在AI中建立信心：在受监管企业中实现编排的运营化

很多公司都急着追求那种舞台时刻，你知道的，比如一个炫目的公告、公关时刻、新闻稿，能让你登上头条或者立刻成为焦点。

A lot of companies are just rushing for that stage moment, know, like a big flashy announcement, a PR moment, a press release that gets you on a first phase or targeted right now,

我得跟你说，真的。

I gotta tell you, like.

AI的承诺是真实的。

The promise of AI is real.

在你的组织中让它发挥作用，同样也是挑战。

So is the challenge of making it work in your organization.

欢迎收听《前进之路》，在这里我们探索AI时代工作的未来。

Welcome to The Path Forward, where we explore the future of work in the age of AI.

我是主持人克里斯蒂安·波茨。

I'm Christian Potts, your host.

今天很高兴邀请到UiPath的西米·古普塔和安德拉达·莫拉尔。

Pleased to be joined today by Simi Gupta and Andrada Morar from UiPath.

我们今天要讨论的是，如何自信地扩展规模？

We're gonna be talking today about how do you scale with confidence?

你怎样才能真正让你的组织达到你需要的标准，从而推动你想要的AI转型呢？

How do you actually make your organization meet the bar that you need to hit so that you can affect the AI transformation you're looking for?

所以前几集我们一直在讨论，我们正处于什么样的编排时刻？

So the first few episodes we've been talking about, what's the orchestration moment that we're in?

我们正处于怎样的AI时刻？

What is this AI moment that we're in?

我们还讨论了，如果你想把AI应用到业务中一个非常复杂的流程，一个每天驱动你业务的流程，比如收入周期管理，那会是什么体验。

We've also been talking about what's the experience like if you're looking to apply this to, like, a really complex process in your business, a process that actually drives your business every day, like revenue cycle management, for example.

我们还请来了丹尼尔，他分享了他对如何规划未来路径的一些看法。

We've also had, Daniel come in and talk about sort of, like, his vision for, like, the kind of the kind of ways you'd certainly need to chart the path towards where we're going.

我们稍后还会讨论人的因素。

And we're gonna talk about people later.

我对即将到来的那次对话感到非常兴奋。

I'm really excited about that conversation that's gonna be coming.

但今天，我真正想聚焦的是，在我们客户每天所处的复杂环境中，这实际上是什么样子的。

But today, what I really wanna focus on is, like, what does this actually look like in the kind of complex environments that our customers are in every single day?

比如金融服务或者医疗保健。

So like, financial services or health care.

这次我们有一位金融服务方面的专家，所以可能先聊这个领域。

In this case, we have an expert here on financial services, so that's probably where we're gonna go first.

但我想在这一集开头问你一个问题，安德拉达，就是当你以你的角色和客户交流时，你知道，你算是我们的客户体验副总裁。

But I wanna start this episode, Andrada, by asking you a question, which is when you talk to customers in a role that you're in, you know, you're sort of the you know, you're the VP of customer experience for us.

你确实和他们关系很紧密。

You really have that, like, that close relationship to them.

当你和他们谈到信任和信心时，最大的担忧是什么？

When you talk about trust and confidence with them, what are the biggest concerns?

哪些事情让他们夜不能寐？

What are the things that are keeping them up at night?

对。

Yeah.

我觉得情况变化很大。

So I think it's changed a lot.

对吧？

Right?

我甚至敢说，可能在这个播客播出之前，情况就会发生变化。

Like, I I would probably dare to say probably it will change before this podcast actually airs.

对。

Yeah.

如果你看看过去一个月，讨论的话题已经从“模型到底能做什么”变成了“我到底放心让模型或代理为我做什么”。

If you look at the past month, conversation has changed a lot from what is a model actually able to do to what am I actually comfortable to allow a model to do for me, an agent to do for me?

所以我觉得在过去几周甚至几个月里，出现了三个主题。

So I think there are three themes that I've seen emerging in the past, let's just say weeks and maybe months.

我不敢再往前追溯了。

I don't dare to go even more back in time.

第一个是关于边界的概念。

But the first one is this idea of boundary.

让模型帮你起草一封邮件是一回事，但让代理去做一个重大决定，比如批准或拒绝理赔，或者向监管机构发送文件，那完全是另一回事。

It's one to work with a model to draft an email and it's something completely different to allow an agent to make a really big decision like denying or saying yes to a claim or sending something to a regulator.

所以我觉得这个边界也取决于每个企业的情况。

So I think like it depends also from enterprise to enterprise where that boundary is.

我觉得审计也是个大问题。

I think also audit is another big thing.

这不仅仅是关于AI做决定是否正确，更重要的是，我想知道是什么让AI做出那个决定。

It goes beyond whether AI was right to make a decision or not, and it goes more into, I want to know what made AI make that decision.

你知道，有没有人参与其中？

You know, was there a human involved?

