自我提升的公司 | Kavak的AI行动指南

我今天在代币上的投入比在知识型员工上更多。

I'm investing more today in tokens than in knowledge workers.

我们可以打造出超越人类能力的智能体。

We could build superhuman agents.

这意味着在所有关键维度上，我们的智能体都会胜过我们曾经雇佣过的最优秀的人类。

This means that by every dimension that matters, our agents would outperform the best human we had ever hired.

那些最有雄心的公司听到这里后，会决定跟进，比如你们就决定为每个客户都构建一个智能体。

The most ambitious companies listening to this will decide to follow suit, which is you decided to build an agent per customer.

是的。

Yes.

每天，会有十万到二十万个智能体被专门创建出来服务这位客户，每个智能体都运行在独立的虚拟机上。

Every day, between one hundred and two hundred thousand agents get instantiated specifically for this customer with its own virtual machine.

很多人都在担心未来的组织会是什么样子，以及人类在其中将扮演什么角色。

There's a lot of people worried about how the organizations of the future are gonna look like and the role that humans are

将会扮演的角色。

gonna play.

如果你还没经历过恐惧，说明你还没真正感受到它，那你肯定还没试过AI。

If you haven't faced fear before, you haven't felt it, then you haven't tried AI.

我们在Kavak内部启动了一个叫作‘绝地学院’的项目。

We launched a program inside Kavak that's called the Jedi Academy.

从CEO到AI工程师，再到机械师，我们全员培训。

From the CEO to, like, AI engineers to mechanics, we train everyone.

六周之后，他们就能把最先进的智能体投入生产环境。

And after six weeks, they launch state of the art agents to production.

你对未来的创始人，或者第一次创业的人有什么建议吗？

What advice do you have to future founders or first time founders that might be listening?

现在行之有效的方法是

What works right now is

大多数公司都在问怎么把AI加到组织里。

Most companies are asking how to add AI to the organization.

Kavak则提出了一个更激进的问题。

Kavak asked a much more radical question.

如果我们从零开始用AI创建公司，我们会做什么？

What would we build if we were starting the company from scratch with AI?

Angela Strange和Gabriel Vasquez与Kavak的首席产品兼AI官Alejandro Ayala坐下来，探讨了公司押注于围绕智能体重建自身后发生了什么。

Angela Strange and Gabriel Vasquez sit down with Kavak's chief product and AI officer, Alejandro Ayala, to unpack what happened when the company bet on rebuilding itself around agents.

如今，每天可以实例化数十万个智能体，处理从销售和融资汽车到维护长期客户关系的一切事务。

Today, hundreds of thousands of agents can be instantiated each day, handling everything from selling and financing cars to maintaining long term customer relationships.

他们讨论了Kavak为什么拆掉了一个已经运作良好的智能体架构，重新开始，以及评估如何成为加速发展的基础，还有当智能体不仅仅为人类工作，而人类有时也为智能体工作时，会发生什么。

They discuss why Kavak tore down an agent architecture that was already working to start again, how evals became the foundation for moving faster, and what happens when agents don't just work for humans, but humans sometimes work for agents.

欢迎回到ACC播客。

Welcome back to the ACC Podcast.

今天，我们有Kavak的AI负责人Ale Maza。

Today, we have Ale Maza, the head of AI at Kavak.

我们今天要讨论的是，Ale在Kavak内部领导的转型，将其转变为一家AI原生的公司。

We're gonna discuss today the transformation that Ale led within Kavak to turn it into an AI native company.

谢谢你，Ale，今天能和我们在一起。

Thank you, Ale, for being with us today.

谢谢邀请。

Thanks for having me.

在加入 Kavak 之前，你经营过一家叫 Opi Analytics 的公司。

Before starting at Kavak, you were running a company called Opi Analytics.

没错。

That's right.

而且你在 ChatGPT 出现之前就已经非常关注 AI 了。

And you were very much into AI before ChatGPT.

能跟我们分享一下这段经历吗？

You wanna tell us a little bit about that journey?

好的，当然可以。

Yes, yes, of course.

那时候我们还叫它机器学习，用的算法家族也不太一样。我们创立这家公司时抱着一个非常宏大的愿景：相信新的机器学习模型会强大到足以解决任何复杂的问题。

Well, we called it machine learning back then, it was a different family of algorithms, and we founded a company with this very ambitious vision there that new machine learning models would be so powerful that they could solve any complex problem.

那是 Transformer 出现之前的事情了，大概是 2013 年。我们就按这个思路开始创业，我觉得当时可能超前了十年左右，但确实做成了一家很棒的公司。

This was pre transformers, This was like 2013, so we started building the company that way, and I think we were like ten years ahead of time, but we built a great company.

我们曾为1450家公司提供风险算法、物流、预测和营销方面的服务。但真正让我们看清方向的是Transformer技术的出现，以及ChatGPT带来的那个转折点，它让我们明确意识到，现在可以创办一家全新的公司，甚至改变整个创业的方式。

We served fourteen, five hundred companies around risk algorithms, logistics, forecasting, marketing, but really the power of what Transformers and then the chajapiti moment, when he arrived, made things very clearly that we could now build a whole new company and way of building companies.

所以我们加入Kavak，和Carlos一起打造这家公司。

And we joined Kavak to and Carlos to build that.

太棒了。

Amazing.

没错。

Right.

所以这期播客的大部分时间，我们都会深入聊聊你是怎么发现 Kavak 这个项目的。

So we're gonna spend the bulk of this podcast talking about exactly how you've identified Kavak.

