Zip 的 AI 超级代理：端到端自动化采购流程

我们可以把大量事务性工作自动化，但要说有人去说服另一个人做某件事、主动发起并推动一个项目——我没看到任何 AI 能做到这一点。

We can automate a lot of transactional work, but someone convincing another person about doing something, making initiative out of it, and running a project — I don't see any AI doing that.

我听到不少犹豫。我的意思是，现在大家都有点怕智能体。你的两个客户是 Anthropic 和 OpenAI，他们自己肯定能做出这个东西。

I'm hearing a lot of reticence. I mean, everyone's kind of afraid of the agents right now. Two of your customers are Anthropic and OpenAI, who certainly could build this themselves.

企业很容易说“我们当然在用 AI”。但现实是，组织必须把 AI 更深入地部署到各个流程里，而不是——你知道，嘴上说在试点 AI 之类的——我记得看到的是 39% 的投资回报率，对比 6% 的投资回报率。

It's very easy for an enterprise to say, "Of course we're using AI." But the reality is an organization has to deploy AI more deeply across their processes, versus just — you know, they say they're piloting AI or whatever — see 39% ROI, if I remember, versus 6% ROI.

你会去想这件事吗，整个经济会怎么变化？

Do you think about that at all, how the overall economy is going to change?

那么，你的背景是什么？你是怎么来到 Zip 的？也讲讲 Zip 的一些背景。它已经存在一段时间了。

So, what is your background? And how did you get to Zip? And give a little bit of Zip's background. It's been around for a while.

好的，没问题。我先简单讲讲我的背景，然后过渡到 Zip。我 2013 年从大学辍学——或者说是 2012 年底吧——去创办我的第一家公司。完全不同的业务，叫 FlightCar。我们会给你免费的机场停车位，办法是把你的车租给别人。全程有保险。如果你还记得，这就是当时所谓的共享经济，那是 12、13、14 年前，Uber、Lyft 这些共享出行公司才刚刚起步——他们在融 A 轮、B 轮之类的。所以那时候人们——现在很难相信——还在一贯地开车去机场。我们大概在 2012 年底创办了那家公司。最后我经营了大约 4 年。那是一个非常考验运营能力的生意。我把在那个生意里赚钱比作挤柠檬——柠檬挤出来的最后一滴，才是你在这类生意里留下的钱。当然，如果你中途搞砸了，你不需要把剩下的柠檬汁都挤出来才知道结局会怎样，对吧？我们在机场有实体网点，覆盖了 17 个以上的机场，有人到处跑着洗车、去机场接人，那个生意里会发生各种糟糕的事，对吧？车被偷、被用来干坏事，各种各样的事情都会发生。最后我们在 2016 年把资产卖给了梅赛德斯-奔驰，那不是我们当初想要的结果，但我从中学到的是：我再也不想做这种运营密集型的生意了。我创办的下一家公司会是一家 B2B SaaS 公司。那还是在 AI 之前，早得很，对吧？2016 年。这就是我得到的教训。

Yeah, absolutely. So, my quick background, and then I'll segue that into Zip. I dropped out of college in 2013 — or late '12, I guess — to start my first company. Totally different business. It was called FlightCar. So, we would give you free airport parking by renting out your car to someone else. Fully insured. If you remember, this was the sharing economy, as it was called, you know, 12, 13, 14 years ago, and Uber and Lyft and all the ride-sharing companies had just recently started — they were raising their Series A's and B's and that kind of thing. So, people were — it's hard to believe, but still driving to the airport consistently back then. And so, that's how we started that company, yeah, some point late '12. Ended up running the company for about 4 years. It was a very operationally challenging business to run. I equate making money in that business to squeezing a lemon — the last drop that comes out of the lemon is the money you keep in a business like that. Of course, if you screw up halfway through, you don't need to squeeze the rest of the juice out of the lemon to know how things are going to end up, right? And we had physical locations at airports. We had over 17 different airports, and we had people running around washing cars and picking up folks from the airport, and all sorts of terrible things happen in that business, right? Cars get stolen, they get used for bad things, all sorts of other things happen. And we ended up doing an asset sale to Mercedes-Benz in 2016, which was not the outcome that we had set out to have, but what I learned from that is, well, I don't think I want to do an operationally intensive business like this ever again. The next company I started is going to be a B2B SaaS company. And that was all pre-AI, well before that, right? 2016. And so, that was kind of the learning that I had.

后来，因为我辍学了、也没在任何地方工作过，我想去一家真正的科技公司工作。于是我去了 Airbnb 工作几年，也是在那里认识了我的联合创始人 Lou。我在那里不得不走采购流程，可以说对采购的痛点、所有相关方以及它的复杂性都更熟悉了。之后，我去 YC 做访问合伙人——我承诺待两个批次。然后 Lou 和我都决定离开，全职投入这家公司。但我们又花了四五个月，在疫情当中，才想清楚要做什么，因为那时我们其实还没有这个点子。然后我们在 2020 年 8 月左右创办了今天这家公司。所以，它已经——

And then, because I dropped out and I hadn't worked anywhere, I wanted to go work for a real tech company. So, I ended up working for Airbnb for a few years, which is where I met my co-founder Lou, and I had to go through the procurement process and became more familiar, I would say, with the pain points of procurement and all the stakeholders and just the complexity of it. And then after that, I ended up working at YC as a visiting partner — I committed to being there for two batches. And then Lou and I had both decided that we would leave and be full-time on the company. But it still took us another four, five more months, just in the middle of COVID, figuring out what we were going to do, because we hadn't really had this idea yet. And then we started the company as it is today in August or so of 2020. So, it's been—

哇。六年——

Wow. Six—

差不多。

Almost.

哇。我记得读到过，你们在那段时间里累计处理了超过一万亿美元的支出。

Wow. And you've processed, I think I read, more than a trillion dollars in cumulative spend in that time.

对，我想大概是 1.2 万亿。没错，这就是通过 Zip 流转、并在所有客户中被批准的金额。

Yeah, I think it's like 1.2 trillion. But yeah, that's how much has gone through Zip and been approved across all customers.

嗯，到今天为止。

Um, to today.

我们创办 Zip 时的打法，一直是去编排采购所涵盖的工作，这样我们就能理解这些工作，最终帮你完成这些工作，对吧？这就把我们带到了今天，带到了超级智能体。如果你想想采购流程实际上由什么构成，你大概可以把 80% 左右的采购流程拆解成一系列相对离散、直接的任务，对吧？就是有大量事务性工作。我举几个例子。如果你是一家公司，正在和一家新供应商做生意，你大概需要处理 W-9 或 W-8 这类文件，确保它合规、没有超过 3 年有效期。如果你组织里有人在采购——他们想买某个新的 AI 产品、某个新的软件产品——最终 IT 部门得有人去审一份 SOC 2 报告，对吧？他们要么直接盖章通过，要么说“等一下，我觉得这东西得再看一眼”。不管哪种，这都是一个相对离散、可以用智能体自动化的任务。

Our play when we started Zip was always to orchestrate the work that procurement encapsulates, so that we can understand the work, so that we can eventually help you do the work, right? And that sort of takes us to today and to the super agents. And if you think about what the procurement process actually comprises of, you could probably break down 80-ish percent of the procurement process into a series of relatively discrete, straightforward tasks, right? There's just a lot of transactional work. And I'll give you some examples. If you're a company and you're doing business with a new supplier, you're probably going to need to process a W-9 or a W-8 type of document and make sure it's compliant, that it's not more than 3 years out of date. If someone in your organization is procuring — they want to buy some new AI product, some new software product — eventually someone in IT will have to review a SOC 2 report, right? They're going to rubber stamp it or say, "Wait a minute, I think we've got to take another look at this thing." Either way, that's a relatively discrete task that can be automated with an agent.