用了什么样的内容来让模型做出改变？

What kind of content was used to allow the model to make a change?

所以审计是个大事。

So audit is a big thing.

然后我觉得还有一个很多供应商其实不太愿意明说的问题，那就是行动太快带来的机构性风险。

And then I think one that a lot of actually like vendors don't say it out loud, but it's truly like the institutional risk of moving too fast.

我想我们都听过这样的故事：在沙盒环境里运行得很好的试点项目，一旦进入实际生产环境，如果你没有做好充分的准备，没有进行大量的测试，没有确保你在漂亮盒子里看到的东西在现实世界中真的有效，那可能会非常糟糕。

I think we've all heard the stories where you have a pilot that works wonderfully in a sandbox, but when it comes to actual production, it can be pretty disastrous if you don't do proper homework and don't do a lot of testing and don't ensure that what you see in a shiny box actually works in the real world.

对。

Yeah.

对。

Yeah.

听你这么描述挺有意思的，因为我觉得，你提到的那些考虑因素，确实存在，比如采用方面的问题。

It's it's interesting to hear you describe that because I think that, you know, there is certainly that those considerations you're talking about adoption.

但我觉得还有一点，我们在和丹尼尔的对话中也提到过，就是如果你真的在考虑部署这个技术，你想要构建的生态系统是什么样的？

But I think another piece, and we touched on this in a conversation we have with Daniel, is what's the ecosystem that you're trying to build if you are someone who's thinking about actually deploying this?

比如说，当然有一些事情需要你操心，但也有很多其他的人可以和你合作。

Like, there are things to keep you up and head, obviously, but there's also there's a there's a host of other people that you can work with.

这里说的人，指的是公司、超大规模服务商等等。

And by people, I mean, like, companies, hyperscales, etcetera.

明白。

Mhmm.

那么，当你考虑这个生态系统时，什么样的组合才是合适的呢？

So when you think about that ecosystem, what is the right mix?

更重要的是，在考虑如何构建那个生态系统时，什么样的组合是错误的？

And more importantly, what is the wrong mix when it comes to how you should think about building that ecosystem?

对。

Yeah.

所以我觉得，由于发展速度，我们现在正处于市场转型期，你看到的是，很多公司都在急于追求那个舞台时刻，你知道，比如一个大型的炫目公告、一个公关时刻、一份能让你登上头条的新闻稿。

So I think like the transition we are in right now with the market because of the speed of development, what you see happening, a lot of companies are just rushing for that stage moment, you know, like a big flashy announcement, a PR moment, a press release that gets you on a

我得先告诉你，我现在非常针对性地讲这个，你知道，或者这里的公关人员，别这么做。

I'm first very targeted right now, I gotta tell you, like, you know, or the comms guy here, don't do that.

Axe社区，你知道，就像只谈论那个公告，但事实是，这不仅仅是那个。

The Axe community, you know, like just talking about that one announcement, but truth is it goes beyond that.

而且我认为，目前市场上正在发生的事情实际上给客户带来了很多困惑。

And I think like what's happening right now in the market is actually creating a lot of confusion for customers.

当有这么多不同的供应商从这么多不同的角度声称做同样的事情时，就会造成很多困惑。

When you have so many different vendors from so many different angles claiming to do the same exact thing, it creates a lot of confusion.

而这种困惑也带来了客户的很多重复工作，对吧？

And with that confusion comes also a lot of duplication for customers, right?

所以我总是想，如果我要接触一个GSI或者技术联盟伙伴，我首先会问，我们能为客户共同承担的真实责任是什么。

So what I always think, like if I approach a GSI or a Tech Alliance partner, I always start with what's the honest stake that we can have together for our customers.

比如，我们真正最擅长的是什么，技术联盟生态系统的边界或交接点在哪里？

Like, where are we honestly the best in and where is the boundary or the handoff between the Tech Alliance ecosystem?

因为你需要最好的模型，最好的云服务商，但企业里还有一些不那么光鲜的部分，数据可能非常混乱，你需要一个数据伙伴来帮你理清这些，你明白吗？

Because you need the best model, you need the best cloud provider, but you also have like an unsexy part of an enterprise and your data can be really messy and you need a data partner in that that can help you figure it out, you know?

还有一些企业根本还没到那种超现代化、把所有东西都智能化的地步，数据太乱了。

And there are still some that are simply not in a place to be so ultra modern and agentify everything and it's just all too messy.

所以当所有供应商都声称能做同样的事情时，结果要么是客户做出妥协，而不是把所有伙伴都拉到桌边来完成全部工作，因为有些供应商会说，哦不，我觉得我可以再进一步。

So when you have all the vendors claiming that they can do the same thing, what happens is either customers make compromises and instead of having the whole 100% of the work getting done by bringing all the partners at the table, because there are vendors who are like, Oh no, I think I can go a bit further.

这部分我能做。

I can do this part.

客户最终可能只完成了70%的工作，而且很可能成本还更高。

The customer will be left with 70% of the work actually done and probably done at a mostly costly basis.

