粘贴数据。

No more bouncing between screens, cut pasting data.

例如，你的房贷团队可以查看为审批你的贷款所需的400多项任务，同时并行处理这些任务，甚至让代理完成其中一些更琐碎的工作，供你后续核查。

For instance, your mortgage team could see the 400 plus tasks that are needed to underwrite your loan, do them in parallel, and even have agents do some of the more mundane ones for you to check later.

第二，我们所熟知的分类将得到扩展。

Second, the categories as we know them are going to expand.

例如，从客户入职、了解你的客户（KYC）、交易监控，到客户与客服团队互动的行为，所有这些数据都可以整合到一个统一的风险平台中。

For instance, customer data from onboarding, KYC, KYC, transaction monitoring, even how those customers behave with your customer service team, could all sit into a single risk platform.

更有效地整合欺诈与风险合规功能。

Brings together fraud, risk compliance much more effectively.

第三，对于构建者而言，最令人兴奋的是，新的赢家规模将扩大十倍。

And then third, most excitingly for the builders, the new winners here will be 10x bigger.

这不仅是因为这些软件类别变大了，还因为软件能够承接大量人类本来就不想做的工作，或者银行和保险公司无法快速招聘到足够人力完成的工作。

Not only because those software categories are bigger, but because software is able to consume a lot of the labor that humans didn't want to do anyways or that banks or insurance companies couldn't hire for fast enough.

所以正如那句老话所说，真正的竞争对手不是人工智能。

So as the saying goes, it's not AI that's competition.

而是你的竞争对手正在使用人工智能。

It's your competitors using AI.

因此，最优秀的银行和保险公司将会修复其底层系统，充分把握机遇，在下一个十年中保持最强竞争力。

So the best banks, the best insurance companies will fix their plumbing and enable them to take full advantage and be the most competitive going into the next decade.

几十年来，企业一直在谈论这个问题。

Companies have been talking about this for decades.

那么，现在有什么不同呢？

Why is it different now?

主要有三个原因。

Primarily three reasons.

第一，我们必须记住，许多这些公司仍在使用几十年前的大型机，它们的系统在现有规模下早已濒临崩溃。

One, we have to remember that many of these companies still live on mainframes, decades old mainframes, and their systems were already on the verge of breaking with the scale.

第二，现在企业意识到，由于无法利用人工智能，他们正在放弃大量收入。

Two, now companies see that they're leaving a lot of revenue on the table by not being able to take advantage of AI.

例如，在保险行业，核保人员有时甚至无法应对他们所面临的业务需求，因为他们无法及时处理这些任务。

For instance, in insurance, underwriters sometimes can't even get to the demand that they have because they're not able to process it fast enough.

他们无法获取文件。

They can't bring in the documents.

他们无法扫描这些文件。

They can't scan them.

如果你能部署正确的系统，并在其上叠加人工智能，就能捕获巨大的收入增长机会。

This is a huge revenue upside that can be captured if you get the right system and you layer AI on top.

第三，现在出现了许多可行的下一代AI优先软件，由深刻理解你行业的创业者打造，他们技术精湛，并彻底重构了你的平台，一方面让你能够扩展规模，另一方面让你能够极其灵活地现在和未来轻松集成人工智能。

Third, there are strong viable options of this next generation of AI first software built by entrepreneurs who deeply understand your industry, are deeply technical, and have entirely rearchitected your platforms to, one, enable you to scale, and, two, be incredibly flexible in terms of how you can add AI on now and in the future.

我在这里看到了大量机遇，潜在地将根据谁成为这些新平台的早期采用者，彻底重塑现有公司的胜负格局。

I see a ton of opportunity here and potentially a dramatic reordering of the winners and losers of incumbent companies based on who become the early adopters of some of these new platforms.

我们已经看到了这种变化。

And we're already seeing it.

有一些银行和保险公司已经开始建立起前瞻性强、易于合作的声誉，特别是在抵押贷款服务等领域，一些公司已经将业务的利润率从5%提升到了50%。

There's some banks and there's some insurance companies that are starting to get the reputation of being forward thinking, easier to work with, wanting to lean companies in some areas like mortgage servicing have been able to turn areas of their business from 5% margin businesses to 50% margin businesses.

如果你能尽快在整个公司范围内实现这样的转变，将比那些需要两三年才能赶上来的竞争对手带来大得多的优势。

And you imagine doing that across your company as quickly as possible is gonna make a much bigger difference against your competitor that maybe takes two or three years to catch up.