所以，我们大约一年、一年多一点之前在 Zip 做的第一代智能体，其实是做很多不同的小型离散智能体，可以这么说，把它们放进工作流里，去完成这些离散的事情。但我们意识到，那些智能体有点像一次性智能体。比如一个 SOC 2 审查智能体——我们有一组固定的提示词，它会产出一个结果，然后人——你可以让它自动运行，或者把人叫进来看一眼、批准或不批准。但很多这类事情需要后续跟进，需要更多来回沟通。所以，几个月前我们发布的超级智能体，我们优先做的是真正离散的角色。举个例子，你可以想象 Zip 里的法务超级智能体，它能自主处理整个端到端的红线批注流程。它通过 MCP 连接到你的邮箱。它可以代表你给供应商或他们的法律顾问发邮件、处理红线批注、等他们回复，等等。这样你真的可以把很多人从工作中拿掉，因为 AI 至少在这个流程里第一次承担了更多端到端的工作，而不是工作的子组件。所以，这是我们一直非常兴奋的事情。

And so, our original iteration of agents in Zip, about a year, year and change ago, was to actually have many different little discrete agents, if you will, that can be dropped into the workflow that do all these discrete things. But we realized that those agents were sort of one-shot agents. So, like, a SOC 2 review agent — we have a certain set of prompts, it would result in an output, and then the human — you could have it autopilot, or the human would be brought in to take a look and approve or not. But there's follow-up that a lot of these things require more back and forth. And so, with super agents that we announced a few months ago, we prioritized really discrete personas. So, for example, you could think of the legal super agent in Zip, which can autonomously handle an entire end-to-end redlining process. It's connected to your email with MCP. It can send an email on your behalf to the supplier or to their legal counsel, process a redline, wait for them to come back, so on and so forth. And you can really take a lot of the human out of the work, because AI, for the first time in this process at least, handles more of the end-to-end work versus subcomponents of the work. And so, that's something that we've been really excited about.

你确实提到了采购，Craig。你说得对。我一直这样描述：世界上每一家企业，按定义都必须采购东西。进来的钱应该是你的营收。而当你考虑出去的钱时，最大的两块永远是薪酬和采购。在某些公司，比如科技公司，大概率薪酬排第一，采购排第二。在更传统的行业，比如制造业等，几乎可以肯定采购排第一，薪酬排第二。然后隔了很长很长很长一段，才是差旅报销之类的东西。但对所有企业来说，采购都是一个非常非常大的支出领域。

You did mention procurement, Craig. You're right. I always describe it this way: every business in the world, by definition, has to procure things. Money coming in should be your revenue. And then when you think about money going out, the top two buckets are always payroll and procurement. In certain companies, like tech companies, most likely payroll comes first, procurement comes second. In more traditional industries—manufacturing, etc.—almost certainly procurement will come first, payroll will come second. And then after a long, long, long gap, you have things like T&E and other things. But it's a very, very large area of spend for all businesses.

世界正变得越来越复杂。每当你的团队部署一个新的云工作负载或 AI 智能体，就又多了一个身份获得对你关键系统的永久访问权限。你知道吗？去年十家组织中有九家遭遇了身份泄露，因为攻击者是用有效凭证登录的，而不是破门而入。非人类身份的数量是人类身份的 109 倍，在云环境中制造了巨大的无人管理的攻击面。传统工具是为管理人类员工而建的，这让现代机器和 AI 访问在很大程度上处于无人管理状态。这正是 Palo Alto Networks 的 Idira 的用武之地。人类、机器、AI——一个身份平台统管全部。96% 的人类用户在使用他们其实并不需要的过度权限。Idira 用动态访问取代永久权限，让你在不拖慢业务的前提下锁定每一个身份。Idira 按需授予访问权限，并在任务完成后立即撤销，从而消除高风险的永久权限。借助 Idira，你获得统一控制：用一个控制平面取代各自为政的 IAM（身份与访问管理）、PAM（现代特权访问管理）和治理工具，覆盖人类、机器和 AI 身份。用 Palo Alto Networks 的 Idira 保护你组织中的每一个身份。了解更多请访问 paloaltonetworks.com/idira。

The world is getting increasingly complicated. Every time your team deploys a new cloud workload or AI agent, another identity gets permanent access to your critical systems. Did you know nine out of 10 organizations suffered an identity breach last year because adversaries are logging in with valid credentials, not breaking in. Non-human identities outnumber human identities 109 to one, creating a massive unmanaged attack surface in cloud environments. Legacy tools were built to manage human employees, which leaves modern machine and AI access largely unmanaged. That's where Idira by Palo Alto Networks comes in. Human, machine, AI—one identity platform for all. 96% of human users operate with excessive permissions they don't actually need. Idira replaces permanent permissions with dynamic access so you can lock down every identity without slowing down your business. Idira provisions access on demand and revokes it immediately when the job is done, which removes high-risk permanent permissions. With Idira, you get unified control by replacing siloed IAM—identity and access management—PAM—modern privilege access management—and governance tools with a single control plane across human, machine, and AI identities. Secure every identity at your organization with Idira by Palo Alto Networks. Learn more at paloaltonetworks.com/idira.

我说这是个有意思的时期，指的是那些从沙箱里逃出来的失控 AI 集群之类的事。跟企业高管交流时，我听到很多迟疑。我的意思是，现在大家都有点怕智能体。我倒是觉得，某种程度上这对你们有利，因为如果一家公司并不真正理解安全与治理，自己造智能体到处跑去做这些事，要危险得多。而你们想必花了大量时间把这些边界都加固好。你能谈谈这方面吗——关于安全和治理？

Well, when I say it's an interesting time, I'm referring to the rogue AI swarms that have broken out of sandboxes and things. And talking to corporate C-suite executives, I'm hearing a lot of reticence. I mean, everyone's kind of afraid of agents right now. I would think in some ways that works to your advantage, because it's much more dangerous for a company to build its own agents to run around and do this if they don't really understand safety and governance. Whereas presumably you guys spend a lot of time bolting down the edges on this. Can you talk about that—about safety and governance?

是的，绝对是这样。首先，你说得完全对。一个组织内的所有用例并不承担相同的风险。我们发现，采购是风险最高、下行影响最大的领域之一。如果你用一个智能体写营销邮件发给别人，内容不对——说了些不准确的话——对方可能会抱怨，但没什么大不了的，没关系。那其实正是适合让 AI 稍微放开手脚的那类用例。但涉及采购，以及任何触及你 ERP 下游总账的东西——那正是支撑你损益表的东西——你就进入了一个对此毫无容忍度的领域。它必须是真的正确。因为你在拿这家公司冒险。出于显而易见的原因，你的损益表不能是“大体正确”，甚至不能差一分钱。对此绝对零容忍。

Yeah, no, absolutely. First of all, you're absolutely right. Not all use cases within an organization carry equal risk. And what we find is procurement is one of the areas with the highest risk and the most downside. If you use an agent and it writes an email for marketing and sends it out to someone and it wasn't correct—it said something incorrect—they might complain, but whatever, it's okay. That is actually the right type of use case for AI to be let loose a little bit. But when it comes to procurement and anything that touches the general ledger downstream in your ERP—that is what's feeding your P&L—you get to an area where there's no appetite for any of that. It's got to be actually correct. Because you're putting this corporation at risk. For obvious reasons, you cannot have a P&L statement that is mostly correct or even off by a cent. There's absolutely no tolerance for that.

所以我们花了大量时间构建的，就是这样一个受治理的框架，让智能体可以在其中生存和运作——我们超级智能体的能力可以活在这个框架里。我举几个例子。我们在 Zip 里有一个关键项，就是对智能体决定要做和实际所做的每一件事、它调用的上下文、它从哪里获取信息，都有完整的可审计性——它当然会引用来源。你能看到某件事是什么时候发生的。你能看到，哦，这个智能体当时被允许——比如我给你的那个 SOC 2 报告的例子——去判断这个供应商按我们的政策拥有足够的 SOC 2。是谁配置了那个能力、授予了那个权限？哪怕那是 3 个月前做的，所有这些信息都集中在一个地方，而且我们通过审计包让内部和外部审计师非常容易下载这些材料并加以验证。这是它的一个核心组成部分。

And so what we've spent so much time building is this governed framework within which agents can live and operate—our super agents' capabilities can live within that. I'll give you some examples. One of the key items we have in Zip is full auditability of everything an agent decides to do and does, the context it drew on, where it's getting it from—it's of course citing its sources. You can see when something happened. You can see, oh, this agent was actually allowed to, for example—that SOC 2 report example I gave you—to say that this supplier has a sufficient SOC 2 according to our policies. Who configured that capability and gave it that permission? Even if it was done 3 months ago, all of that information lives in one place, and we make it really easy with audit packages for internal and external auditors to download this stuff and validate it. That's a core component of it.

