Masterclass: How FDEs make $1M/yr deploying AI agents

频道: Greg Isenberg
视频: https://www.youtube.com/watch?v=1a5HxU52vCQ
原文语言: en
统计: 共 96 轮 · Greg 38 · Vas 51

[0:00] Greg

Why are people getting paid a million dollars as for deployed engineers? It sounds crazy, right? But if you think about it, it isn't. Cuz if you're deploying agents and you're able to drive 5, 10, $25 million of value for companies, [music] would they be willing to give you a slice of that pie? It turns out yes. So the question becomes, how could you become a foreploy engineer? How could you deploy agents to make companies run more efficiently, drive revenue, lift margins? Well, today I brought on Boss from Veric Agents for an inside look at how this works. He's sharing examples from client engagements with the details changed that you just don't get to see publicly. This is for the first time ever on the internet. And that's really [music] cool. I think a lot of people have talked about for deployed engineering on the internet, but they haven't gone into concrete examples for how you [music] can actually do this. By the end of this episode, you're going to understand how to find the work worth automating. [music] You're going to be able to decide where agents belong. You're going to understand open- source versus closed source. [music] You're going to understand where Muse, Grockbot, and Dots been in in all of this and how you can start putting this into practice.

为什么有人当 FDE（forward deployed engineer，前线部署工程师）能拿到一百万美元？听起来很离谱，对吧？但你仔细想想，其实并不离谱。因为如果你在部署 agent，能给公司带来 500 万、1000 万、2500 万美元的价值，他们愿不愿意把这块饼分你一块？事实证明，愿意。所以问题就变成了：你怎么才能成为一名 FDE？怎么通过部署 agent 让公司运转得更高效、拉动营收、抬高利润率？今天我请来了 Varick Agents 的 Vas，带大家从内部看看这件事到底是怎么运作的。他会分享客户项目里的例子——细节做了改动——这些东西你在公开场合根本看不到。这是全网第一次公开，真的很酷。我觉得网上聊 FDE 这门活儿的人不少，但都没有深入到具体例子，讲清楚你实际上该怎么做。看完这期，你会明白怎么找到值得自动化的工作，能判断 agent 该放在哪儿，会搞懂开源和闭源怎么选，也会明白 Muse、Grok Bot 和 Dots 在这整件事里各自处在什么位置，以及你怎么开始把这些付诸实践。

[1:15] Greg

This is a master class for how to become a Ford deployed engineer. I did one other episode with Voss not too long ago, but we kept it high level and by popular demand, we're going deeper. So send this to a friend, like and comment and I'll see you at the end of the episode. One of the most common questions I get nowadays is, Greg, how are you using voice AI in your everyday life? Well, today I'm going to break it down in 60 seconds. I'm going to give you my voice AI toolkit right now. And this section is sponsored by Google. So Google's actually been crushing it with voice AI. Um, I've been using Gemini 3.8 Live recently. So what's really cool about 3.8 Live is you can just go and have a phone call basically with it and say something like, "Hey, you know, how am I doing with my launch campaign?" And it's connected to all my tools and it's basically running my business in the background. So I think about it as if my it's almost like my chief of staff. It's smart. It's intelligent. It's connected to my tools. And it allows me to live my life while having uh voice AI help me run my business. The second is Gemini 3.5 Transcribe. Now, what's really cool about this is I can go and leave voice notes and it'll go and parse those voice notes. It removes the ums, the likes, and you know how I use it is, you know me, I've got a lot of ideas. I got a lot of startup ideas and it just allows me to go on walks and just basically give those those ideas and get back text that is clear. Um, and that also just, you know, allows me to remember cuz frankly I forget things. Um, the third is Gemini 3.8 flash text to speech. Now, what's really cool about this is you can go and say, "Hey, build me a voice that has a Brooklyn accent." Um, and it'll go and do it. And you might be thinking, well, how can you actually use that? Well, you know, there's so many products that need a voice. You know, for me, recently, I've been building a mobile app and I

这是一堂教你怎么成为 FDE 的大师课。前不久我和 Vas 录过一期，但那次聊得比较宏观，应大家的强烈要求，这次我们往深了讲。所以，把这期转给朋友，点赞、留言，我们节目结尾见。最近我被问得最多的问题之一是：Greg，你日常生活里是怎么用语音 AI 的？今天我就用 60 秒讲清楚，现在就把我的语音 AI 工具箱交给你。这一段由 Google 赞助。Google 在语音 AI 上确实做得相当猛。我最近一直在用 Gemini 3.8 Live。3.8 Live 特别酷的地方在于，你基本上可以直接跟它打个电话，说一句「嘿，我那个发布推广活动做得怎么样了？」它连着我所有的工具，等于在后台帮我打理生意。所以我把它当成——差不多就是我的幕僚长（chief of staff）。它聪明、有脑子，连着我的工具，让我一边过自己的生活，一边有语音 AI 帮我经营公司。第二个是 Gemini 3.5 Transcribe。它酷在我可以随手留语音备忘，它会去解析这些语音，把「嗯」「那个」之类的口头语去掉。我是怎么用的呢——你们了解我，我点子特别多，创业点子一大堆——它让我可以边散步边把这些点子说出来，拿回来的是清清楚楚的文字。而且这也能帮我记住事情，因为说实话我挺健忘的。第三个是 Gemini 3.8 Flash text-to-speech（文字转语音）。它酷在你可以直接说「嘿，给我做一个带布鲁克林口音的声音」，它就去做了。你可能在想，这到底能怎么用？其实需要声音的产品太多了。就拿我来说，最近我在做一款手机 app，我