作为一名投资者，我之所以对基础设施如此兴奋，是因为它提供了优美的基础设施，从而支撑了出色的消费者体验和商业体验。

One of the reasons as an investor that I get so excited about infrastructure is that it's beautiful infrastructure that enables beautiful consumer experiences and beautiful business experiences.

例如，为什么你的银行会向你推销你已经拥有的产品？

For instance, why does your bank market products to you that you already have?

这是因为你的客户数据分散在各个不同的系统中。

It's because your customer data sits in all of these different sectors.

为什么当你致电咨询银行业务时，客服人员A无法回答关于客服B的问题？

Why can't customer service agent A answer questions about customer service B if you call in about your banking operations?

想象一下未来拥有统一的数据层，再加上一群极其聪明、并由智能代理辅助的人才，这些代理能够理解你的需求，帮助你使用你已有的任何产品，并预测你未来的需要——这将为客户和企业都带来绝佳的体验。

Now imagine the future of a unified data layer and incredibly smart people supplemented by agents that can understand your needs, help you with any product you already have, anticipate your needs in the future, that would be a beautiful experience for both customers and businesses.

到2026年，任何构建了面向这一庞大行业的全新AI优先平台的公司，都将看到显著的加速发展。

In 2026, we're gonna see a dramatic acceleration for any company that has built a new AI first platform that sells into this large industry.

但这个机会是巨大的。

But the opportunity is massive.

因此，如果你是一位创始人，深刻理解或对银行业或保险业的任何陈旧方面充满好奇，现在就是机会。

So if you're a founder who deeply understands or is deeply curious about any archaic aspects of banking or insurance, the opportunity is now.

你可以更快地开发软件，而客户也准备好购买。

You can build your software faster and customers are ready to buy.

安吉拉阐述了为什么现在会发生这种情况。

Angela makes the case for why this happens now.

现代平台统一了数据，代理可以并行运行任务，从而改变客户体验和经济模式。

Modern platforms unify data and agent can run work in parallel, changing both the customer experience and the economics.

接下来是产品影响。

Next is the product implication.

这种编排层在软件内部是什么样子的？

What does this orchestration layer look like inside the software itself?

亚历克斯·伊曼纽尔描述了垂直型人工智能进入多人协作模式，在这种模式下，多个真人和多个代理在工作流中协作，并有明确的信任规则和一个指挥中心界面，区分代理可以执行的操作和人类需要审查的内容。

Alex Imerman describes vertical AI moving into multiplayer mode, where multiple humans and multiple agents collaborate inside a workflow with explicit trust rules and a command center interface that separates what agents can execute for what humans need to review.

这是亚历克斯。

Here's Alex.

我对2026年的主要构想是，垂直型人工智能将从信息检索与推理演进到多人协作模式。

My big idea for 2026 is vertical AI is gonna evolve from information retrieval and reasoning to multiplayer mode.

垂直型软件正当时，但在ChatGPT出现之前，垂直型软件就已经很酷了。

Vertical software is having a moment, but vertical software was cool before ChatGPT.

Shopify、Viva、Procore、Toast都已成长为市值数百亿甚至上千亿美元的公司。

Shopify, Viva, Procore, Toast have all scaled to tens or even hundreds of billions of market cap.

庞大的企业。

Huge companies.

但垂直型AI公司的发展速度更快，远超我们过去在SaaS领域所见的历史先例。

But vertical AI companies, they're growing faster, faster than historical precedents that we saw in SaaS.

我们所有人讨论AI时的一个有趣方面是，智能代理正在取代人力。

One of the cool aspects that we're all talking about with AI is how agents are replacing labor.

取代律师比取代通才更容易。

It's easier to replace a lawyer than it is to replace a generalist.

为垂直领域构建，为特定类型的员工设计，意味着需要深度集成、专有数据和专用界面，而像我如此喜爱的ChatGPT这样的通用平台在这方面无法匹敌。

Building for a vertical, building for a specific type of employee means deep integrations, proprietary data, specialized interfaces that a horizontal, as much as I love ChatGPT, is not gonna be as good at.

我们观察到垂直AI经历了三个阶段的演进。

We've observed vertical AI evolve across three phases.

第一阶段是信息检索。

First was information retrieval.

你阅读一些文档，提取信息，并可能对其进行总结。

You read some documents, you extract information, and you might summarize it.

第二阶段出现在今年，即2025年。

The second came this year in 2025.

推理。

Reasoning.

推理能力对垂直软件企业产生了巨大影响。

Reasoning capabilities have been really impactful for vertical software businesses.