但我们也发现，能够说“嘿，在这些风险参数下，我永远想要——”很重要，比如我们跟很多银行合作——“我永远想要有人来审核并签字确认。”而我们也有一些科技客户更愿意说：“不，在这些参数下，我其实会让它自动运行，让它处理情况并做决定，但在这些参数下，我仍然要有人。”系统里就有这种级别的治理。这变得非常关键，因为像 Zip 这样的系统会连接你的 SAP 或 Oracle 环境并实际推送数据，我们会改动那个系统里的记录。所以你最好确保无误，因为那是整个组织的真相来源。大量治理必须投入其中。这就是我们花了这么多时间真正在构建的东西——这个框架，让你能够启用 AI，真正从人手上分担更多工作。

But what we also find is it's really important to be able to say, “Hey, under these risk parameters, I always want—” we work with a lot of banks, for example—“I always want a human to review and sign off.” And we have some tech customers that are more willing to say, “No, under these parameters, I'll actually let it run on autopilot and let it handle the situation and make a call, but under these parameters, I still want a human.” And you have that level of governance in the system. This becomes really critical because a system like Zip connects to your SAP or Oracle environment and actually pushes data, and we change records in that system. So you better be sure, because that is the source of truth for the entire organization. A lot of governance simply has to go into that. That's what we've spent so much time actually building—this framework so that you can enable AI to actually take more work off the human's plate.

是的。我读到你的两个客户是 Anthropic 和 OpenAI，他们当然有能力自己造这个。所以他们用你们，是相当大的背书。其实就这一点，你们从他们的治理研究中受益多少？因为肯定有关于在他们系统上使用这套东西的对话吧？

Yeah. And I read that two of your customers are Anthropic and OpenAI, who certainly could build this themselves. So the fact that they use you guys is a pretty big endorsement. Well, actually on that point, how much do you benefit from their governance research, because there must be a dialogue about using this on their systems?

是的，绝对有对话。我举一个我们遇到的例子——它其实非常难处理。

Yeah, no, there absolutely is a dialogue. And I'll give you an example that we saw—it's actually really complicated to handle.

假设你有一个用户，对吧？比如我们的客户在 Zip 里有上市前的机密数据。如果你是我们的客户，正准备上市，你会把这类信息分享给你合作的咨询公司，因为你可能需要共享这类信息。这些信息可能存在于发票里，可能存在于合同里，也可能存在于工作说明书中，因为这是 IPO 准备工作流之类的东西，对吧？那么这份文件几乎肯定会存在于 Zip 的某个地方，之前可能经过了机密请求和审批流程。现在，如果我作为用户问：“嘿，我有个问题”，而这个问题恰好涉及那份数据，你当然不希望 AI 用这些其他数据来回答我问的另一个问题。但我不应该看到那个来源，我不应该知道那些信息，因为我不该有权限访问它，对吧？所以在 Zip，我们构建了 350 种不同类型的权限，这些权限同时适用于人类和智能体。它的运作方式是：如果一个智能体代表我行事，那个智能体只能看到我能看到的东西。所以人类的权限提供了边界，智能体只能存在于这些权限的子集之内。这一点，第一，构建起来相当有挑战性；第二，测试、验证并确保它正确运行同样有挑战性，甚至更有挑战性，对吧？而且这类事情的下行风险太多了。所以我们从这两家公司以及它们最近关于治理的思考中获益良多。企业里并非所有领域都是平等的。

So, say you have a user, right? Our customers have pre-confidential data in Zip, for example. If you're a customer of ours and you're going to go public, you shared that with a consultancy that you work with because you may need to share that type of information. That information might sit on an invoice, it might sit in a contract, it might sit in a statement of work because it's an IPO readiness work stream or something, right? Now, that document is almost certainly going to be sitting somewhere in Zip, having gone through maybe a confidential request and approval process at some point. Now, if I as a user ask, 'Hey, I have some question,' and it happens to touch that data, you of course don't want AI to use this other data to give me the answer to some other question that I asked. But I should not be seeing that source, I should not know that information because I shouldn't have access to it, right? And so, in Zip, we've built 350 different types of permissions, and those permissions apply to both humans and to agents. And the way it works is that if an agent is acting on my behalf, that agent can only see the stuff that I can see. So the human's permission sort of provides the boundary, and the agent can only reside within the subset of those permissions. And that's one, quite challenging to build, but two, also equally challenging and more challenging to actually test and validate and make sure it's working correctly, right? And there's just so much downside with this stuff. And so we've absolutely benefited a tremendous amount from both of those companies and their most recent thinking around governance. Not all areas of the enterprise are created equal.

是的。那这一切是建立在、或者说基于一两个模型之上吗？你们底层运行着多少个模型？它们是 Anthropic 和 OpenAI 的模型，还是你们在架构中使用了其他专有模型或开源模型？

Yeah. And is all of this sitting on, or based on, one or two models? How many underlying models do you have running, and are they Anthropic and OpenAI models, or do you use other proprietary models or open-source models in the architecture?

是的，你可以想象，这是一个快速变化的领域，我会这么说。但我们针对不同用例持续测试不同的模型。比如，我们有客户会在 Zip 里处理发票。结果发现，解析一张发票与让 AI 在 MSA（主服务协议）上生成修订稿，是完全不同的用例；或者让 AI 仅仅判断某件事是否在正确的——实际上，是判断某件事是否有预算，通过尝试找出正确的预算代码并验证这类逻辑，或者一系列其他事情。所以我们实际上是在一个近乎持续的基础上审视这件事。我可以说每天都会实际测试不同的模型。我们在测试 Zip 拥有的所有上下文。我们也会在有 Zip 上下文和没有 Zip 上下文的情况下测试，看有效率是多少。我们针对许多不同类型的用例追踪这一点。所以对发票来说，可能是行级编码，也可能是表头级——我们追踪采购订单的表头级、行级，合同里则在条款级追踪各种其他东西。我们一直在测试这些东西，而且我们绝对会在底层和幕后做调整，根据用例和复杂度重新路由。比如，我们可能有一个客户有 1200 页的发票，这可能难以置信，但我们有些客户从某些类型的供应商那里购买大量单件商品，而整张发票甚至可能不是英文的，对吧？那个用例与一页纸的简单发票非常不同，所以即便是那种情况也会以不同方式路由、以不同方式测试。所以这就是我们做的事。但没错，我们与——我们使用两家领先实验室的模型。

Yeah, as you can imagine, this is an area that's rapidly changing, is what I would say. But we continuously test different models for different use cases. So for example, we have customers that will process invoices in Zip. Encoding an invoice turns out to be a wildly different use case than having AI produce a redline on an MSA, or having AI simply determine whether something is in the right—actually, whether there's budget for something by trying to figure out what the right budget code might be and validating that type of logic, or a series of other things. And so we actually look at this on an effectively continuous basis. So I'd say like a daily basis, where we actually test different models. We're testing all of the context that Zip has. And we also test with Zip context and without, to see what the efficacy rate is. And we track this for many different types of use cases. So for invoices, it might be the line-level coding. It might be the header level—we track header level, line level for purchase orders, you track all sorts of other stuff at the clause level for contracts. We're always testing these things, and we absolutely make tweaks under the hood and behind the scenes to re-route depending on the use case and the complexity. For example, we might have a customer with a 1,200-page invoice, which might be hard to believe, but we have some customers that are buying tremendous amounts of individual goods from certain types of suppliers, and that whole invoice might not even be in English, right? That use case is very different from a one-page simple invoice, and so even that gets routed differently and tested differently. And so that's what we do. But yes, we work with—we use both of the models of the leading labs.

那你们会在新模型变得更强时换用新模型吗？

And are you swapping in new models as they become more capable?

百分之百。我们发现——有意思的是——对某些用例，比如处理一张简单发票，甚至更复杂的发票，你其实不需要用最新的模型。你不需要 Fable，甚至不需要 Opus，来做这些事。嗯，不是所有这些事，而是其中一些事，比如发票。实际上更重要的是上下文。所以我们也帮助管理这一点，因为当然这一切都有成本因素。

100%. And what we find—what's interesting—is that for certain use cases, like processing a simple invoice, even a more complex invoice, you don't need to actually use the most up-to-date. You don't need Fable, you don't even need Opus, for any of this stuff. Well, not any of this stuff, but for some of this stuff, for like an invoice. And it's actually the context that matters way more. And so we help manage that too, because of course there's a cost component to all this stuff.