不过咱们先简单开场一下，Kavak 到底是做什么的？你在里面担任什么角色呢？

But maybe just to start, what does Kavak do, and what is your role there?

Kavak 最初是一个二手车交易平台。

Kavak started out as a used car marketplace.

简单来说，我们收购车辆，翻新之后再进行销售和提供金融服务。

So we buy cars, we refurbish them, and then we sell them and finance them.

但要做到这一点，我们还必须建立一家金融科技公司、一家物流公司，还有CARFAX，基本上拉美地区根本没有这些基础设施，所以我们不得不把整个链条垂直整合起来，才能以正确的方式服务客户。

But to do that, we also had to build a FinTech and a logistics company and CARFAX and basically all the infrastructure for this to work didn't exist in LATAM, so we had to build everything vertically so we could serve our customers the right way.

我想先从一个框架说起，就是整个架构是什么样的。

I'm gonna sort of start with the framing of what the architecture looks like.

比如一个消费者进来，说，我想卖掉我的车。

So a consumer comes in and says, I wanna sell my car.

他们会接触到多少个代理人？

Like, how many agents do they touch?

这个流程是怎么设计的？

What's the harness look like?

给我们讲讲你是怎么设计这个的细节。

Ground us in how you design this.

对。

Right.

所以我们把整个公司的赌注押在了转型上，变成一家由人工智能代理驱动的公司。

So we bet the company in transforming to a company run by agents.

我们问自己的问题是，如果到了2035年，用相当于Fable 10或GPT-ten级别的人工智能来重新打造Kavak，我们会怎么建？

The questions we ask ourselves is, how would we build Kavak in 2035 with Fable 10 or GPT-ten level intelligence?

实际上，那样的公司形态跟我们当时建起来的，或者当时的样子会有很大不同。

And actually that company looks very different than what we had built or what we had back then.

所以，现在客户一进来，系统就会为这位客户专门生成一个Agent，并分配独立的虚拟机。

So when a customer comes in right now, Agent will get spawned specifically for this customer with its own virtual machine.

它会记住该客户过去几年与Kavak的所有互动细节，比如浏览过网页的哪些内容，两年前打过什么电话，把所有信息都记在记忆里，然后制定策略，设定长期目标以最大化客户的终身价值，并采取一切必要措施让客户满意，并在不同时间段内引导他们使用我们的各类产品。

It'll remember years of interaction of these customers with Kavak, what they visited in the webpage, or the call they had two years ago, remember everything in its memory, come up with a strategy, and set a long term goal to maximize the lifetime value of this customer, and do whatever it takes to make the customer happy, and convert them into all our different products across time.

这是一套全新且极具突破性的规模化架构。我觉得吧，因为现在大家还在用专家系统来搭建多智能体系统，而我们意识到，押注那些有明确硬目标的长周期智能体——而不只是跑流程——才能真正最大化客户满意度，也自然能提升他们的终身价值。

And this is a completely new and groundbreaking architecture at scale, I think, because people are still building multi agent system with experts, and we realize to bet that long running agents with hard goals, not just workflows, could maximize our customers' satisfaction and obviously their lifetime value.

好的。

Okay.

那我们就深入聊聊其中的细节。

So we're gonna jump to the nuances of that.

不过，这可能跟很多公司说的不太一样，他们说嘿。

But maybe versus many companies that say, hey.

我们想实现智能体化，他们尝试了一些工作流。

We wanna be agentic, and they try some workflows.

是的。

Yes.

你们就直接放手去做了。

You guys took the just rip.

我们不得不让这个方案奏效。

We had to make this work.

你们不得不大幅缩减规模。

You had to downsize dramatically.

有一年时间都没成功。

It didn't work for a year.

对。

Right.

那你愿意聊聊吗？显然，你们为了让它奏效，调整了很多东西。

So do you wanna talk through obviously, you had to tune a lot of things to make that work.

比如，描述一下当时的工具环境，你们用的是哪些模型，具体一点。

Like, describe the harness at that time and what models you were using and sort of specifically.

对。

Yeah.

我们当时必须做出三个关键决定。

There were three main decisions that we had to make.

第一个决定，也是我觉得现在很多公司卡住的地方，就是第一反应是：好，我们采用AI吧。

The first, and this is where I think many companies are stuck right now, is the first instinct is, okay, let's adopt AI.

然后基本上保持原有结构不变，直接把ChatGPT或Claude交给团队，结果效率没有提升，客户的问题还是老样子，什么都没发生，对吧？

And you basically leave your structure as it is, and just give ChatGPT or Claude to your team, and then there's no efficiencies, your customers have the same problems, and nothing happens, right?

所以你需要围绕这些智能体和未来的能力，重新设计整个公司。

And so you need to redesign your whole company around the agents and around the future capabilities.

这意味着真的要重建大部分API，重建你的系统，让智能体能够使用它们来执行任务。

And this means really rebuilding most of your APIs, rebuilding your system so the agents can use them to perform.

然后你需要开始生成数据和反馈循环，来微调这些智能体。

Then you need to start generating the data and the feedback loops to fine tune these agents.

要让它们真正发挥作用，唯一的办法就是去教它们，那具体怎么教呢？

The only way to really make them work is if you teach them, like how do you teach them?

把它们放到真实场景里，放到客户面前，收集数据、获得评估，然后再训练你的智能体。

You put them out in the open, you put them in front of customers, you get that data, you get those evals, and then you train your agents.

这是我们押注的第二件事，就是要打造出超越人类的智能体。

And this is the second bet that we made, that we could build superhuman agents.