对。

Yeah.

对。

Yeah.

挺有意思的。

It's funny.

和丹尼尔聊的时候，我们探讨了一个想法，就是如果你做了错误的决定，这个错误决定可能有很多不同的方面。

The conversation with Daniel about sort of like the moment, we explored this idea of if you make the wrong decision, the wrong decision can has so many different sort of aspects to it.

这正是我特别感兴趣的地方，我想听听你怎么看，因为如果你考虑错误的答案，有一种错误答案是导致你技术堆得太多了。

This is where where where I'm where I'm I'm really interested in, like, what what do you have to say about this, because if you think about the wrong answer, there's the wrong answer that leads to you have too much technology.

你欠了太多债。

You have too much debt.

你有一堆东西散落在你的技术栈里，就像我说的，可能你只用了70%，剩下的30%可能花了你几百万美元却根本没用上。

You have all this stuff that's just, like, littered all over your stack, and you're just, like I said, 70% of that maybe you're using is 30% that probably is costing you, like, a couple million dollars you're just never using.

但还有另一种风险，就是如果你没有考虑到监管问题、合规问题会怎么样。

But there's also the other risk, which is what happens if you are not thinking about the regulatory concerns, compliance concerns.

所以从这个角度来想。

So thinking about that angle.

那么，Simi，对你来说，我们谈到了信心。

So, Simi, for you, we talked about confidence.

当你和金融行业的公司打交道时，这些公司受到严格监管，有很多合规方面的顾虑，每天都在想怎么保护我的数据？

When you're when you're dealing with, like, companies in the financial sector, these companies who are heavily regulated, who have a lot of compliance concerns, who are thinking every day about how am I protecting my data?

怎么保护我的系统？

How am I protecting my systems?

怎么保护我的客户？

How How am I protecting my customers?

怎么保护我的利润？

How am protecting, like, my bottom line?

当他们想有能力地应用AI时，他们在想什么？

What are they thinking about when they wanna apply AI with competence?

更重要的是，他们可以参考哪些基本原则来行动，以确保他们这样做的时候，是以正确的方式进行的？

And more importantly, like, what are the sort of, like, the first principles that they can look to act on to make sure that when they're doing that, they're doing it in the right way?

好问题。

Good question.

克里斯，我想先说，这个问题问得对。

Chris, what I'll start by saying is this is the right question.

因为我质疑的是像“我们如何自信地应用AI”这样的问题。

Because what I challenge is questions like how can we apply AI with confidence?

这个问题通常会被以另一种方式理解。

It is usually heard in a different way.

大多数人听到的是“AI现在够好了吗”。

Mostly people hear that is AI good enough yet.

对吧？

Right?

那是个模型问题。

And that's a model question.

在受监管的环境中，模型很少是瓶颈。

That's in regulated environments, model is rarely a bottleneck.

所以我会从三个原则开始。

So I would start with three principles.

第一条原则是，信心是设计出来的。

Principle number one is confidence is engineered.

不是凭空发现的。

It is not discovered.

我的意思是，你不需要等着对模型感到放心，也不需要到处去找一个更好的模型来让自己有信心。

By that, what I mean is you don't wait to feel comfortable with with the model, or you don't go out and shop around for a better model and feel confident about it.

信心是通过在模型周围搭建运营框架来建立的。

Confident is manufactured by building the operational scaffolding around it.

也就是说，有哪些控制措施？

That is, what are the controls?

可追溯性怎么样？

What is the traceability?

有哪些人工介入的环节？

What are the human touch points?

模型能做什么？

What the model can do?

这些就是让你能够负责任地部署它的那些东西。

Those are the kind of things just so that you can deploy it accountably.

对吧？

Right?

所以信心是制造出来的，不是找到的。

So confidence is manufactured, not found.

第二点我想强调的是，编排就是执行的治理。

So the second thing which I want to focus on is orchestration is the governance to execute.

我的意思是，你的政策不能只存在于没人看的PDF里，对吧？

By that, what I mean is that your policy cannot be living in a PDF which nobody reads, right?

它必须在系统内部强制执行，这样你才能确保它在什么数据上运行、谁在批准、以及在什么限制范围内。

It has to be enforced within the systems so that you can make sure that what is running on what data it is running, who is signing off on it, and within what limits.

对吧？

Right?

第三点也是最后一点，目标从来不是完全自主的AI。

And the third and the last point is the goal was never autonomous AI.

目标是负责任的AI。

The goal was accountable AI.

对。

Mhmm.

对吧？

Right?

所以每个决策都应该能够重现。

So every decision should be reconstructable.

而且对于监管者来说，如果不能重现，他们就会说这没发生过。

And as far as regulators are con concerned, they'll say it did not happen if you can't reconstruct it.

对吧？

Right?

所以最后，回答你的问题，从哪里开始？

So finally, to answer your question, where do I start?