是的。关于上下文窗口——你说上下文，是指进入上下文窗口的内容，对吗？还是说你只是指让智能体理解它们工作所处的语境？

Yeah. On the context window—when you say context, you're talking about what goes into the context window. Is that right, or are you just talking about having the agents understand the context in which they're working?

不，在这个情况下我觉得是前者。而且是的，这是一个我们真正投入过的领域，因为这是一个高体量的领域，尤其是发票。

No, I think the former in this case. And yeah, I mean, it's an area that we really invested in because this is a high-volume kind of area, especially invoices.

是的。而且有很多讨论说，随着上下文窗口增长，所需算力会呈二次方扩张。也有很多研究在探讨如何应对或管理这一点。这对你们来说是个问题吗？我是说，如果你有一张 1500 页的发票，把所有这些都放进上下文窗口，很多模型会相当困惑。

Yeah. And there's a lot of talk about the sort of quadratic expansion of the compute required as the context window grows. And there's a lot of research into how to wrangle that or how to manage that. Is that an issue for you guys? I mean, if you have a 1,500-page invoice, a lot of models would get pretty confused if you put all that in the context window.

我不觉得这对我们是个问题。坦白说，我在内部也不是这方面的专家。我跟你坦诚这一点。但就我所见——我们正在做的——它没有成为问题。不过话说回来，也许我漏掉了什么。

I don't think it's been an issue for us. I'm also candidly not the expert in some of this, just internally. I'll be open with you about that. But what I've seen—what we're doing—it hasn't been an issue. But again, maybe I'm missing something.

你说 AI 原生，你们构建公司的方式有什么不同，是 Coupa、SAP、Ariba 之类的公司如果建立在遗留系统之上就必须做得不一样的地方？

When you say AI native, what's different about the way you built your company that a Coupa or an SAP or Ariba or somebody like that would have to do differently if they're building on a legacy system?

这是个好问题。我认为有几个组成部分。第一，是你拥有的人，对吧？归根结底，我们的公司就是我们拥有的人。我们确保产品、工程和设计——我们只在这些地方招聘这些岗位：湾区、纽约市和多伦多。所以我们努力招最好的工程师、设计师。我们拥有的大多数工程师——实际上几乎全部——都来自所谓的遗留型背景。他们更多来自当前这一代科技公司，很多有消费者背景，因为我们总是说：“嘿，我们是在为最终用户解决问题。我们是在为请求者解决问题。如果我们让他们的生活更轻松、为他们的体验服务，那就会帮助采购。”

That's a great question. I think there's a couple of components of it. I think one, it's the people that you have, right? Ultimately, our company is the set of people that we have. And we make sure that product, engineering, and design—we only hire for those roles here in the Bay Area, in New York City, and in Toronto. And so we try to hire the best engineers, designers. Most of the engineers that we have—nearly all, actually—come from a legacy type background, if you will. They come from more of a current generation of tech company, a lot of consumer backgrounds, because we always say, 'Hey, we're solving for the end user. We're solving for the requester. If we make their life easier and we solve for their experience, it's going to help procurement.'

这会帮助法务，帮助 IT，帮助所有跨职能的利益相关者。所以拥有这种 DNA，我认为可能是最难的事。也是最难改变的事。所以这是我们首要的任务。但现实是，我们没有 20、30、40、50 年的技术债务和极其广泛、庞杂的用例面。我认为这个领域的一些公司就是这样。就变成了——你在服务所有这些不同的行业，我们也是，但你必须非常自律和勤勉地约束用例，不为客户定制任何东西。在旧世界里，你就不能——它们无法升级，你会有所有这些发散的产品分支。就很难把它变成真正利用 AI 的东西。这几乎是不可能完成的任务，幸好我们没有这些。而且我认为我们有正确的文化。它始于人。

It's going to help legal. It's going to help IT. It's going to help all these cross-functional stakeholders. And so having that DNA, I think, is maybe the hardest thing, actually. And the hardest thing to change. So that's the first order of business for us. But the reality is we don't have 20, 30, 40, 50 years of technical debt and an extremely wide, sprawling surface area of use cases. I think that's what happens with some of the companies in the space. It just becomes—you're servicing all these different industries, and we do too, but you have to be really disciplined and diligent about constraining the use cases and not building any bespoke or custom fit for a customer. In the old world, then you can't—they can't be upgraded, and you've got all these fanned-out branches of the product. It just becomes really hard to take that and turn that into something that really utilizes AI. It's almost an impossible task, and thankfully we don't have any of that. And I think we've got the right culture. It starts with the people first.

你提到你们的系统处理 300 多项权限。这些是如何设置的？我的意思是，如果你在引导客户上线，他们是否必须浏览一个包含 300 多项权限的切换列表，还是有某种自动化方式？

You mentioned that your system handles 300 plus permissions. How are those set? I mean, if you're onboarding a customer, do they have to go through a toggle list of 300 plus permissions or is there some way that that's automated?

这是个好问题。权限的工作方式是——在 Zip 中，你可以创建不同类型的角色。你可以把它们看作角色，对吧？比如管理员、请求者、法务审核员等等。所以我们提供了一组预先配置的推荐默认角色，然后你可以点击进入一个角色，看到，好的，这个权限是开的，这个是关的，这个只是查看，这个是读取，这个是编辑，对吧？以此类推。这些也是基于最佳实践以及我们从数百上千的客户那里学到和看到的情况默认设置的。所以这些都是为客户配置好的。但我们发现，较小的组织可能会说——他们实际上会说，‘是的，这看起来不错。我们就用这些吧。’但有时我们与财富 50 强公司合作，他们想要——他们会请来审计师，他们会逐一检查每一个权限，因为这对他们来说非常重要，他们想这么做。当然，我们完全支持这两种方式，但我们希望让人们容易上手并有默认设置，如果他们选择审计每一个，他们完全可以这样做。这就是权限的工作方式。

It's a great question. So the way that the permissions work—in Zip, you can create different types of roles. You can think of them as personas, right? As admin, requestor, legal reviewer, so forth. So we come with a set of recommended default roles that are pre-configured, and then you can click into a role and see, okay, this permission's on, this one's off, this one's view only, this one's read, this one's edit, right? So on and so forth. And so those are also defaulted based on best practice and frankly just what we've learned and seen across hundreds and hundreds and hundreds of customers. And so that's all configured for the customer. But what we find is a smaller organization might say—and they'll actually say, 'Yeah, this seems fine. We'll go forward with these.' But sometimes we work with a Fortune 50 company and they want to—they'll loop in their auditor and they'll actually go through every single one because it's that important to them and that's what they want to do. And of course we're completely supportive of either approach, but we want to make it easy for people and have defaults, and should they choose to audit every single one, they absolutely can. That's how the permissions work there.

对。是的。在——超级智能体在没有人参与的情况下能端到端完成多少工作？再次，我想这是由客户设定的，但有没有某些事情是你们不允许智能体在没有人工干预的情况下做的？

Right. Yeah. In the—how much can the super agents do end-to-end without a human in the loop? Again, I imagine that's set by the customer, but are there certain things that you don't allow the agents to do without human intervention?