这意味着在每一个关键维度上，比如转化率、客户终身价值、客户体验，我们的智能体都要超过我们曾经雇过的最优秀的人。

This means that by every dimension that matters, like conversion, lifetime value, customer experience, our agents would outperform the best human we had ever hired.

而且我们会让它们去处理最难的问题。

And we'd put them in front of the hardest problems.

最后，你要开始改变衡量公司成功的方式。

And finally, you start to change how you measure the success of the company.

Kavak 是一家交易型公司。

Kavak was a transactional company.

我们过去衡量的是买了多少辆车、卖了多少辆车、需要采购多少刹车片。

We used to measure how many cars were bought, how many cars we sold, how many brake pads we needed to buy.

然后我们转型成了一家以客户关系为核心的公司，现在我的数据库里有一千万客户，而且大部分客户都有专属的客服人员，他们的任务就是最大化客户的终身价值。

And we moved to a relational company, where now I have 10,000,000 customers in my database, and I have agents assigned to most of them with the task of maximizing their lifetime value.

现在我们卖的是汽车、个人贷款，还有那些高客单价的产品。

Now, we're selling cars and personal loans and very high ticket items.

所以只要激活这1%的客户，如果做得好，那就是几亿美元的收入。

So just activating 1% of this customer base, it's hundreds of millions of dollars if we do it the right way.

所以这个赌注对我们来说是有意义的，因为我们的行业、我们的客单价，而且说到底，客户买二手车的时候，需要跟公司建立信任。

So it's a bet that made sense for us because of our industry, because of the ticket, and because at the end of the day, customers need to build trust with a company because they're buying a used car.

而建立信任的方式就是了解他们，并且规划、培养一段长期的关系。

And the way to build trust is to know them and to plan and nurture a long term relationship.

对。

Yeah.

Ale，我想深入聊一下，就是关于评估和客服演示之间的区别。

Ale, I just wanted to double click on something, evals over agent demos.

对。

Yeah.

你可能经常收到各种代理的推销，而且现在做东西比以往任何时候都容易。

You probably get pitched a lot of agents, and it's never been easier to build things like before.

但问题之一是，你们是怎么评估这些代理的？

But one of the questions is, like, how do you guys go about evaluating this?

因为不是每个人都会在90%的客户互动中去测试它们是否真的有效，而你们，我相信大约98%的互动现在都是由代理处理的。

Because not everybody tests them across 90% of the customer interactions to see if they're really working, and you guys, believe it's about 98% of the interactions or something like that are now handled by agents.

是的，完全正确。

Yes, totally.

所以给你一个规模的概念，比如96%的互动都是由代理处理的，所以那里没有人工。

So to give you a sense of the scale, like 96% of all interactions are handled by agents, so no humans there.

95%的交易完全由代理处理。

95% of all transactions are completely handled by agents.

当然，当你取车的时候你会见到真人，比如有人在那里给你钥匙，但其余的旅程体验都是由代理处理的。

Obviously you meet a human when you pick up your car, like there's someone physically there to give you the keys, but the rest of the experience of the journey is handled by an agent.

每天，有十万到二十万个代理被实例化。

Every day, between one hundred and two hundred thousand agents get instantiated in a day.

它们醒来，工作有时候三分钟，有时候八小时，有时候三天，然后为下一个任务设好闹钟，再回去睡觉。

They wake up, they work sometimes for three minutes, sometimes for eight hours, sometimes for three days, and they set an alarm clock for their next task, and they go back to sleep.

所以这个规模真的非常惊人，而且确实在运作。

So the scale of this is just amazing, and it's working.

那么，你们是怎么让它在这么大规模下运转的呢？

Now, how do you get this to work at scale?

答案你提到了，就是评估。

And the answer you mentioned it is evals.

我喜欢行动得非常快。

I like to move extremely fast.

但要想快，你必须有刹车。

But in order to move fast, you need to have brakes.

想象一辆车，只有刹车靠谱，你才敢踩油门。

Imagine a car, you'll hit on the gas just if you have the right brakes.

AI非常强大，我见过很多公司搞砸了，就是因为他们没有合适的刹车，所以只能慢慢来。

And AI is super powerful, and I've seen many companies get this wrong because they try to go slow because they don't have the right bricks.

所以我就反着想了。

So I thought about it the other way around.

比如，我们能多快呢？

Like, how fast can we go?

这个嘛，要看我们评估的质量了。

Well, it depends on the quality of our evals.

这里有个很好的经验法则：我们在构建评估上投入的工程师时间、代币和金钱，和构建智能体几乎一样多。

So a good rule of thumb here is we spend about the same amount of time, engineer time, tokens, and money on building the evals than building the agents.

这样的话，你才能不断进步，而不是把评估当作事后才补充的东西。

And this is how you get better and better and better, not letting evals as an afterthought.

那我们衡量什么呢？

So what do we measure?

首要的是业务结果。

First and foremost, the results for the business.

如果客户满意，他们就会买车，贷款获批，或者把车卖给我们——这就是第一道检查。

If my customer's happy, they'll buy a car, they'll get their loan approved, they'll sell a car to us, and that's the first check.

有没有转化？

Did it convert?

大多数问题都出在这儿。

And that's where most things break.

我看到有些公司只统计通话次数、通话时长，或者一些表面上的KPI，这些虽然能提供一些信息，但实际用处不大。

I see companies measuring number of calls, or minutes during the call, or some superficial KPIs that give you some information, but that doesn't really work.

关键问题是，这位客户到底有没有完成转化？

The important thing is, did this customer convert?

它有没有给客户带来价值？

Is it bringing value to the customer?