[3:13] Greg

just included an onboarding voice into the mobile app. And the last thing is speechtoech. You know, not everyone speaks English. So, if you're doing business in Mandarin in China, you get this realtime translation via Gemini 3.5 Live Translate and it just opens up new markets. So, these are just four tools that I've been uh obsessed with lately. Uh, shout out again to Google for sponsoring this part of the episode, and I'll include links for where you can play with these tools and models in the description. Have fun with it. Have a creative day, and I'll see you at the next 60cond master class.

就直接在这款 app 里加了一段新手引导语音。最后一个是 speech-to-speech（语音转语音）。要知道，不是所有人都说英语。所以如果你在中国用普通话做生意，可以通过 Gemini 3.5 Live Translate 拿到实时翻译，一下子就打开了新市场。这四个工具就是我最近特别上头的。再次感谢 Google 赞助本期的这一段，我会把能上手玩这些工具和模型的链接放在简介里。玩得开心，祝你今天创意满满，我们下一堂 60 秒大师课见。

[3:53]

[music]

（音乐）

[3:58] Greg

Voss, by the end of this episode, what are people going to learn?

Vas，看完这期，大家能学到什么？

[4:02] Vas

People are going to hopefully learn the endto-end workflows that we're seeing as part of AI transformation and better understand how they can go ahead and do this themselves.

希望大家能学到我们在 AI 转型里看到的那些端到端工作流，也更清楚自己该怎么上手去做这件事。

[4:10] Greg

Yeah, because there's lots of talk and I'm sort of guilty for this too of just keeping it high level. Um, sometimes I'm just like, here's an interesting topic. Here's an interesting business model. Just apply AI. Um, but the big question is, well, how do you actually apply AI? What is this forward deployed engineer piece look like? So, we're going to actually get into the nitty-gritty of that. All right. So, that's the commitment you're going to make to the person listening to this. By the end of this episode, uh, they'll understand what it actually means to apply AI. What does it actually mean to forward deploy into something and where are the business opportunities and monetization opportunities and how big is this thing? Is that the commitment you are going to make today?

对，因为外面空谈很多，我自己多少也有这毛病，老是停在宏观层面。有时候我就是说：这有个有意思的话题，这有个有意思的商业模式，应用一下 AI 就行了。但真正的大问题是：到底怎么应用 AI？FDE 这一块具体长什么样？所以我们今天要真的钻到细节里去。好，那这就是你要对正在听这期的人许下的承诺：听完这期，他们会明白应用 AI 到底意味着什么，前线部署到一个业务里到底意味着什么，商业机会和变现机会在哪儿，这件事的盘子有多大。这是你今天要许下的承诺吗？

[4:57] Vas

100%. You have my word. Let's do it.

100%。我向你保证。来吧。

[5:00] Greg

All right.

好。

[5:02] Vas

Cool. So, you know, I put out an article a few weeks ago um and it was titled don't apply AI. Obviously, a play on, you know, applied AI, etc. And again, the point of that is to to go into, you know, AI isn't something that can be applied like a coat of paint. It's something that really involves process re-engineering. Um, and that's the goal of of today's uh kind of presentation. And thanks for having me on. Um, so as I'm sure everyone already knows, and you know, I know you, Greg, you talked about this a week or two ago. Um, this is already happening, right? Rolloffs is a huge part of the private equity playbook these days where you'll buy a firm that runs very much on people, outdated processes, maybe outdated software as well. Um, for example, accounting firms, IT shops, law practices, etc. And these AI holding companies or AI transformation companies are buying up these, you know, portfolio companies and putting AI engineers forward deployed engineers inside of it. So what they do is they'll find the processes. They'll map out the systems of record, the exceptions, what happens where, what cycle time is occur is recurring as a result of, you know, handoff between two different pods of people. And they'll rebuild that process with agents from the ground up. Um, Thrive, for example, comes to mind. They have 35 engineers across 70 firms as of, you know, me creating this and, uh, doing some research. I'm sure that might be even higher now. Um, and what they're seeing is that it's actually quite successful.