是的，这实际上是个好问题。所以超级智能体可以在 Zip 中运行端到端流程。我举个例子。如果有人提交了购买 XYZ 类型商品的请求，你可以让超级智能体在那里说，‘好的，我们知道你想要这个。我去找几家我们已经合作的供应商，然后也许一两家我们没有合作但会是好供应商的。’并运行端到端的 RFP 流程。它会实际找到供应商，无论它们是在你的供应商主数据中——那是你 ERP 中供应商的真相来源。它当然可以运行网络搜索，找出谁最适合帮助你，联系他们，创建 RFP，发送 RFP，等待结果，计算结果，做出判断或询问它认为基于成本和价值的正确决策是什么，然后返回供应商选择，对吧？这是一个例子。另一个端到端流程的例子是当事情进展更多，我们到了法务步骤，说，‘好的，我手头有客户完整的法务操作手册。我们有 MSA。我将实际对其进行注释，做完整的红线标记，并给供应商发邮件，让供应商的律师回复。当然，在一定轮次后或某些点仍未解决时，它会举手说，‘嘿，我想我需要人类律师打电话和这些人谈谈。’但否则，它可以处理那个端到端流程。所以这些是一些例子。这些是相当复杂、漫长的流程。我们构建了一个产品，首先，就像我使用 Claude 或其他东西时，它会和我确认说，‘嘿，你对这封邮件满意吗？你要我发送吗？’而我作为人类可以说，‘下次我想开启自动驾驶时就不要问我了。’你可以做出那个决定，它会端到端运行。系统中更多的制衡是在我们任何时候将数据推送到源系统时，特别是像 ERP 系统，当有这样的集成推送时，我们强烈建议有人在环路中审查，以确保某些类型的记录是正确的，因为那自然会有更多风险。但坦率地说，在企业中，我们看到几乎每个客户都希望这样。

Yeah, that's actually a great question. So a super agent can run an end-to-end process in Zip. So I'll give you an example. If someone puts in a request to buy XYZ type of good, you can have a super agent in there say, 'Great, we know you want this. I'm going to go find a few suppliers that we already do business with and then maybe one or two that we don't do business with that would be good suppliers.' And run an end-to-end RFP process. It'll actually find the suppliers, whether they're in your vendor master, which is the source of truth for your vendor suppliers in your ERP. It can of course run a web search, go figure out who would be best positioned to help you with this, reach out to them, create an RFP, send them an RFP, wait for the results to come in, compute the results, make a judgment call or ask what it thinks the right decision is based on cost and value, and come back with a supplier selection, right? That's one example. Another example of an end-to-end process would be when something gets more progressed and we're at the legal step to say, 'Okay, great. I've got the customer's full legal playbook, of course, in hand. We've got the MSA. I'm going to actually annotate it, do a full redline, and send an email back to a supplier, and let the supplier's counsel do their turn back and come back. And of course, after a certain number of turns or certain points remain unresolved, it will raise up a hand and say, 'Hey, I think I need a human lawyer to get on the phone and talk to these guys.' But otherwise, it can handle that end-to-end process. So those are some examples. And those are pretty complicated, lengthy processes that exist. We've built a product so that—first of all, just like when I use Claude or something else, it will check with me and say, 'Hey, are you fine with this email? You want me to send it?' And I as a human can say, 'Just don't ask me next time when I want to turn this thing on autopilot.' And you can make that decision and it will run that end-to-end. Where there are more checks and balances in the system is anytime we're pushing data into a source system, especially like an ERP system, when there's an integration push like that, we really strongly recommend there being a human in the loop review just to make sure that certain types of records are correct, because there can just be naturally more risk with that. But candidly, in the enterprise, we see pretty much every customer wanting that.

那么，如果其中一个智能体确实谈了一个糟糕的条款或批准了不该批准的支出，谁来承担责任？是客户吗？是 Zip 吗？还是模型提供商？可能是模型层面的幻觉。

Well, if one of the agents does negotiate a bad term or approve spend that shouldn't have, who takes the hit? Is it the customer? Is it Zip? Or is it the model provider? It could have been a hallucination at the model level.

这是个好问题。我对此的思考方式是，如果我现在使用 Claude 来发送——当然我开着它在运行一些东西——但如果我用它给你或某人发送一封邮件，而它包含大量错误信息，语气不对或其他问题，最终我给了它权限去做这件事，那么责任在我。我不能怪 Claude，因为它只是做了它能做的，而我给了它权限并要求它去做这件事。

It's a great question. The way I think about this is if I use Claude right now to send—and I've got it open of course running something—but if I use it to send an email to you or someone and it's got a whole lot of incorrect information, doesn't have the right tone or whatever, ultimately and I gave it the permission to just go do this, that responsibility sits with me. I can't blame Claude because it just did what it could do and I gave it the permission and asked it to go do this thing.

所以，我也是这么看待这个问题的，对吧？最终，责任确实要落在某个人身上，而在今天这个显然还在演变的框架下，责任很可能落在那个决定让智能体去做 X、Y 或 Z，并授权它去做 X、Y 或 Z 的人身上。也许这会随时间改变，但至少我是这么看的。

And so, that's how I think about this, too, right? Ultimately, the responsibility does sit with someone, and it likely today—in the framework that is clearly evolving—sits with the human that made the decision to have the agent do X, Y, or Z and gave it the permission to do X, Y, or Z. And maybe that changes over time, but that's at least how I see it.

我觉得很有意思的是，你刚才说这些智能体会发出 RFP——智能体自己创建 RFP、生成文档，然后发给供应商？这些都是自主行动吗？

I mean, it's fascinating that these agents you were talking about sending out an RFP—the agent creates the RFP, the document, and sends it to the vendors? Are those all autonomous actions?

对，没错。Zip 里的超级智能体带有默认配置、默认提示词。比如，对于法律操作手册，它会有一个默认设置。对于采购——发起 RFX 类型的事件——它也会有默认设置。现在，我们与客户合作。我们有一个前置部署的工程团队与客户一起工作。当然，他们自己也能做，但许多客户更愿意和我们合作，让我们为他们处理和配置一些东西。我们可以让智能体利用公司内部的采购或 RFX 操作手册，针对不同类别的支出。随着公司规模变大，他们会有专门的品类经理负责某些领域，比如 IT 支出或销货成本中的某些支出类型。我们与一家大型快消品公司合作，当然，他们有专门负责糖采购的品类经理，对吧？还有可可之类的。所以这些操作手册可以输入到 Zip 中，供采购超级智能体、RFX 超级智能体使用。然后，基于这些，它绝对会创建 RFP。它会创建 RFP，并把它发出去。它可以处理所有这些。在实践中，我们发现——你的 100 万或 1000 万，不管你的阈值是多少，像大额的，比如你的 5000 万美元 AWS 合同——你会让人类来做，因为太贵了。可能是你最大的合同之类的。这些东西，我们发现客户就是希望由人类来做。但中间有一大块，还有长尾。长尾甚至都没被触及。中间部分也真的没被触及，因为工作量太大。如果是 25 万美元的采购，甚至 100 万美元的采购，对许多组织来说，做这件事太费劲了。但如果你有超级智能体，为什么不呢？我认为这就是我们的客户看到的，那里有真正的杠杆。我的意思是，这些钱没有被持续地拿出来招标。公司在多付钱，可能没有为他们的钱获得最高价值的产品或价值。是的，我可以帮忙。所以那里有真正的机会。

Yeah, correct. So, the super agents in Zip come with default configurations, default prompts. For example, for a legal playbook, it'll have a default in there. For sourcing—putting out an RFX type of event—it'll have a default. Now, we work with customers. We've got a forward deployed engineering team that works with customers. Again, they can do this themselves, but many customers prefer that they just work with us and we handle and configure some of the stuff for them. We can have the agent utilize a company's internal sourcing or RFX playbook for different categories of spend. As companies get much larger, they end up having specific category managers that own certain domains like IT spend or certain types of spend in cost of goods sold. We work with a large CPG company and, of course, they've got a category manager for just sugar purchasing, right? Versus cocoa versus whatever. And so those playbooks can be fed into Zip for the sourcing super agent, the RFX super agent, to utilize. And yeah, and then based on that, absolutely it will create an RFP. It will create an RFP and it will go send that thing out. And it can handle all of this. In practice, what we find is—your $1 million or $10 million, whatever your thresholds are, like the big ones, like your $50 million AWS contract—you're going to have a human do that just because it's so expensive. It's maybe your largest contract or whatever. That stuff, we find customers just want to have humans do it. But there's a whole middle and then there's a long tail. And the long tail is not even being touched. The middle is also not really being touched because it's just too much work. If it's a $250,000 purchase, even a million-dollar purchase, from many organizations, it's just too much effort to do this. But if you have a super agent, well, why wouldn't you do it, right? And I think that's what our customers are seeing, and there's real leverage there. I mean, these are dollars that are not being put out to bid consistently. Companies are overpaying for stuff and not maybe getting the highest value product or value for their dollars. And yeah, I can help with that. So there's real opportunity there.