绝对不是从模型开始。

Definitely never the model.

所以我的起点通常是终点。

So where do I start is I say usually I start at the end.

也就是说，今天必须满足什么条件，我才能理解这个决定，并且让审计人员明白我为什么做出这个决定。

What that means is what has to be true today for me to be able to understand the decision and make the auditors understand why I made this decision.

对吧？

Right?

然后从那里开始，你倒着往回构建。

And from there, you start building backwards.

对吧？

Right?

一旦你这么做，你就会发现，信心不再是一种盲目的跳跃，而是变得更加可理解，变成了一个工程问题。

And the moment you do that, you will realize that the confidence lifts up or stops from becoming a leap of faith and starts becoming more understandable and starts becoming an engineering problem.

而我们都擅长解决工程问题。

And we all are good at solving engineering problems.

我真的很喜欢这个回答，因为它把一件事非常清晰地聚焦了，那就是你不能把这个时刻误认为是追逐新鲜事物的冲动。

I really love that answer because I think it it puts something very sharply into focus, which is this you can't mistake this moment for shiny object syndrome.

你不能因为觉得某个模型看起来对，就随便选一个。就像安达拉之前说的，有时候它看起来合适只是因为更花哨，或者品牌宣传得好，或者其他什么原因。

You can't just run to a model because you think the model it know, to to the point Andrada was making earlier that it's it looks like the right one because it's flashier, because it you know, it's branded the right way or whatever.

你真的得认真审视一下，你打算围绕这个模型搭建什么样的基础设施，然后再做决定。

Like, you really have to take a hard look at, like, what is the infrastructure that you're looking to build around this and then make a model decision.

那么，既然这样，为什么编排这么难呢？

So in that vein, why is orchestration then so hard?

为什么偏偏这是最难的部分？

Why is that specifically the hardest part?

我想先说，以前模型是最难的部分。

What I would start by saying is model used to be the hardest part.

对。

Yeah.

对吧？

Right?

以前是这样。

Previously.

现在已经不是了。

It isn't anymore.

如今在大公司里，选模型只是个采购决定。

These days with large organization, model is just a procurement decision.

你可以今天下午就找到一个非常棒的模型，它能发挥出最佳性能。

You can go out and find a really amazing model which operates to the best of its capabilities this afternoon.

但你找不到的是：它被允许做什么、基于哪些数据、在什么限制内，以及当它行动时谁负责。

What you cannot find is the answers to what it is allowed to do, on what data, within what limits, and who is accountable of when it acts.

对吧？

Right?

所以这些都是编排的问题。

So that's all orchestration.

那么什么是编排呢？

So what is orchestration?

它是介于“模型能做什么”和“我们被允许做什么”之间的那一层。

It is the layer which sits between the model can do this and we are allowed to do this.

在受监管的环境中，这个差距非常大。

And within regulated environments, that gap is enormous.

对吧？

Right?

那我给你举个例子。

So let me give you an example.

一个简单的工作流程可能涉及客户数据、核心银行应用、第三方系统，甚至还需要人工审批。

A single workflow can touch customer data, can touch a core banking application, can touch a third party system, and can even have a human sign off.

对吧？

Right?

在所有这些环节中，智能部分可能只占10%。

So in all of that, the intelligent in an intelligence in any of that step is, let's say, 10%.

剩下的90%都是协调工作。

The remaining 90% is the coordination

嗯。

Mhmm.

就是护栏，还有贯穿其中的审计追踪。

Is the guardrails, is the audit trails across it.

对。

Mhmm.

对吧？

Right?

这里我有一个很简单的类比，我特别喜欢。

I have a very simple analogy to give around here, which I really like.

模型就像是发动机。

Like, model is the engine.

对吧？

Right?

但编排就像是刹车、方向盘和仪表盘。

But the orchestration is the brakes, the steering wheels, the the instrumental panel.

没有人会对一辆性能出色但没有刹车的车感到满意。

And nobody is impressed with a car which performs really amazing but does not have the brakes.

在我们的世界里，我们甚至没法让它上路。

In our world, we cannot even get that to road.

对吧？

Right?

所以我想说的是，当人们痴迷于模型时，我想把他们拉回来，让他们明白，模型并不是风险所在。

So what I would say is when people get obsessed over model, I would like to kind of get them back and ground them on that model is not where your risk is living.

模型也不是优势所在。

Model isn't where your advantage is living.

这两者都存在于编排之中，这就是为什么它很难。

Both of them are living in the orchestration there, and that's the reason it is hard.

对。

Yeah.

对。

Yeah.

对吧？

Right?

我喜欢这个想法，因为它又回到了我问你的那个问题，安德拉达，关于生态系统的问题。

I like that idea because it kinda gets back to the question I asked you, Andrada, about the ecosystem.

因为我觉得，当你想到那种，呃，而且我顺便说一句，我很喜欢那个汽车的比喻，有几个原因。

Because I think that when you think about sort of like the the and I love the car analogy, by the way, for a couple of reasons.