客户过段时间还愿意再次和我们互动吗？

And is the customer happy to reengage with us after a while?

一旦把这些评估体系对接好，剩下的就是优化合适的智能体架构，赋予智能体相应的能力，从而实现规模化扩展，服务数百万客户。

And once you get those evals connected, then it's just optimizing the right agentic architecture and giving the agent skills to scale this and cater to millions of customers.

这真的非常令人惊叹。

It's really, really amazing.

说到这个，就是当你建立了正确的评估体系，系统运转正常的时候。

And related to this is like, okay, so you create the right evals, you know, it's working.

不过你知道吗，有些公司还是有点保守，不敢直接把它们推给客户去执行那些最能产生价值的工作，比如你们这种情况下的销售环节。

You know, some people, some companies still feel a little bit risk averse in putting them in front of the customers and being able to perform the highest leverage task, which in your case would be selling.

你们的智能体真的能直接跟客户做销售吗？

Your agents really sell to customers?

是的。

Yes.

所以我们从来没做过客服或售后支持类的智能体，我们做的是销售类智能体。

So we never built customer support or customer service agents, we built like sales agents.

在拉美地区卖车真的特别难。

It's extremely hard to sell a car in Latin America.

你想想看，想买辆车的人面对两万个SKU选项，简直无从下手。

So imagine someone wanting to buy a car, they can choose amongst 20,000 SKUs.

接下来他们还得选金融方案、走贷款流程，再搞定保险和保障条款，而且很多人还要置换旧车，所以我们也得给那辆旧车估价。

Then they need to pick financing and go through the financing process, insurance and coverage, and then they're probably trading in their car, so we need to quote that car.

所以这是一个流程，如果由人来操作，或者像Kavak在2020、2021年那样做的话，你需要精通15个不同的环节，并且要有15个不同团队里的15位专家。

So it's a process that if someone does it, or the way Kavak did it back in 2020, 2021, was you need to be extremely good at 15 different things, and have 15 different experts in 15 different teams.

通常这个人得去跟融资专家、汽车顾问专家、采购专家以及保险专家分别沟通。

And usually the person would go and speak with the expert in financing, the expert in car advisory, the expert in buying, the expert in insurance.

然后他们再打包方案，完成购车。

And they'll build a package and buy a car.

这做起来非常难。

That's extremely hard to do.

我们第一步做的是，看看能不能让每个环节的智能体都比专家更厉害，然后把它们整合起来，打造一个在保险、融资等方面都精通的超级专家，最后把这个超级专家推给客户。

But the first thing we did was, okay, can we get an agent to be better than the expert in each of these things, and then put it together and have a mega expert that's an expert in insurance, financing, etcetera, and that's who we put in front of the customer.

客户的体验非常棒，我们把智能体放在前端后，NPS和客户满意度分数翻了三倍。一开始它的转化率比人工团队高出50%，现在更是达到了2.1倍。

So the experience for the customer is amazing, we tripled NPS and customer satisfaction score by putting the agent in front of the customer, and first it converted 50% more than our human team, and now it's converting over that 2.1x more.

这完全是一家不同的公司了。

So it's a completely different company.

所以你们的智能体是更厉害的推销员。

So your agents are better sellers.

确实好得多。

Totally better.

你能明白这个逻辑吧？

And you get this, right?

因为他们是专家，而且特别有耐心，了解你的所有历史记录，能为你做长远规划，还永远不会累。

Because they're experts, and they're infinitely patient, and they know all your history, and they can plan for the long term, and they never get tired.

如果他们犯了错，就会吸取教训，第二天不只是他们自己，其他20万位代理都会从这次错误中学到东西。

And if they make a mistake, they learn it, and the next day, not just them, but the other 200,000 agents will have learned from that mistake.

这就是我们建立的反馈循环，它直接体现在我们的增长、成果以及客户满意度上。

So that's the feedback loop that we engaged, and that's showing in the growth and results and satisfaction of our customers.

说到 Kavak，我觉得有两点特别酷。首先，大家现在对 AI 处理客服已经习以为常了。

Of the well, two of the very cool things I think about Kavak is I think the world has gotten comfortable with AI can do customer service.

但要把这件事做好，依然非常困难。

It's still very hard to do well.

不过，就像 Gabe 提到的，很多人还是觉得，客户大概不愿意用 AI 来购买高价值的商品吧。

But, you know, as Gabe said, there's still a view that, well, customers aren't gonna wanna buy expensive things

从

from

AI，而你正在证明他们是错的。

AI, and you are proving them wrong.

是的。

Yes.

下一层是，嗯，你实际上无法用AI端到端地做受监管的金融服务。是的。

The next layer on that is, well, you're not actually gonna be able to do regulated financial services Yes.

端到端地用AI。

End to end with AI.

但如果你看看你们在做什么，你们正在为信用记录薄或没有信用记录的客户承保。

But if you walk through what you're doing, you are underwriting a thin or no file customer

客户。

customer.

是的。

Yes.

对。

Yes.

给他们正确定价。

Pricing them correctly.

是的。

Yes.

处理服务。

Doing servicing.

那你能不能说说，你是怎么写评估的，让自己对这件事有信心的？

So so maybe talk through how did you how did you write the evals to get comfortable with that?

然后相比之下，我不知道，去银行网点或者甚至金融科技公司，那种体验怎么样？

And then versus, I don't know, going to a bank branch or or even a fintech, How is that experience

对。

Yes.

所以我们推出的第一个金融产品是车贷。

So that much the first financial product that we launched was a car loan.

在墨西哥和一些新兴市场，通常要等两个月甚至更久才能批下汽车贷款。

And usually in Mexico and in some emerging markets, it'll get two months or more to get a car loan approved.