让我着迷的一件事是——假设这是一个智能体起草 RFP 并发送给公司的案例。此时或很快，另一侧就有软件来读取 RFP。我之前和某人聊过 FDA 临床试验提交——设计临床试验——你创建一份 6 英寸厚的文档，包含所有这些程序和方案，然后发过去，但生成那份文档的是 AI。他们把它发给 FDA，而 FDA 直接把它输入 AI。我的意思是，到某个时候，这些智能体会直接传递数据。为什么要费事创建人类可读的文档？比如，你认为在 RFP 上会发生这种情况吗？

One of the things that fascinates me is—let's say this is a case where the agent is drawing up the RFP and it sends it to a company. At this point or very soon, there's software on the other side that reads the RFP. I was talking to somebody about FDA clinical trial submissions—designing a clinical trial—and you create a document that's 6 inches thick on all of these procedures and protocols, and they send it over, but it's an AI that is generating that document. They send it to the FDA and they just feed it into an AI. I mean, at some point these agents will just be passing data. Why go through the trouble of creating a human-readable document? Do you think that will ever happen on RFPs, for example?

我认为那绝对是我们正在走向的世界。我的意思是，我们完全使用很多 AI 工具为我们自己——是的，我们每年自己回应数百个 RFP，所以当然我们在那个过程中使用 AI。

I think that's absolutely the world we're headed in. I mean, we entirely use so many AI tools for our own—yeah, we respond to hundreds of RFPs ourselves a year, and so of course we use AI as part of that process.

在你们团队发来的材料中，有一个 UCI Health 的合同谈判案例。你能谈谈那个案例吗？

In the materials that your team sent over, there's a UCI Health contract negotiation. Can you talk about that case?

是的，是的，所以 UCI——更准确地说是 UCI Health——是我们的客户之一，我知道他们运行了——他们最终利用 Zip 的谈判智能体用例，为他们这个 IT 特定计划带来了 2000 多万美元的成本规避。现实是，像 UCI Health 这样的公司会有如何运行不同类型谈判的操作手册。而且可以是不同类型的操作手册，对吧？针对风险、法律、实际的美元和美分谈判以及价值等等。最终他们意识到，有了智能体，他们可以——与我们刚才的话题非常相关——覆盖更多。覆盖更多支出。如果你触及中间部分，甚至不是长尾，而是中间，它很快就会累积起来，因为即使你能在这个上削减 5%、10%，在那个上削减 20%，在另一个上削减 2%——我的意思是，所有这些都会累积。这是非常真实、非常非常真实的节省。归根结底，我觉得你也会感觉更好——你会想，‘好吧，为什么，我觉得我确实尝试过了，我们已经挤出了长尾中间剩下的大部分汁水。’所以这是我们和客户关注的一大领域。

Yeah, yeah, so UCI—UCI Health rather—is one of our customers, and I know that they ran—so they ended up utilizing a negotiation agent use case with Zip, and that drove a 20-plus million-dollar cost avoidance for this IT-specific initiative that they had. The reality is, a company like UCI Health will have playbooks for how to run different types of negotiation. And it can be different types of playbooks, right? For risk, for legal, for actual negotiation of the dollars and cents and the value and so on and so forth. And ultimately what they realized is with the agent they can—very related to the topic we were just on—cover more. Cover more of their spend. And if you get to the middle, not even the tail, but the middle, it adds up pretty quickly because even if you can shave 5%, 10% off on this one, or 20% on this one, or 2% on this other one—I mean, that all adds up. And it's very real, very, very real savings. And at the end of the day, I think you also feel better—you're like, 'Okay, why, I feel like I have actually tried and we've squeezed most of the juice that was left here in the middle of the tail.' And so that's a big area of focus for us and customers.

节省的钱——我看到了 2000 万这个数字。那是累计的。不是一份合同里的。

The money saved—I saw the number 20 million. That's cumulative. That's not in one contract.

是累计的。是的，在一段时间内——我得查一下是 6 个月还是一年之类的。但是的，是在一段时间内计算的。

Cumulative. Yeah, over a period of—I'd have to check whether it was like 6 months or a year or something like that. But yeah, it was computed over a period of time.

你们有一份报告调查了超过一千位支出负责人。其中哪个发现最违背你关于 AI 如何在这些组织内落地的假设？

You have a report that surveyed over a thousand spend leaders. What finding in it most contradicted your assumptions about how AI is landing inside these organizations?

是的，我觉得让我们惊讶的是——回想起来，也许本不该惊讶——但真正最重要的是部署的深度。企业很容易说，‘哦，我们当然在用 AI，我们在试点 AI。’但现实是，我们在数据中发现，那些在流程中更深入部署 AI 的组织——相对于只是说他们在试点 AI 之类的——获得了 39% 的投资回报率，如果我没记错的话，而另一组是 6%。

Yeah, I think what surprised us was—and thinking back, maybe it shouldn't have been a surprise—but it's the depth of the deployment that actually matters the most. It's very easy for an enterprise to say, 'Oh, we're of course we're using AI, we're piloting AI.' But the reality is, what we found in our data is that organizations that deploy AI more deeply across their processes—versus just saying they're piloting AI or whatever—see 39% ROI, if I remember, versus 6% ROI.

所以，比如，你知道，6 倍左右，6.5 倍。这其实也不应该让人意外，对吧？因为归根结底，如果你只是让 AI 运行，辅助人类完成一些基本任务，你只是在试点，你并没有真正获得杠杆效应，因为人类可能仍然在做他们原本就在做的 95%、99% 的工作，而你实际上增加了接触面。你必须给人类更多杠杆，而只有深入解决某些用例，你才能做到这一点。

So, like, you know, 6x or whatever, 6 and 1/2 x. Which again should be no surprise really, right? Because ultimately, if you just have AI running and you're assisting a human with some basic tasks and you're piloting something, you're not actually getting leverage because the human's probably still doing 95%, 99% of the work that they were already doing and you actually have increased the surface area. You have to give the human more leverage and you can only do that if you go deep in solving certain use cases.

我们还意识到，客户可以使用智能体，他们可以使用超级智能体和 Zip。他们可以开启它们，里面有默认设置，他们可以选择更改配置并编写自己的技能并修改技能，但你知道，他们可以直接使用超级智能体。但我们发现，很多企业，呃，你知道，我们一直在，呃，我们创建了这种我们称之为 AI 支出自动化的产品。所以，我们作为企业来找你，说：“嘿，看，我们查看了你的数据，在过去 12 个月里，你们在 Zip 中完成了 85,000 种不同类型的审批。我们查看了你们都在批准什么，对吧？我们查看了所有不同的审批和步骤，我们相信我们可以与你合作，创建这些，比如 25 个，我随便说的，用配置好的超级智能体解决这 25 个用例，比如配置成你拥有的定制用例，这应该会给你带来 X、Y 或 Z 的投资回报率，对吧？无论是缩短周期时间、节省更多成本，因为你把更多请求拿去招标，可能还有一些人员节省，通过 BPO，你知道，大型企业做的那种工作，呃，我们会为你配备一名 AI 前线部署工程师，实际帮你配置这些东西。再说一次，他们没有做任何你无法用产品做的事情。他们不碰代码，但我们发现，你知道，我们知道如何测试智能体，我们可以重写提示。这都是流程，我们可以为你做。这真的引起了共鸣，我们有很多客户实际上正在与我们进行这类转型，因为企业在采用 AI 方面存在最后一英里的问题，所以我认为这是这份报告的主要收获，或者主要收获之一。

And what we've realized too is, you know, customers have access to agents, they have access to super agents and Zip. They can turn them on, there's defaults in there, they can change the configuration if they choose and write their own skills and modify the skills, but you know, they can just use super agents as they are. But what we find is a lot of enterprises, uh you know, we've been uh we've created sort of this we call uh AI spend automation offering. So, we come to you as an enterprise and say, "Hey, look, we looked at your data and in the last 12 months, uh you all have done, you know, 85,000 different types of approvals in Zip. And we looked at what were you all approving, right? We looked at all the different approvals and the steps and we believe that we can work with you to create these, you know, 25, I'm making it up, solve for these 25 use cases with configured super agents, like configured to the bespoke use cases that you have and that should result in you know, an ROI for you of X, Y, or Z, right? Whether it be cycle time reduction, more cost savings just because you're putting more requests out to bid, there may be some head count savings with BPO, you know, type work uh that large enterprises do uh and and we will staff uh a forward an AI FDE AI forward deployed engineer with you to actually help configure this stuff for you. Again, they're not doing anything you can't do with the product. They're not touch code, but we find that, you know, we know how to like test agents and we can rewrite the prompt. It's all process and we can do that for you. And that has, you know, really resonated and we've got a bunch of customers actually doing these types of transformations with us because there's a last mile problem with with with AI in the enterprise and and so that's that I think was the main takeaway from uh or one of the main takeaways from this report.