第一，我女儿刚拿到驾照，所以那是另一个话题了。

Number one, my daughter just got her license, so that's a whole other discussion.

但我真正喜欢的是，它确实触及了这种，几乎像是流水线的方式，零件开始沿着生产线移动。

But what I really like about that is it really does, like, speak to this whole like, there's almost, like, this assembly line approach where the the the the parts start moving down the line.

然后你心里想，好的。

And you're thinking to yourself, okay.

当我组装这个的时候，我需要怎么一层一层地、一块一块地把这些零件放进去，这样到最后，生产线末端的那辆车看起来就像底特律车展上展示的那辆，或者像光鲜宣传册里的那辆，等等。

As I build this, what are the plea how do I need to layer these pieces in bit by bit, piece by piece so that when I get to the end, it the car that is at the end of the assembly line looks like the one that is showed at the Detroit Auto Show or that's in a glossy brochure or whatever.

对。

Yeah.

所以从生态系统的角度来看，你觉得当人们在考虑构建这个的时候，无论是为你自己构建一个生态系统，还是为一个行业构建一个生态系统，他们有多频繁地遵循这种从结果出发、然后倒推的思路呢？

So thinking about this from an ecosystem standpoint, how often do you think when people are thinking about, like like like building this, whether it's like an ecosystem just for you or like an ecosystem for an industry, how often do you think people are sort of following this idea of, like, start with outcome and work your way backwards?

我其实觉得这才是健康的思考方式。

I actually think that's the healthy way to think about it.

我们刚才在聊现在市场上正在发生什么。

We were just talking like what's happening now in the market.

大家都在追逐那些花哨的东西，但实际上我看到的大多数成功案例，都不是从该用哪个模型、该用哪个编排工具、该用哪个云服务商开始的。

It's chasing that flashiness, but actually where I do see most of the success is when the conversation doesn't start with which model we should use, which orchestration we should use, which cloud provider.

更多的是像这样：这是我今天需要解决的问题。

It's mostly more like, this is the challenge I need to fix today.

我怎样才能把解决方案推向生产？

How do I get the solution to production?

而这就是每个参与者都有自己角色的地方，因为你需要最新的世界级模型来帮忙，你也需要云服务商，需要数据，需要编排，还需要GSI来让它在特定企业的运作方式中真正发挥作用。

And that's where every single player has a role to play because you need the latest world class model that can help with that, you also need the cloud provider, you need the data, you need the orchestration and you need the GSI to actually make it work in the way a certain enterprise operates?

因为我觉得那才是瓶颈所在。

Because I think that's where the bottleneck is.

你可以采用最好的技术，但如果它不符合公司的运作方式，你就无法推向生产。

You can adapt the best technology, but if it doesn't fit in the way a company operates, you're not going to get to production.

对。

Yeah.

对。

Yeah.

这引出了我想问你的一个新问题，Simi，就是当你和这些客户交流时，当他们考虑这些决策，比如构建他们的汽车时，谁在谈判桌上，谁又应该在那里？

And it really sets up a new question I wanted to ask you, Simi, which is when you're talking to these customers, when when they're thinking about they're they're they're making these decisions about, like, about building their car, who's at the table, and who should be at the table?

所以我觉得是这样的，你知道，以前采用云服务的时候，想想最初人们开始使用电子邮件，到后来开始使用云服务，把SaaS引入业务，那主要是IT部门的决策。

So I think the idea is like, you know, it used to be when you're adopting cloud, you know, think back think back to things like when people first started using email to when people first started rolling using cloud services, they started rolling out SaaS into their business.

主要是一个IT决策。

Primarily an IT decision.

涉及风险。

Risk was involved.

是的。

Yes.

合规性主要不是，而且主要是一个IT决策。

Compliance was primarily not and primarily an IT decision.

但因为这涉及面太广了，这个决定到底该由哪些人来参与呢？

But because this touches so much, who needs to be at the table for this decision?

这个问题问得好，因为如果我只给你一个答案，那肯定是错的。

That's a great question because the wrong answer would be if I give you any one of them.

那绝对是个错误的答案。

That'll be definitely a wrong one.

我在客户那里看到的最常见的失败模式，就是把责任当成烫手山芋。

The failure mode, which I see mostly in our customers, is when we treat ownership as a hot potato.

也就是说，谁最后接住它，谁就得负责。

That means whoever gets stuck holding it owns it.

这样一来，就会出现一些没人完全理解的代理，而且一旦出问题，他们很容易就推卸责任。

And that is where you get agents which nobody fully understands, and they can very easily, you know, kind of disown it when anything goes wrong.

对。

Mhmm.

对吧？

Right?

所以，刻意地，非常刻意地，我会说，所有权是一个需要分层建立的模型。

So, deliberately, very, very deliberately, I would say, the ownership is a very layered model which you need to put in place.