我们通常在三分钟内就能批下来，这挺酷的，因为我们掌握了客户和汽车的所有数据。

We usually approve it in under three minutes, which is pretty cool, because we have all this data around the customer and the car.

如果客户还不起车贷了，他们可以直接把车还给我们，我们给他们换一辆更便宜的车，这样他们每个月付的钱就少了，也就摆脱了困境。这就是垂直整合业务的好处。

And if the customer can't pay for the car anymore, they'll just return it to us, we can give them a cheaper car, and then they pay a smaller amount each month, and they get out of the water, which is amazing about the vertical integration of the business.

但当我们开始推出其他金融产品时，我们意识到这对客户来说是个非常重要的决定。

But then when we started launching other financial products, we realized that this is a very important decision for the customer.

他们通常要花三到四个月才能下定决心买车并申请贷款，或者申请我们提供的个人贷款，比如大额个人贷款。

They usually take three to four months to make up their mind in buying a car and getting a loan, or getting a personal loan, like a large personal loan that we also do.

所以如果你在整个过程中了解你的客户，并让流程对他们来说变得简单，那么你的转化率和留存率就会飙升。

So if you get to know your customer throughout this process and make the process easy for them, then just your conversion and retention metrics start going through the roof.

这不仅仅是交易，而是要个性化地了解每个客户，在他们准备好时促成转化，通过深度个性化利率、风险和最高贷款额度，当然，这要符合整个投资组合的利益，但也要根据风险水平以及客户可能获得的其他优惠来优化。

It's not just the transaction, it's understanding each customer personally, and get them to convert when they're ready, with a very deep personalization of the interest rate, the risk, the max amount of the loan, in a way that makes sense for the portfolio as a whole, obviously, but that's optimized to the risk level and probably the other offers that the customer is is getting.

然后，也许给我们讲讲，你知道，评估一直是个热门话题。

And then maybe give us just to be you know, evals are always a very hot topic.

你一开始就提到了这个。

You kind of led with that.

那么，举个例子，评估中比较难设计的部分，或者你需要花很多时间去处理的部分，有哪些呢？

What is a like, what is an example of maybe a a hard to design area for evals or one where you had to spend extra amount of time?

尤其是考虑到有真金白银和个人信息泄露的风险。

What's just given the fact that there's real money PII at risk.

对。

Right.

所以当我们决定围绕AI重新设计公司时，你会问：AI能不能胜任这个工作，甚至是CEO这样的领导岗位？

So when we decided to redesign the company around AI, you asked the question, okay, is AI going to be able to do this job, like even the CEO job, or jobs where the leadership is?

老实说，答案很可能是肯定的，比如按照现在的进步速度，到2035年它就能做到。

And the answer honestly is probably yes, like in 2035 with a rate of improvement, it will be able to do.

所以我们就说，好，那我们现在就试试看。

So we said, okay, let's try it now.

我们来试试打造一个AI CEO吧。

Let's try and build an AI CEO.

所以我们在墨西哥划了一个城市，叫库埃纳瓦卡，然后我们把一个智能体放在我们的一个框架里当CEO。

So we carved out a city in Mexico, it's Cuernavaca, and we put an agent in one of our harnesses as the CEO.

它开始学习，开始做决策，并且根据这些决策进行评估。

And it starts learning, and it starts making decisions and evaluating on those decisions.

到现在才运行了六个星期。

And it's only been running for six weeks now.

第一个月的目标是把Cordova Can的利润翻倍。

The goal of the first month was to double the profits of Cordova Can.

虽然没有达到，但做到了1.5倍，也就是利润增加了50%。

It didn't reach it, but it was 1.5 X, like 50% more profits.

哇。

Wow.

太厉害了。

That's amazing.

就像管理整个城市一样，这太疯狂了，对吧？

Just like managing the city, which is crazy, right?

太不可思议了。

It's amazing.

而且它还是CEO。

And it's the CEO.

人们觉得那是AI最后才会取代的工作。

People were That was the last job AI was supposed to take.

其实也不是。

No, it isn't, really.

那这一切是怎么发生的？

And how did this happen?

就像一位极其聪明的人，聪明到菲尔兹奖级别的那种，会去分析每一个数字、每一个客户，做出精准预测，然后事无巨细地管理每天需要执行的所有事情来完成计划。

Like a very smart person, like Fields Medal level smart, like going into every single number, single customer, making the perfect forecast, and going to micromanage every single things that needs to be executed every day to reach a plan.

所以它真的会给库埃纳瓦卡的所有体力工人发消息，告诉他们当天的工作安排，并要求他们发语音消息回来汇报进展。

So he'll literally send messages to all the physical workers in Cuernavaca, with their plans for the day, and ask them to send voice notes back to know their progress.

所以客户满意度提升了，库存管理更好了，周转更快了，融资渗透率也更高了，基本上所有关键指标都开始改善。

So customer satisfaction grew, we got a better inventory, we rotated better, better financing penetration, like every KPI started to improve.

所以这非常酷，也非常令人兴奋。

So it's super cool and super exciting.

那么，我们现在还认为哪些工作需要继续培训和雇佣真人来做呢？

Now, what are the jobs where we think we're still training and hiring humans?

这些工作都是跟物理世界相关的。

Those are related to the physical world.

比如说到维修技师，Kavak在墨西哥大概有800名技师。

So when we talk about mechanics, Kavak has around, I think in Mexico, around 800 mechanics.

这里面涉及大量的灵活操作和感官判断，很难被替代。

There's lots of dexterity and senses that's super hard to substitute.