在Hebia中，你正在分析财务报表并构建模型。

With Hebia, you're analyzing financial statements and building models.

使用 Basis，你可以对试算平衡表进行对账。

With Basis, you're able to reconcile trial balances.

使用 Elise AI，你可以诊断出维护问题并联系合适的供应商。

And with Elise AI, you're able to diagnose what the maintenance issue is and contact the right vendor.

问题是，所有复杂的工作都需要协作。

The problem is that with all complex work, there's collaboration.

需要多人模式。

Multiplayer mode is required.

2026 年，多人模式将正式发挥作用。

2026 is when multiplayer mode comes into gear.

如果你想完成的不只是一个独立的任务，而是整个工作，你就必须能够与他人协作。

If you wanna accomplish not just a discrete task, but the full job, you need to be able to collaborate with others.

因此，多人员与多智能体协作正在到来。

So multihuman and multi agent collaboration is on its way.

随着这一点的实现，这些平台的价值将不断提升。

And with that, the value of these platforms increases.

而且切换成本上升，这让我们在思考这些平台的可防御性时感到非常振奋。

And the switching costs rise, which is really exciting as we think about defensibility of these platforms.

垂直类应用一直受到批评，认为在人工智能时代它们的可防御性不强。

Vertical apps have been criticized that they're not very defensible in this AI era.

它们能经受住时间的考验吗？

Will they stand the test of time?

最优秀的那些绝对可以。

The best ones absolutely will.

我对垂直类应用会关注几个特性。

A couple attributes that I look for with vertical apps.

第一，品牌。

One, brand.

在垂直市场中，推荐性非常高。

There's a high referenceability in vertical markets.

客户都会参加同样的会议。

The customers all go to the same conferences.

他们一起吃饭。

They go to dinner together.

因此，Elise AI 已成为物业管理领域的品牌。

And so, Elise AI has emerged as the brand in property management.

所有客户，所有大型物业管理公司一想到人工智能就会想到他们。

All the customers, all the large property managers know them when they think of AI.

第二种模式是专有技术或知识产权。

A second mode is proprietary technology or IP.

国防领域的 Anderol，公共安全领域的 Flock Safety，自动驾驶领域的 Waymo 或 Applied Intuition。

Anderol in defense, flock safety in public safety, Waymo or applied intuition in autonomy.

构建这样的技术非常困难，难以复制。

Really difficult to build technology, difficult to replicate.

然后回到网络效应。

And then coming back, network effects.

在多人模式下，随着越来越多的代理和人类在平台上获得越来越多的价值，转换成本随之上升，没有人会离开这个平台。

With multiplayer mode, as more agents and more humans find increasing value on the platform, switching costs arise and no one's leaving the platform.

我们预计这将融合，并成为2026年故事的重要组成部分。

We expect that to merge and be an important part of the 2026 story.

实现多人模式最大的障碍之一是建立信任。

One of the biggest obstacles to getting to multiplayer mode is building trust.

需要有AI操作协议，明确在什么情况下代理可以代表人类行动，或在什么情况下需要向人类报告问题。

There needs to be AI operating agreements, understandings of when an agent can act on behalf or when they need to flag an issue to their human.

起初，它们可能只能为你安排会议，但未来，随着信任度不断提升，它们将能站在前线进行谈判。

Initially, they might be able to schedule a meeting for you, but in the future, as they build more and more trust, they can be on the front lines negotiating.

让我们描绘一下这个场景。

So let's paint that picture.

你正在参与一笔并购交易。

You're in an M and A transaction.

你的代理已经建立了信任。

Your agent has built up trust.

它们有责任去进行谈判。

They have the responsibility to go negotiate.

你已经设置了参数。

You've set parameters.

如果你是卖方，正在出售一家企业，你会设定你愿意达成协议的最低价格。

So if you're the sell side, you're selling a business, you set the minimum price that you're willing to come to terms on.

而买方的代理则会设定他们愿意支付的最高价格。

And then the buy side agent, well, they'll set the max they're willing to pay.

如果这两个价格交叉，那就很好，可以达成一个高层次的协议。

And if those two cross, great, you can get to a high level agreement.

但还会有一些未解决的问题，比如交割时的营运资金安排，或如何处理或有事项或盈利支付。

But there's going be outstanding questions, like what's the working capital arrangement at close or how to deal with contingencies or earn outs.

代理可能没有足够的信息来代表你进行谈判，因此这些问题会被提报上来。

The agent may not have the information to negotiate on their behalf, so that gets flagged up.