这意味着业务方负责结果。

What that means is the business owns the outcomes.

也就是说，这个代理要做什么，以及它看起来是什么样的？

That means what does the agent do and what does it look like for this agent?

IT部门负责能力。

The IT owns the capabilities.

我能做到吗？

Can I do this?

我该怎么做？

How can I do this?

我如何在系统之间进行集成？

How can I integrate between systems?

以及它的性能如何？

And how does it perform?

而风险和合规部门将负责界定边界。

And risk and compliance will own the boundaries.

也就是说，它被允许做什么，以及我们如何证明这一点？

That means what it is allowed to do, and how can we prove that?

对吧？

Right?

而且我告诉你，这三个合作伙伴，没有一个可以当甩手掌柜。

And let me tell you this, all these three partners, none of them can be a silent partner.

他们必须积极发声。

They have to be vocal.

他们必须大声说出来。

They have to be loud.

也就是说，一个没有业务负责人的智能体就是一个科研项目。

That means an agent without a business owner is a science project.

一个没有IT支持的智能体，是无法维护的。

An agent without without IT, it is unsupportable.

没有风险管控，就无法上线。

Without risk, it is unshippable.

对吧？

Right?

我就是这么看的。

That's how I look at it.

实际上，我采用的模式，就是给客户提建议，这跟我们今天运作高风险项目的方式非常接近。

And then the model the model I actually use is giving suggestions to the customer is very close to when we how we run high stake projects even today.

也就是说，业务方要有一个负责任的负责人，真正对结果负责，对吧？

That means you want an accountable owner from the business side who is really responsible for the outcomes, right?

他们不能把这个责任推给别人，而且必须在风险团队设定的、IT 执行的边界内运作，对吧？

They cannot delegate that away and they need to operate within the boundaries which are set by the risk teams and which IT enforces, right?

所以我想说的是，当你准备在生产环境中部署智能体时，你应该能够非常清晰地划分业务、IT 和风险之间的所有权和问责制。

So what I would say is when you are ready for deploying the agents in production, you should be able to very clearly map the ownership and the accountability between business, IT, and risk.

如果你做不到这一点，那你就还没准备好部署它。

And if you are not able to do this, you're not ready to deploy it.

我们已经讨论了各个组件。

So we've talked about the components.

我们也讨论了生态系统。

We've talked about the ecosystem.

我们还讨论了基本原则。

We talked about sort of first principles.

现在我们要进入实际操作的阶段了。

So now we're getting to the point where actually you're go you're gonna press the button to go on this.

好的。

Okay.

所以我想先问你一个问题，Andrada，那就是有没有客户已经在正确地做这件事了？

So I wanna I'll ask you this question first, Andrada, which is there are are customers out there who are doing this the right way.

是不是只要采纳我们的建议就行了？

Is it just as simple as just take our advice?

我们有没有遗漏什么？

Did we miss anything?

有没有一些客户今天应该做的事情，能让它们成功，而我们还没提到的？

Is there something customers should be doing to make them successful today that we haven't talked about?

所以我觉得，从各个行业来看，我看到了三个主题，决定客户是成功还是挣扎，或者被困在沙盒里。

So I think, like, across the board and industries, three themes that I see emerging, what makes a customer successful versus stumbling or like stuck in the sandbox.

对。

Yeah.

我觉得第一点是，AI战略不能只由一个孤立的AI团队来推动，就像那些酷孩子，他们只做自己的事情，因为如果有些人只做酷的事情，也会引起很多不满，对吧？

I think number one is when this AI strategy is not just driven by an AI group that isolated, that's just like the cool kids, they're just doing their own thing because there is also a lot of resentment if somebody just isolates the cool stuff, right?

而且你还必须确保组织的其他部分都准备好，并且兴奋地和你一起踏上这段旅程。

And you also have to make sure that the rest of your organization is ready and excited to go on the journey with you.

所以我实际上认为，这给公司领导层带来了紧迫感，让他们聚在一起，就AI战略达成一致，因为这和我过去经历的其他转型有点不同，那些转型只涉及公司的一个部分。

So I actually think it creates sort of this urgency for the leadership of a company to all come together and agree on this AI strategy because it's a bit different than any other transformation I think I've experienced in the past where it touches just one piece of the company.

它实际上涉及整个公司。

It actually goes across the board.

所以每个人都必须感到参与，感到自己有发言权。

So everyone has to feel engaged, has to feel like they have a say.

所以大家一起做，这是第一点。

So just doing it all together, that's number one.

我觉得我们聊了很多关于模型的事，但别只盯着模型，别误会，我很喜欢模型。

I think we talked a lot about models, like just don't focus on the models and don't get me wrong, I love the models.

我晚上都在玩它，比如设置我的Apple Mini，做各种事情，我知道不只是我这样，但对一个企业来说，只关注模型是不够的。

I'm spending nights just playing with it, like setting up my Apple Mini and like just doing everything and I know I'm not talking just about myself, but just focusing on the model is not enough for an enterprise.