你如何看待这改变商业或经济？嗯，我想在那之前，你们的渗透率如何？有竞争对手吗？嗯，你知道，你们在争夺市场份额吗？还是说这是一个开放领域，你们只是在增长过程中尽可能多地占领市场？还要多久每个企业都会使用这样的东西？

How do you see this changing commerce or the economy? Well, I guess before that, what kind of penetration do you have? Are there competitors? Um, you know, are you fighting for market share? Are you uh is it sort of an open field and you're just picking up as much of the market as you can as you grow? How long before every enterprise is using something like this?

所以，Zip 创建了 AI 采购编排这一类别，对吧？这有点像是用 AI 捕获请求，当然还要处理该请求在所有跨职能团队和系统（如预算、风险、法律、IT）中的路由，而且如今越来越多地使用 AI 自动化尽可能多的工作。但是，我们也做，我们称之为从接收到采购。嗯，这里有很多采购术语，但我们也做从接收到采购到付款。采购到付款是传统的类别，或者，你知道，提交采购申请，它被批准、批准、批准，变成采购订单，发票进来，匹配发生，发票得到，你知道，或者金额支付给供应商。这是已经存在了，你知道，50 年甚至更久的现有类别。所以，Zip 也做这个。除了采购和编排，我们有自己的产品做采购到付款。嗯，我们是世界上唯一一家实际上两者都做的公司。呃，我们很幸运今年首次作为远见者被纳入 Gartner 魔力象限的现有采购到付款类别。所以，我想说的是，实际上，你知道，我们发现，在企业中，公司真的越来越想转向平台。呃，他们想要一个非常 AI 优先、AI 深度嵌入其中的平台。呃，当你这样看问题时，世界上没有其他方这样做。呃，顺便说一下，我也想过为什么，因为我一直在想，为什么那样？我认为现实是，其中一些东西太复杂了，嗯，并且要求没有犯错余地，对吧？比如，我们必须处理，嗯，你知道，巴西国家的税务要求，因为我们有很多客户在巴西做生意，税务必须正确处理。显然，不能出错，因为如果多了，你就多付了。呃，如果少了，你可能有监管问题。呃，所以，它必须正确，对吧？而这只是一个例子。呃，所以，所有的权限和所有深度，呃，真的非常非常难做。所以，我们一直，你知道，继续专注于为产品构建更多深度，为 Zip 的 AI 组件构建更多深度，呃，然后抓住端到端的机会。

So, Zip created the uh AI procurement orchestration category, right? Which is sort of the capturing the request with the AI, handling the of course the routing of that request across all the cross-functional teams and systems that they sit in like budget, risk, legal, IT, and, you know, today, much more increasingly so, automating as much of that work as possible with AI. But, we also do we call that intake to procure. Um there's a lot of procurement lingo here involved, but we also do intake to procure to pay. And procure to pay is the traditional category of purchase or, you know, putting in a purchase requisition, it gets approved, approved, approved, it turns into a purchase order, invoice comes in, the match happens, the invoice gets, you know, or the amount gets paid out to the supplier. That's sort of the incumbent category that's been around for, you know, 50 years or or maybe even longer. And so, Zip has does that as well. We have our own product that does procure to pay in addition to procure and orchestration. Um and we're the only company in the world, actually, that does both. Uh, and we were fortunate to have been included in the Gartner Magic Quadrant for the incumbent procure-to-pay category for the first time as a visionary this year. And so, what I would say is to actually, you know, we find that in the enterprise, companies really want to move more and more to a platform. Uh, and they want a platform that is very AI-first and has AI deeply embedded sort of throughout it. Uh, and when you look at the the problem sort of that way, there is no other party in the world that does this. Uh, and I I thought about why, by the way, too, cuz I've been like, well, why is that? And I think the reality is some of this stuff is so complicated um, and requires there's no margin for error, right? Like, we have to handle, um, you know, tax requirements in the country of Brazil, because we have many customers that do business in Brazil, and the tax has to get handled correctly. Obviously, it can't be wrong, because if it's over, you just overpaid. Uh, if it's under, you might have a regulatory problem. Uh, and so, it's got to be correct, right? And And that's just like one example. Uh, and so, it's all of the permission and all of that depth uh, is really, really hard to do. And so, we've been, you know, continuing to be focused on just building more depth into the product and more depth into the AI components uh, of of Zip and and and then capturing the end-to-end opportunity.

在整个市场，我的意思是，这是一个无尽的市场吗？呃，或者你看到它在哪里增长？

In the market overall, I mean, is it sort of an endless market out there? Uh, or or where where do you see this growing?

呃，我，你知道，我认为，呃，有趣的是，你知道，我们解决的问题是如此横向。比如，我们有，你知道，一些领先制药公司的科学家，呃，他们使用 Zip 购买，呃，他们正在测试和创造的药物所需的化学品，你有，呃，实验室的研究人员，你知道，使用 Zip 采购他们日常工作中需要的任何东西，呃，你有，嗯，试着给个好例子，比如，你知道，一家制作巧克力棒的公司通过 Zip 采购进入巧克力棒的所有东西，对吧？所以，我们拥有的用例范围如此广泛。所以，你知道，在某种意义上，它确实是横向的，它确实实际上是无尽的。感觉确实如此，因为世界上每家公司，至少超过一定规模，也许最低端是几百名员工，都有这个问题。它太普遍了。对，每家公司都必须采购东西，随着经济，你知道，进步，经济中有更多专业化，公司最终会采购更多东西，因为你可以只专注于你的秘方和比较优势。为什么你要，你知道，为什么你不采购所有其他东西，因为它不会让你的业务特别。你可以专注于让它特别的东西。

Uh, I I You know, I think uh, what's been interesting is, you know, it's it's so horizontal, the problem that we solve. Like, we've got uh, you know, scientists at some of the leading pharma companies uh, that utilize Zip to buy uh, chemicals for the drugs that they're, you know, testing and creating to you have uh, researchers that you know, at the labs uh, using Zip to procure whatever they need in their day-to-day jobs to uh, you have um Try to give a good example like a you know a company that makes chocolate bars procuring everything that goes into the chocolate bar through Zip, right? So it's just such a wide breadth of of of sort of use cases that that we have. So you know in a sense it truly is horizontal and it truly actually is kind of endless. It certainly feels that way because every company in the world at least above a certain size maybe a few hundred employees at the absolute low end has this problem. It's just so universal. Right every company has to procure things and as economies you know progress and there's more specialization in the economy companies end up procuring more things because you can just focus on the thing that is your special sauce and your comparative advantage. And why should you you know why wouldn't you procure all the other stuff because it doesn't doesn't make your business special. You can focus on the thing that does.

对。

Right.

你知道，我们正处在一个特定的时刻，对吧？我得不断提醒自己这一点。所以像 Zip 这样的东西，10 年或 20 年后——不管是不是 Zip，还是那个流程、那个软件层——都会在市场上增长，并且如你所说，在深度上也会增长。Harry 谈到最终全球经济会建立在这些智能体构成的基底之上，由它们来完成所有这些事。你会思考这一点吗？你觉得这会发生吗？你认为这是一个错误的预测吗？以及，让这些看不见的 AI 智能体来处理所有这些日常事务，你觉得影响是什么？

And you know, we're at a moment in time, right? I have to keep reminding myself this. So something like Zip in 10 or 20 years — whether or not it's Zip or that process, that software layer — will have grown both in the market and, as you say, in its depth. Harry talks about eventually the global economy resting on this fabric of agents that do all this stuff. Do you think about that? Do you see that coming? Do you think that's a misguided prediction, and what do you think the implications are, again, for having these unseen AI agents taking care of all of this routine work?