因此，软件不仅仅是一个聊天界面，你可以把它看作是一个指挥中心。

And so software won't be just another chat interface, but you can think of it as a command center.

有一份正在谈判的活动清单，代理可以完全自主地去执行。

There is a list of activities that are being negotiated on, that agents have full ability to go and act.

然后还有一个独立的部分，即标志项，需要人类介入并采取行动。

And then there's a separate section, the flags, where humans need to engage and take action.

我非常期待这些新的用户界面，但我更兴奋的是，工作将越来越少地涉及执行，而更多地转向审查。

I'm really excited to see these new user interfaces, but what I'm more excited about is where work becomes less about doing and more about reviewing.

亚历克斯展示了AI驱动的工作流程在实践中是如何运作的：协作、运营协议以及围绕审查和升级设计的界面。

Alex shows what AI runs the workflow becomes in practice: collaboration, operating agreements, and interfaces designed around review and escalation.

要达成交易，我们需要商业过滤器：哪些AI系统真正会胜出并持续存在？

To close, we need the commercial filter: which AI systems will actually win and persist?

大卫·哈伯认为，最强大的公司是那些AI强化了商业模式、推动收入和成果（而不仅仅是降低成本），并通过工作流程所有权和专有成果数据建立防御性的公司。

David Haber argues the strongest companies are the ones where AI reinforces the business model, driving revenue and outcomes, not just cost reduction, and building defensibility through workflow ownership and proprietary outcomes data.

以下是大卫。

Here's David.

嗨。

Hey.

我是大卫·哈伯，十六C的普通合伙人，我协助领导AI应用基金。

I'm David Haber, general partner here at a sixteen c, and I help co lead the AI Apps Fund.

我对2026年的主要想法是寻找那些AI能强化商业模式的公司。

My big idea for 2026 is looking for companies where AI reinforces the business model.

我认为，现在有很多说法认为AI能帮助自动化工作并降低成本。

You know, I think there's a lot of narrative around AI helping automate work and reducing cost.

但在AI真正强化商业模式、推动收入增长的情况下，客户采用这项技术的潜力几乎是无限的。

But I think in instances where AI is actually reinforcing the business model in driving revenue, there's really no limit to the amount that customers may wanna adopt that technology.

因此，像这样的市场推动力远比单纯的成本削减故事要强劲得多。

And so the market pull in examples like that are just so much stronger than those where it's just a cost reduction story.

我担任一家名为Eve的公司的董事会成员，该公司从事原告法律领域。

I sit on the board of a company called Eve, which operates in the plaintiff law space.

原告法律的独特之处在于，这些律师并不按小时收费。

And what's unique about plaintiff law is that those attorneys don't charge by the hour.

他们采用风险代理模式，这意味着只有胜诉时他们才能获得报酬。

They operate on a contingency basis, which means that they only get paid if they win.

因此，尽管AI正在帮助自动化他们大量的起草和推理工作，但最终的核心还是帮助他们承接更多客户、赚取更多收入。

And so again, while AI is helping automate a lot of the drafting and reasoning work that they do, ultimately it's really about enabling them to take on more clients and make more money.

因此，它并没有侵蚀计时收费模式。

So it doesn't erode the billable hour.

它真正强化了他们的商业模式。

It really reinforces their business model.

因此，Eve这类AI工作空间的市场需求极其强劲。

And as a result, the market pull for Eve's kind of AI workspace has just been tremendous.

我们投资组合中的另一个例子是一家名为Salient的公司，它从事贷款服务业务。

Another example in our portfolio is a company called Salient, which operates in the loan servicing space.

他们将语音代理应用于汽车贷款领域，但已扩展到整个消费贷款产品生态系统，语音代理能以50种语言进行合规沟通，追踪UDAP，完成欢迎电话和还款提醒。

So they're applying voice agents to they started in auto lending, but they've expanded to a whole ecosystem of kind of consumer lending products where a voice agent can speak in 50 languages, fully compliantly, track UDAP, do welcome calls and payment reminders.

当然，这其中确实存在降低成本的故事。

And obviously, you know, there is a cost reduction story in that.

对吧？

Right?

它正在帮助许多拥有大型呼叫中心的银行和非银行贷款机构提升效率。

It is helping drive efficiencies in many of these bank and nonbank lenders who have large call centers.

但我认为他们发现的惊人之处在于，语音代理实际上提高了催收率。

But I think what they found, which is so remarkable, is that the voice agents are actually driving better collection rates.

对吧？

Right?

这不仅仅是一个降低成本的故事。

So it's not just a cost reduction story.