所以在那方面，我们也构建了同样的智能助手，用的是同一套可扩展的技术框架，而技师们就有了一个副手。

So there, we also build these agents with the exact same harness that's scaling, and the mechanics have the sidekick.

我之前跟你们说过，就像电影《料理鼠王》里那只老鼠，它其实是个大厨，跟人类一起合作，差不多就是那种感觉。

I was telling you guys earlier, it's like movie Ratatouille, like the mouse that's actually a chef collaborating with a human, it's kind of like that.

我们管这个副手叫El Mike，它会告诉他们怎么检查车子，给出建议，展示怎么操作。

So it's a sidekick, we call it El Mike, and it tells them how to inspect a car, and gives them tips, and shows them the way to do it.

而且检查的质量又大幅提升了，我们检查得更快，维修也更快，成本更低，但最重要的是，我们交付的汽车质量更高了。

And the quality of inspections, again, went through the roof, we're inspecting faster, we're repairing faster, it's cheaper, but most importantly, we're delivering higher quality cars.

保修索赔下降了，自推出以来大约降了26%，客户满意度也再次上升。

Warranties came down, around 26% since we launched, and customer satisfaction again went up.

所以关键就在于，如果你从零开始设计你的组织，拥有便宜又超级智能的能力，你会怎么设计，然后就去实现它。

So it's about this, like how would you design your organization from scratch with abundant super intelligence that's cheap, and just go build it.

Nelly，这正好引出了硅谷现在的一个热门话题，很多人担心未来的组织会是什么样子，以及人类在其中扮演什么角色。

Nelly, this is a good segue to a key topic right now in Silicon Valley where there's a lot of people worried about how the organizations of the future are gonna look like and the role that humans are gonna play.

是的。

Yes.

我觉得你刚才提到了一点，我很想听听，你们是怎么考虑这个问题的，以及……

And I think you touched a little bit on that, I'd love to hear, yeah, like how you guys are thinking about that, and the

组织。

organizations.

完全同意。

Totally.

所以三年前我们就非常重视这个问题，事实上每个人的工作都会发生变化。

So we took that question very seriously three years ago, and the truth is that everyone's job will change.

几年前我们做的那些事情，很可能以后由AI代理来完成得更好。

And what we were doing a couple of years ago will probably be performed better by an AI agent.

那这意味着什么呢？

So what does this mean?

我们需要对所有人进行培训。

We need to train everyone.

所以我们在 Kavak 内部启动了一个叫 JEDI Academy 的项目，面向公司里的每一个人，从首席执行官到普通员工，我们都特别欢迎。

So we launched a program inside Kavak that's called the JEDI Academy, where anyone from Quebec, like from the CEO We love this.

对。

Yeah.

这个项目会告诉大家，不管你是CEO、AI工程师还是修车技师，想要进这个学院学习都非常不容易。

And it tells them, like, from the CEO to AI engineers to mechanics, like, going to the academy, it's super hard.

我自己也亲自带过他们。

Like, I I've led them myself.

这个项目是你设计的。

You designed the program.

是我设计的。

I designed the program.

但得不断更新。

But constantly.

对，得不断更新，因为一切变化太快了。

Constantly, because you need to be upgrading the program, because everything's changing so fast.

而且你不能把这些人送到斯坦福去学这些，因为这是新东西，对吧？

And there's, you can't send these people outside to Stanford to learn this because it's new stuff, right?

所以我们培训所有人，六周后，他们就能把最先进的AI代理投入生产。

So we train everyone, and after six weeks, they launch state of the art agents, AI agents, to production.

无论是机械师、财务人员还是工程师，每个人都能做到。

And it's mechanics and finance guys and engineers, like everyone can do it.

这样做的结果是，也许这个人不会成为AI工程师，虽然有些人确实成了，但他们知道如何与这项新技术协作。

And what this generated is maybe this person won't become an AI engineer, some of them have, but they know how to collaborate with this new technology.

所以我们当时怎么看这个问题呢，大家听好了，没有回头路了。

So the way we looked about it was, guys, there's no way back.

这就是Kavak要走的方向。

This is the way Kavak is going.

这就是公司未来的样子。

This is the way the company will look like.

这些是工程团队、财务团队、产品团队要面对的变化。

These are the changes for the engineering team, the finance team, the product team.

这些就是将要改变的东西。

This is what's going to change.

你有两个选择：要么接受培训，掌握新技能，适应这个新现实、新世界；要么如果觉得不适合你，可以离开Kavak。但这就是我们要走的路。

You have the choice to train and get the skills to perform in this new reality in this new world, or maybe leave Kavak if this is not for you, but this is where we're going.

结果效果非常好。

And it worked great.

我们强化了公司文化，大家也都非常兴奋。

We strengthened the culture, and we were super excited.

大家真的很会构建这些智能体系统。

People really know how to build these agentic systems.

再看看现在的Kavak，任何流程都是智能体和人类协作的，有时候智能体是人类的上级，有时候人类在设计智能体。

And then if you look at Kavak now, any process, it's really a collaboration of agents and humans, and sometimes agents are the bosses of humans, and sometimes humans are designing the agents.

但我觉得我们真的做到了，并且改变了这一切。

But I think we managed to really build this and change this.

这源于我们每天都要学习，事情会不断变化。

And it's through this idea that we need to be learning every day and things will continue to change.

要保持竞争力，唯一的方法就是每个月或每几个月提升自己的技能。

And the only way to continue being relevant is to upgrade your skills every month or every couple of months.