真正难的地方，也是成功所在，是把运营模式整合起来，而不只是今天能用什么酷炫的模型。

Where the hard stuff is, is actually where the success lies, like bringing together the operational model around it, not just what's the cool model that I can use today.

再说回我在节目开头说的，事情变化太快了。

Again, going back to what I said at the beginning of the episode, things change so fast.

你不能只押注在一个模型上。

You cannot just place your bets on just one model.

对。

Yeah.

还有第三点，你也提到了，就是大家都急着走向自主，比如让智能体全权处理。

And I think third is, you mentioned it as well, like there is this rush going to like autonomous, like an agent doing it all.

而且你想想模型的演变，我们一开始是，来挑战一下这个模型，看它能不能给我一个答案。

And if you think also about the evolution of the models, we started with like, let's challenge this model to see if it has an answer for me.

然后我们变得更聪明了，我们说，让模型假装它有这些技能，能给我一些东西。

Then we got smarter and we said, Let's have the model pretend they have these skills and they can give me something.

它确实可以，但说实话，那还是理论上的。

They can, but truth is, it's still going to be theoretical.

现在你看行业里发生的事，我们已经不只是为了提示而提示了，你懂我的意思，我们更多是为了循环而提示，而且人必须参与其中，因为在一个企业里有效的方法，在另一个企业里不一定有效。

And now if you look at what's happening in the industry, we went past the prompting just for the sake of prompting, you read my mind, but we're prompting more for the loops and humans have to be part of it because also what works in one enterprise doesn't work in another one.

事实是，AI会给你一个关于策略的建议。

Truth is like AI will give you a recommendation on what the strategy could be.

但如果没有具备相应机构知识的人来验证这个想法，并真正把它落地，那你还差得远呢。

But if you don't have the humans with the right institutional knowledge to validate the thinking and to actually bring it to life, you are not far along the way.

对。

Yeah.

听你这么说很有意思，因为就像你说的，我们都在用这些模型做实验。

It's interesting to hear you say, because I've like you're saying, we're all experimenting with these models.

我们都在玩Claude、GPT或者Gemini之类的。

We're all playing with Claude or GPT or Gemini or whatever.

我们团队在罗马尼亚做过一个练习，花了一天半的时间基本上就是在玩这些模型。

We had this exercise as a team where we were in Romania, and we were we spent a day and a half basically playing with models.

当时想的是，我们能做些什么？我是做市场营销的。

And the idea was like, what are the things that we could do I work in marketing.

在市场营销方面，有哪些事情可以用AI来改造？

What are things that we could do in marketing that we could potentially transform using AI?

然后我们每个人都有一个个人练习，基本上就是说，好。

And so we all had this, like, individual exercise where you you were basically said, okay.

选一个你每天都会做的流程，你的团队每天都会用的，然后看看能不能把它变成……

Pick a process that you do every single day that your team uses every single day, and try to figure out if that's something you could actually turn into.

你可以构建一个代理，让代理来做这件事。

You could build an agent, that agent can then do it.

最有趣的是，是的，你可以把它放进提示里，问几个问题，然后它就会给你一个建议。

And what's really interesting is, yes, you could put that you put it in the prompt, and you ask a couple questions, and you and it gives you a recommendation.

然后你开始去试探它。

And then you start to poke at it.

对。

Mhmm.

然后它也会反过来试探你。

And then it pokes back.

我觉得这才是真正让人兴奋的地方，因为到了那个阶段，你才能真正开始理解从试点到生产之间，成功和失败的区别在哪里。

And that's where I think it gets really exciting, because then you get to the point where, like, that's where you can really start to understand the difference between success and failure from pilot to production.

我自己经历过这个过程，所以特别想听听你的看法，Simi。因为你经历过这种体验，我猜你可能跟客户聊过几十次，甚至上百次了。

So having gone through that experience myself, I'm really curious to hear, like, what you would have say me Simi, because you've been through this experience, I imagine probably, you know, dozens of times, just done hundreds of times in talking with customers.

那些能成功扩展AI的组织有什么特点？

What are the organizations that will successfully scale AI?

他们怎么才能摆脱那种试点阶段的困境？

How will they get out of that pilot, like, purgatory?

他们怎么才能进入生产阶段？

How will they get to production?

对。

Yeah.

我见过很多次了。

I've seen it multiple times.

所以我想说三点。

So I would say three things.

而且我要说清楚，这三点都和AI无关。

And let me be very clear, none of them is about AI.

对吧？

Right?

第一点是他们做事的顺序。

So the first thing is the order in which they did all the things.

也就是说，那些卡在试点阶段的团队，他们先建能力，然后再试图加上治理。

What that means is the teams which are stuck in pilot, they build the capability first, and then they try to put the governance around it.