它实际上为最终客户带来了更好的结果。

It's actually delivering better outcomes for their end customers.

因此，它强化了贷款机构的商业模式。

And as a result, it's reinforcing, you know, the lender business model.

最终，人工智能应用中的复合竞争优势来源在哪里？

Ultimately, where do the sources of compounding competitive advantage reside in AI applications?

Eve 是一个非常独特且典型的例子。

Think Eve is a really unique example and case study for this.

最终，Eve 的创始团队有一个愿景，即掌控从接洽到结果的全流程。

Ultimately, the founders of Eve had a vision for owning the end to end workflow from intake to outcome.

深入融入你的客户，让他们每天都在产品中使用，以此构建护城河。

Think deeply embedding yourself within your customer, having them live within the product every day as a source of defensibility.

我认为他们还在创造一种非常独特的数据资产。

I think they are also creating a really unique data asset.

最终，通过能够从接案一直处理到结果，这些结果数据并非公开的。

Ultimately, by being able to process cases, again, from intake all the way to outcomes, that outcomes data is not public.

对吧？

Right?

这种信息来源并不是模型公司和实验室能在公开互联网上用来训练的数据。

That is not a source of information that, you know, model companies and labs can actually train on in the public Internet.

因此，这些结果数据被用来更好地指导更智能的接案，让EVE能够告诉客户：这个案件具有这些特征，可能值5万美元。

And so ultimately, that outcomes data is used to better inform smarter intake so that EVE can tell their customers, look, this case has these characteristics to potentially be worth, you know, $50,000.

这个案件可能值500万美元。

This case is potentially worth $5,000,000.

这就是你应该如何分配人力和时间的建议。

Here's how you may wanna triage, you know, your labor and your time.

最终，考虑到这一对方，你希望在催告函中加入哪些特征，以实现更好的结果？

And ultimately, given this counterparty, what are the characteristics that you may want to put into a demand letter to actually affect better outcomes?

因此，我认为Eaves处理的案件越多，平台就越智能、越强大，最终进一步巩固了其客户的商业模式。

And so I think the more cases that Eaves processes, the smarter and more powerful the platform becomes, again, ultimately reinforcing the business model for their clients.

以下是这四个想法之间的连接纽带。

Here's the connective tissue across all four ideas.

CEMA：从孤立工具向协调的智能代理团队转变，以及为什么上下文成为关键制约因素。

CEMA, the shift from isolated tools to coordinated agent teams and why context becomes the gating factor.

安吉拉：当统一数据和并行工作流释放了速度与利润率时，传统替代加速的转折点。

Angela, the turning point where legacy replacement accelerates because unified data and parallel workflows unlock speed and margin.

亚历克斯：软件在实际中变成了什么——围绕审阅构建的多人协作、信任规则和指挥中心用户体验。

Alex, what the software becomes in practice, multiplayer collaboration, trust rules, and command center UX built around review.

大卫：什么在商业上获胜？

David, what wins commercially?

端到端嵌入的平台，可衡量地提升结果，并强化客户创造价值的方式。

Platforms embedded end to end that measurably improve outcomes and reinforce how customers create value.

这就是企业编排层。

That's the enterprise orchestration layer.

不是聊天机器人，也不是一个功能，而是一种全新的公司工作流程方式。

Not a chatbot and not a feature, but a new way workflows through the company.

感谢收听本集的a16z播客。

Thanks for listening to this episode of the a 16 z podcast.

如果你喜欢这集，请务必点赞、评论、订阅、给我们评分或留言，并分享给你的朋友和家人。

If you like this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family.

如需收听更多集数，请前往YouTube、Apple Podcasts和Spotify。

For more episodes, go to YouTube, Apple Podcasts, and Spotify.

在X上关注我们@a16z，并在a16z.substack.com订阅我们的Substack。

Follow us on x at a sixteen z, and subscribe to our Substack at a16z.substack.com.

再次感谢收听，我们下集再见。

Thanks again for listening, and I'll see you in the next episode.

提醒一下，此处内容仅作信息参考，不应被视为法律、商业、税务或投资建议，也不应用于评估任何投资或证券，且并非针对任何a16z基金的投资者或潜在投资者。

As a reminder, the content here is for informational purposes only, should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a sixteen z fund.

请注意，a16z及其关联方可能也持有本播客中讨论的公司的投资。

Please note that a sixteen z and its affiliates may also maintain investments in the companies discussed in this podcast.

如需更多详情，包括我们的投资链接，请访问a16z.com/discoveries。

For more details, including a link to our investments, please see a 16z.com forward slash disclosures.