但你确实有，或者说曾经有，你知道，成千上万的员工。

But you do have, or did have, you know, thousands of people.

现在大多数事情都由智能体来做。

Now agents do most things.

是的。

Yes.

那Kavak的组织结构是什么样的？

So, like, what is the org structure of Kavak?

中层管理这个概念还存在吗？

Like, does the middle management concept even exist anymore?

你们的组织架构是怎样的？

Like, what does your org look like?

对。

Right.

现在的结构是扁平化团队，团队非常资深，而且被充分授权。

So the way it looks like now is very flat teams, very senior teams, super empowered.

一个团队里，有工程、AI、运营，什么都有。

If you look at a team, you'll have engineering, AI, operations, everything.

他们要么在开发智能体，要么在为智能体工作，要么直接面对客户。

And they're either building the agents, working for the agents, or being in the physical world in front of the customer.

我们大部分组织都是这样的。

Most of our organization looks like that.

所以这真的是围绕未来组织形态的理念来构建的，并且以AI为核心，充分利用这项新技术。

So it's really built around the idea of how organizations will look like in the future, and around AI and really harnessing this new technology.

显然，这需要大量的再培训，因为在2022年或2023年，没有人构建代理，没有人帮助代理，也没有人听从代理的指令。

Obviously, this required lots of retraining because in 2023, or 2022, no one was building agents, No one was helping agents or taking orders from agents.

而且，应对实体世界或客户的方式，与代理告诉你该做什么或帮助你提升工作效率的方式是完全不同的。

And the way you cater to the physical world or the customers was in a different way than if an agent's telling you what to do or helping you make your job better.

所以，这和我们两年前的结构完全不同。

So it's a completely different structure than we had just two years ago.

解释一下，我们之前聊过，为代理工作是什么样的。

Explain, we talked about this before, what working for the agents look like.

我记得你描述的方式是，一个单一基因系统，然后当它失败时，就像，哦，被踢出来，交给人类处理，但之后信息就丢失了。

I think the way you described it was a unigenetic system, and then sometimes when it fails, it's like, Oh, that's kicked out to kind of a human cue, But then that's lost.

那你是如何把这些整合起来的？

So how have you brought that together?

所以，我们看待人在回路中的方式，以及目前在生产环境中运行的大规模代理系统，通常，如果代理遇到障碍或无法继续执行，它会将这个案例或客户转交给二级支持，然后就不再管了。

So, the way we see human in the loops and most of these agentic systems in production right now, large scale agentic systems, usually if an agent hits a wall or can't perform anymore, it'll send this case or this customer to a tier two support, and forget about it.

这样其实行不通，因为闭环没有闭合，你就没法生成数据来训练智能体做得更好。

That doesn't really work, because you don't close the loops, so you don't generate the data to train the agent to do this better.

现在有效的方法是，我们为每个客户都配备一个专属智能体，这样的智能体有数百万个。

What works right now is we have an agent that's obsessed with each of the customers, like millions of this.

它们能访问所有API和技能，而这些技能是由人类和智能体共同构建的。

They have access to every single API, every single skill, and we have agents building those, humans building those skills for them.

当智能体遇到障碍或取消操作时，它会调用API，说“我需要帮助”。

And then if an agent hits a wall or cancels something, it'll call this API, saying, I need help.

而在另一端，不是智能体或软件，而是真人在协助它们。

And on the other side, it's not an agent or software, it's a human helping them out.

但如果用组织架构图来看，实际上是人类团队在管理智能体。

But if you map this out in an org chart, it's really human teams that have an agent.

我得到的结果越来越好。

I'm getting better results.

这非常清晰。

It's super clear.

对，这说得通。

Like, it makes sense.

这其实是个很好的过渡。

That's actually a perfect segue.

而且我知道有很多大型机构的领导都会主动联系你。

And I know you get lots of leaders at larger institutions inbounding to you.

所以这也许能帮你省下不少电话。

So maybe this will save you many phone calls.

但我觉得，很多公司领导其实心里都明白这个道理。

But I think rationally, many leaders of companies intuitively understand this.

但在他们整个组织里部署AI，还是很难。

It is very hard still to deploy AI through their organization.

模型本身已经够好了。

Like, the models are good enough.

你也知道这一点。

You know that.

这其实是个组织问题。

It's a it's an org problem.

这是个心理问题。

It's a psychology problem.

那么你有什么建议，或者我有什么……

Like, what advice do you have, or what I have you

我觉得有两件事。

think it's two things.

第一，这必须自上而下推进，原因就在这里。

The first is it has to be top down because of this.

如果只是让大家自由采用，那不会有任何进展，因为很难让下面的人自发产生这种品味或战略，去决定该建什么、怎么建，最后弄出对公司真正有用的东西。

If you just get adoption, it won't go anywhere because it's hard to generate this taste or strategy for people to bottom up, decide what to build and whatnot, and come up with something that works for the company.

所以变革必须自上而下，领导者要带头采用，而且他们对要建什么必须有非常清晰的规划。

So the transformation has to be top down, and leaders need to adopt, and leaders have to have a very clear plan on what to build.

我见过太多公司，就是那种，哦，我们搞个黑客松，大家想各种用例，然后我们资助其中一些用例。

I've seen so many companies, it's just like, oh, we're doing a hackathon, people are coming up with use cases, we're sponsoring some of these use cases.

那行不通。

That doesn't work.

要非常清楚公司在三五年后会是什么样子，然后开始朝着那个方向建设，并且要非常垂直地引导你的团队朝着那个目标前进。

Be very clear on what the company will look like in three or five years, and then start building that, and be very vertical in guiding your troops towards that.