结果发现，到那时架构已经因为一些假设而变得非常僵化，而这些假设并没有和审计团队协调过。

To realize that the architecture is so hardened at that point based on the assumptions, which are not coordinating with the audit teams.

所以现在他们要重建它。

So now they're going to rebuilding it.

对吧？

Right?

而那些已经规模化并设计得很好的团队，他们先设计审计、控制和参数，然后在这些参数范围内构建能力，对吧？

And the teams which have scaled and designed, scaled well, they have designed the audit, the controls, the parameters first, and then they are building the capability within those parameters, right?

所以你看，组件是一样的。

So if you look at it, the components are same.

顺序不同，结果就完全不同。

The orders are different, and the outcome is totally different.

所以这就是我要说的第一点，关于你首先做什么，以及你做事的顺序是什么。

So that is the first thing which I'll say in the sense of what you are doing first and what is the order of order of things on how you are doing it.

你的重点顺序是什么？

What is your focal point of orders?

对吧？

Right?

第二点我想说的是，那些把这件事当作运营模式变革，而不是技术项目的公司。

The second thing which I'll say here is that companies who treated this as an operating model change versus a technology project.

对吧？

Right?

我的意思是，试点项目可以证明技术是可行的。

And by that, what I mean is a pilot can prove the technology.

它证明了技术是有效的。

It proves that the technology works.

但真正要推广开来，那就是组织层面的变革了。

But now scaling that is an organizational change.

这意味着什么呢？就是说，谁来负责这件事？

What it means is, like, who owns it?

责任归属在哪里？

Where is the ownership?

控制权在哪里？

Where is the controls?

谁批准它？

Who approves it?

谁在生产环境中监控它？

Who monitors it in production?

所以，如果你想把组织问题当成技术问题来解决，你永远也解决不了。

So if you are trying to make this organizational problem a technology problem, you'll never get there.

对吧？

Right?

你会一直做四五个试点项目，但你还是原地踏步，因为你永远无法通过一项新技术来扩展规模，而这项新技术是你之前无法通过首次成功的技术来扩展的。

You'll be running four four or five pilots, and you'll still be there because you're never gonna scale anything by a different technology, which you were not able to scale by a first successful technology.

这就是我要说的第二点。

So that's the second thing which I'll say.

然后第三点，这一点有点被低估了，但我称之为“翻译网格”。

And then the third thing, and this one is a little underrated one, but I call it the translation grid.

我的意思是，模型的瓶颈并不在那里。

By that, what I mean is that the bottleneck around around the model is not there.

有人会训练模型，让模型变得智能，这些人已经存在了。

There are people who will train the model, who will make the model intelligent, those people are existing.

但那些投资于过渡层人员的公司，这些人能坐在风险人员和机器学习工程师之间，让他们互相沟通，这才是关键。

But the the companies who invested into those transitional layers who can sit between the risk people and the ML engineers and have them communicate with each other is the key.

对吧？

Right?

所以那些能够投资于这些人员的组织，扩展得相当好。

So the organizations who were able to invest in those people well scaled pretty well.

这就是我要说的三件事。

So those are the three things which I'll say.

但如果我们快进几年，公司之间就不会再有差距，说谁有好的AI或坏的AI，因为这对每个人来说都是一样的。

But now, if I fast forward a couple of years, there won't be any gap where companies will have a good AI or a bad AI, because that's going to be same for everyone.

每个人都能接触到惊人的模型。

Everyone is gonna have access to amazing models.

但区别在于，谁能让它成为一个组织性的可扩展项目，而谁还把它当作IT部门的转型项目。

But what'll differentiate them is who was able to make it an organizational scalable program versus who were still sitting in IT as a transformation project.

对。

Yeah.

对。

Yeah.

我很喜欢这个回答，因为它确实回到了我们一开始聊的那个点，那就是这是一个变革性的时刻，是大写的变革。

I love that answer because it's a real it really does get back to, I think, where we started this conversation, which was this is this is a transformational moment, it's transformational with capital t.

确实如此。

It really is.

这不只是，你知道，像等等之类的。

It's it's not just, you know, like, etcetera.

你不只是安装一个新软件。

You're not just plugging in a new piece of software.

你不只是增加一个新合作伙伴。

You're not just adding in adding a new partner.

你不只是，你知道，招一个新团队，或者只是让风险和IT部门第一次碰面。

You are you know, you're not just hiring in a new team or just, like, having it having, you know, having risk and IT meet for the first time.

你把这些事情全都做了。

You're doing all of the things.

是的。

Yeah.

而且你是同时做的，是有意为之的。

And you're doing them all at once, you're doing them with intention.

罗斯，我觉得这是我想传达的一个非常有力的信息。

I think that's a really powerful message I wanna get, Ross.

我真的很感谢你们。

I really wanna thank you guys for this.

这真的很有趣。

This is really fun.

非常棒的对话。

Really great conversation.

这里有很多东西大家可以借鉴。

There's a lot of stuff here people can take away.