我确实认同这一点。在很大程度上，我们会——我绝对相信，我们的客户会越来越多地把这套工作流自动化，构成采购流程的那些审核也会由 AI 来完成。这些工作中有很大一部分相当事务性。它很重要，但它是事务性的。你能看到越来越多这样的工作被自动化。但这意味着，围绕它的治理、治理框架——权限管理、可审计性、控制措施——只会变得越来越重要，对吧？有一点我没提到：我们大多数客户都把 Zip 纳入了 SOX 范围。SOX 是一个极其关键的框架。它显然源自 2000 年代初的安然丑闻。在《萨班斯-奥克斯利法案》上，你没有任何犯错余地，对吧？所以，如果你要给 AI 更多空间，这个框架就变得越来越关键。要给更多空间，你就必须有一套围绕它的框架和控制措施。所以我认为框架和控制会变得更加重要。而且我其实很期待的是，我们将能够专注于更令人兴奋的工作。是的，很多人会说战略性工作，我完全同意。但我也愿意相信，正是这类工作实际上更有回报。它更有趣。它比大量事务性工作更有成就感。我看到公司、公司的员工、组织——

I do subscribe to it. For the most part, we will — and I absolutely believe that we will have our customers just automating more and more of this workflow, and the reviews that comprise the procurement process will be done with AI. So much of that work is fairly transactional. It's important, but it's transactional. And you can see more and more of that getting automated. But what that implies is that the governance around this, and the framework for that governance — the permissioning, the auditability, the controls — only becomes more and more important, right? One thing I didn't mention: the majority of our customers have Zip in SOX scope. SOX is an extremely critical framework. Obviously it came out of the Enron scandal in the early 2000s. It's like, you have no margin for error on Sarbanes-Oxley. Right? And so it really makes the framework more and more crucial if you're going to give more rope to AI. To give more rope, you have to have the framework and the controls around it. And so I think the framework and the controls become more important. And I actually think what I'm excited for is that we'll be able to focus on more exciting work. Yes, a lot of folks say strategic work, and I totally agree with that. But I also would like to believe that it's that type of work that's actually more rewarding. It's more fun. It's more fulfilling than just a lot of that transactional work. I see companies, and employees of companies, and organizations—

这也会对经济产生影响，因为它让定价变得更高效。但很多人靠这种低效为生，你知道吧？很多企业能拿到那部分超支，是因为系统低效，因为他们和买方是朋友，或者买方和供应商是朋友，于是他们就不太较真。这其实也很有意思——随着时间推移，经济会变得越来越高效，但与此同时，很多所谓的“寻租者”最终会失去他们的优势或竞争力。你会思考这一点吗，整个经济会如何变化？

It's also going to have an impact on the economy in that it makes pricing much more efficient. But a lot of people live off that inefficiency, you know? A lot of businesses capture that overspend because the system is inefficient, because they're friends with the buyer, or the buyer is friends with the vendor, and they sort of don't press too hard. I mean, that's interesting, too — how over time the economy will become increasingly efficient, but at the same time a lot of the rent takers, as they're called, will eventually lose their advantage or their edge. I mean, do you think about that at all, how the overall economy is going to change?

我确实认同这一点。我完全同意。而且我认为这对世界是件好事，对吧？销售流程中的一些低效——不管是因为“我认识个人”，还是因为某些产品，甚至像我们这样的产品，需要被销售，需要企业销售，最终整个流程都相当低效、讲政治，诸如此类。但你可以想象，如果你从中挤出更多效率，为世界各地的组织带来更多价值，他们就能拿到那些资金、那些节省下来的钱，而且他们会的，经济会奖励那些把这些资金进一步再投资于研发、让产品变得更好的人。最终这很可能导致为其他工作和岗位招聘更多人。而且无论经济中发生过什么样的宏观变化，似乎事情只是被重新分配，实际上会创造出更多机会。所以，在 AI 的影响方面，我肯定是个乐观主义者。不过我也想说，我在某处读到过——我觉得最后一个被取代的工作会是企业销售。不幸的是，我同意这一点。因为我们可以自动化很多事务性工作，但让一个人说服另一个人去做某件事，并把它变成一个倡议、运行一个项目——我看不到任何 AI 能做到这一点。

I do subscribe to that. I totally agree. And I think it's a good thing for the world, right? I think some of that inefficiency in the sales process — whether it be because, you know, I know a guy, or because certain products, even like ours, need to be sold, and you need enterprise sellers, and ultimately that whole process is fairly inefficient and political and this and that. But you have to imagine that if you squeeze more out of that, and there's more value for organizations around the world, they can take those funds, they can take those savings, and they will, and the economy will reward those who then reinvest those funds further into research and development, making their products better. And ultimately that's going to likely result in hiring more people for other jobs and other roles. And whatever these types of macro changes have occurred in the economy, it seems like things just get moved around, and there actually are more opportunities that get created. And so I'm certainly an optimist when it comes to the implications of AI. I would also say, though, I was reading something somewhere that said — I think it was like the last job that'll be replaced is that of an enterprise seller. And unfortunately, I agree with that. Because we can automate a lot of transactional work, but someone convincing another person about doing something, and making an initiative out of it and running a project — I don't see any AI doing that.

直到超级智能到来，那时它们能说服我们任何人做任何事。是啊，但这确实很迷人。我是说，你身处这个新品类，而且如你所说，它是横向的，你能捕获的市场似乎无穷无尽。但与此同时，你是这场全球转型的一部分，而我们真的才刚刚开始。你是其中的一部分。这真的很有意思。那么，超级智能体、Zip 的下一步是什么？

Not until the superintelligence arrives, when they'll be able to convince any of us of anything. Yeah, but it is fascinating. I mean, you're in this new category, and as you say, it's horizontal and seemingly endless how much market you could capture. But at the same time, you're part of this global transformation that we're really only at the very beginning of. And you're part of it. I mean, it's really interesting. So what's next for super agents, for Zip?

我们感到兴奋的是，继续与客户合作，构建越来越多的超级智能体，让它们承担越来越多的角色。我们先优先做了前五个。我们打算接下来优先做十个。在此基础上，我们正在大量招聘更多前置部署工程师。这是我们之前有点落后的地方，目的是与客户合作，用 AI 端到端地改造他们的流程。我一直在为我们在各行业做这类转型的工作感到兴奋，因为我也学到了非常多。但这是多么激动人心的时刻。我同意你的看法。我们进入这场世界性变革其实才四年——差不多四年。虽然现在感觉过了很久，但我确信 10 年后，会发生多得多的事情。

What we're excited about is continuing to work with customers to build more and more super agents that can take on more and more personas. We prioritized the first five. We intend to prioritize the next 10. And then, layering that in, we're very much hiring more forward-deployed engineers. That's something we've been a little bit behind on, to work with customers to transform their processes end-to-end with AI. And I've been excited about the work we're doing with that type of transformation across all industries, because I've also been learning a tremendous amount. But what an exciting time. I agree with you. It's actually only been four years — almost four years — that we're into this transformative change in the world. And while it feels like a long time right now, I'm sure in 10 years there'll be so much more that will have happened.

没错。而且我必须问一下——你和你合伙人有了这个想法。你们是从哪里——招人来做这种东西是很难的。你的团队是从哪里来的？还是说你们俩都是程序员、工程师、开发者，到了能自己做出概念验证的程度？

That's right. And I have to ask — so you and your partner had the idea. Where did you — it's hard to hire people to build stuff like this. Where did you get your team? Or were you guys both coders, engineers, developers to the point where you could build a proof of concept on your own?

当然，过去六年我们招了很多人。但是——

Of course, we've hired a lot over the last six years. But—

我是说一开始。

I mean at the beginning.

对对对。不，一开始我们其实招了一批非常亲近、很棒的工程师、设计师和产品经理，都是我们在 Airbnb 共事过的人——或者我们非常了解的人，我们心想，我们知道他们很棒。

Yeah, yeah, yeah. No, at the beginning we actually brought on a number of really close, awesome engineers and designers and PMs that we worked with at Airbnb — or people that we just knew really well, and we were like, we know they're awesome.

他们当时也非常需要押一注，因为我们那时候基本什么都没有，对吧？没有客户，基本没有产品。所以，我很感激他们当年就愿意冒险，选择和我们合作。我们就是这么起步的。靠的是我们信任、并且知道绝对是 10x 工程师的那些人。

They also very much needed to take a bet too, because we had pretty much nothing happening at the time, right? No customers, basically no product. And so, you know, I'm grateful that they took the leap of faith and chose to work with us way back when. But that's how we got going. It was people that we trusted who we knew were absolutely 10x engineers.

好，我快没什么可问的了。有没有什么我没覆盖到、但你想让听众听到的？

Okay, I'm running out of stuff to ask. Is there anything I haven't covered that you want listeners to hear?

没有，我觉得聊得很好。谢谢，Greg。

No, I think that was great. Yeah, thanks, Greg.