如果每个人都对战略和战术有自己的想法，然后上战场，想干什么就干什么，那军队是没法运作的。

An army doesn't really work if everyone comes up with ideas on the strategy and tactics, and goes to the battlefield, and does whatever they want.

你需要一个非常清晰的战略。

You need a very clear strategy.

而我们现在需要的就是这个，正处于转型阶段。

And that's what we need now, it's at transformation stage.

第二点，你需要衡量真正重要的东西。

The second one is you need to measure what really matters.

就是评估，但也要是正确的评估。

And it's evals, but it's also the right evals.

我看到很多公司现在投入巨资，然后说，好吧，我得到了采用。

So I see a lot of companies spending now huge amounts, and they say, okay, I got adoption.

我现在光是花在令牌上的钱就有好几亿美元了。

I'm just spending hundreds of millions of dollars in tokens now.

那又怎么样呢？

What about that?

令牌的质量很重要。

There's quality in the tokens.

我这里有一个框架，也挺有用的。

So I have a framework here that's also useful.

第三层令牌，也就是最有价值的，是那些能让你算出每个具体令牌投资回报率的智能体。

Tier three tokens, the most valuable, are these agents where you can get the ROI of each specific token.

我现在就能做到这一点。

And I can do that now.

这对我来说是个好消息。

That's great news for me.

因为我在成长，而且我知道每个令牌的投资回报率，因为它们都用在执行组织任务的智能体上。

Because I'm growing, and because I know the ROI of each token, because it goes to agents that are performing the job of the organization.

这些是最好的代币。

These are the best tokens.

第二层代币是那些可以间接衡量的东西。

Tier two tokens are things that you can measure indirectly.

我能在代码库中看到开发者，至少能间接评估这些代币的价值，然后把它们推向生产环境。

Do I see devs in the codebase, and I can evaluate the value of these tokens at least indirectly, and then push those to production.

第一层，大多数公司所处的阶段，就是人们只是用Plugcode、ChatGPT或者Cowork之类的工具。

Tier one, when most companies are, is people are just using Plugcode or ChatGPT or Cowork or whatever.

那些工具后来怎么样了？

What happened with those?

我完全不知道。

I have no idea.

所以这不仅仅是采用的问题，更重要的是要有非常清晰的愿景，然后衡量你花的每一个代币是否带来了那些好处，然后从那里不断迭代、迭代、再迭代。

So it's not just about adoption, it's really about having a very clear vision, and then measuring that each token you spend is bringing you those benefits, and just iterate, iterate, iterate from there.

我们之前稍微提到过这一点，但我觉得值得深入探讨一下，因为也许收听这个节目的最有雄心的公司会决定效仿，那就是你决定为每个客户构建一个代理。

And we touched on this a little bit, but I think it's worth a dive as maybe the the most ambitious companies listening to this will decide to follow suit, which is you decided to build an agent per customer

是的。

Yes.

而不是按任务来。

Versus per task.

是的。

Yes.

然后在这个过程中发现，每一个这样的代理都需要自己的一个微型虚拟机。

And then discovered along the way that each one of those agents needs its own sort of micro virtual machine.

所以能不能给我们讲讲，是的。

So maybe kind of walk us through Yes.

那些架构上的决策。

Those decisions at architecture.

我们现在看到这些结果了，但当时确实是个很冒险的赌注，因为大家通常都是从工作流开始，如果我能给所有人建议的话，别去构建代理式工作流，而是转向图、函数或者目标。

And and I think we're seeing these results now, but it was a really risky bet because people usually go from workflows, if I could advise everyone, don't build agentic workflows, to graphs, or functions, or objectives.

我们构建了那些，这些是多代理系统，能够为我们的复杂目标执行完整的功能。

And we built that, these are multi agent systems that can perform a whole function for our complex goals.

就像我之前跟你说的，卖一辆车需要处理融资、购买、推荐这些环节。

Like the ones I told you that to sell a car, need to do financing, purchasing, recommendations, etcetera.

去年十二月，我们有成千上万，甚至好几万个这样的智能体在规模化运行整个业务。

And we had thousands, like tens of thousands of these agents working at scale running the business back in December.

但后来 Opus 4.5 出来了，我意识到这个模式已经不再适用了。

But then, Opus 4.5 came out, and I realized this isn't the right paradigm anymore.

现在的智能水平已经不需要那种图结构和多智能体的复杂框架了，因为那反而会限制这种智能水平。

The intelligence now doesn't need the graph and the multi agent lattice work and harness, because it will constraint this level of intelligence.

所以我们决定把两年来一直在建的东西全部推翻，尽管它运行良好，让我们实现了盈利，带来了惊人的增长，然后从头开始，用一个我们认为更稳健、更可扩展的框架，并且能利用递归自我改进，或者每个月推出的更智能的新模型。

So we decided to destroy everything we had been building for two years, that was working, that brought us to profitability, that brought us amazing growth, and start over with a harness that we thought would be robust and scalable, and leverage recursive self improvement, or new models, more intelligent models coming out every month.

所以现在的样子是，一个虚拟机上跑着一个智能体，它能访问记忆和评估系统，还有命令行接口，可以调用我们公司所有的工具和 API，以及长期目标。

So the way this looks like it's a virtual machine with an agent, with access to memory and evals, and the CLI where they can access every tool and every API in my company, and the long term goal.

我每天都会启动成千上万个这样的智能体，它们带着长期目标，比如最大化客户终身价值。

And I instantiate hundreds of thousands of these each day with long term goals, like maximizing the lifetime value.

对。

Yeah.