好。是这样，几周前我发了一篇文章，标题叫《Don't Apply AI》（别「应用」AI），显然是拿 applied AI（应用 AI）这类说法玩了个梗。那篇文章想讲的是：AI 不是像刷一层漆那样「涂」上去的东西，它真正牵涉的是流程再造。这也是今天这场分享的目标。也谢谢你邀请我来。我相信大家都已经知道了——Greg，我知道你一两周前也聊过——这件事已经在发生了，对吧？roll-ups（并购整合）是如今私募股权打法里很大的一块：你去买一家非常依赖人力、流程陈旧、可能软件也陈旧的公司，比如会计师事务所、IT 服务公司、律所等等。这些 AI 控股公司，或者叫 AI 转型公司，正在大举买下这类公司作为投资组合公司，再把 AI 工程师——也就是 FDE——放进去。他们的做法是：先把流程找出来，把记录系统（system of record）、例外情况、每件事发生在哪儿都梳理清楚，还有两拨不同的人之间交接拖出了多长的周期时间。然后用 agent 把这套流程从头重建。比如我会想到 Thrive，截至我做这份材料、做调研的时候，他们有 35 名工程师，分布在 70 家公司里，我相信现在这个数字可能更高了。而他们看到的是，这件事其实相当成功。

[6:30] Vas

Uh, the numbers move. You know, tax returns 30% faster at 98% accuracy. Agents are actually able to take work off of people's plates and it is feasible. It's possible. The only caveat being is that it's a lot more involved than people had maybe initially surmised. Um, so one example is gross margin at one call center firm was 60% and above, which is actually quite high. uh and as a result of that you know when they buy the firm let's say they buy it for a billion dollars and they increase the margin uh and they double it for example in the best case um that actually translates into the valuation of the entire company. So all of a sudden you can buy it for a billion implement AI across you know 3 6 12 24 months whatever that is and then you can sell it for 2 billion or 4 billion or 8 billion um and that's really the goal of these of these companies.

数字确实在动。比如报税快了 30%，准确率 98%。agent 确实能把活儿从人手上接走，这事是可行的，做得到。唯一要提醒的是，它比大家一开始以为的要费工夫得多。举个例子，有一家呼叫中心公司的毛利率做到了 60% 以上，这其实相当高。结果就是，他们买下这家公司的时候——假设花了 10 亿美元——把利润率提上去，最理想的情况下比如翻一倍，这会直接体现到整家公司的估值上。于是你可以 10 亿买进来，花 3 个月、6 个月、12 个月、24 个月——不管多久——把 AI 落地，然后以 20 亿、40 亿或者 80 亿卖出去。这才是这些公司真正的目标。

[7:21] Greg

Cool. Let's keep going.

好，我们继续。

[7:22] Vas

So here you can just see a few a few different examples of that. Thrive holdings with Josh Kushner, General Catalyst, AI enabled roll-ups and then the people who are doing those steps of two and three are those four deployed engineers.

这里你可以看到几个不同的例子：Josh Kushner 的 Thrive Holdings、General Catalyst、AI 赋能的 roll-ups。而真正在做其中第二步、第三步的人，就是这些 FDE。

[7:34] Greg

And I think like we'll get into that by the end, but like I think that's the big question a lot of people have is like I think people hear this and they're like, "Okay, cool." But like you're basically saying like buy a business, add AI and you know question mark question mark profit. You know what I mean? That [laughter] mean so I think like the big qu it's like okay but how do you actually do this?

这个我想我们到最后会讲到，但我觉得这正是很多人心里的大问题——大家听到这些会说「好，挺酷」，可你基本上是在说：买一家公司，加上 AI，然后——问号、问号——利润。你懂我意思吧？（笑）所以我觉得最大的问题就是：行，可这到底怎么做？

[7:59]

We will talk about that

这个我们会讲到。

[8:01] Vas

100%. I'm not going to keep it super high level. We're going to get into it. Um and the analogy that I like to give everyone as you're thinking about what it means to be a forward deployed engineer is your life is already complicated right? You have five different inboxes, four file stores, Google Drive, Notion, iCloud Drive, the desktop. You even have five messaging apps. You have iMessage, WhatsApp, Slack, Google Calendar, Outlook Calendar, all this stuff. And it's actually very hard to understand exactly how you like to use your systems, right? So, which app is the easy part. You can make a simple tool call for example to a API uh of, you know, whatever system of record of your choosing. But if a client emails you, you know, it's a Gmail, it's work. But if a family messages you and it's iMessage or it's WhatsApp, um, and understanding this is quite complicated even at a personal level. So imagine at a company and this is to your point, right? This question mark question mark question mark. People are so annoyed at at this like, you know, AI is going to fix everything. Just just use AI because they know that it's actually quite complicated. Um, imagine a company, right? This company has acquired eight other companies in the past. So now they're existing in five different regions. They're in Sa Paulo, they're in Bangalore, they're in Australia, they're in Sydney and they have 23 different systems of record and each region is doing things differently.

100%。我不会只停在很宏观的层面，我们会真的讲进去。在你琢磨当一名 FDE 意味着什么的时候，我喜欢给大家打这么个比方：你自己的生活就已经够复杂了，对吧？你有五个不同的收件箱，四个存文件的地方——Google Drive、Notion、iCloud Drive、桌面。你甚至有五个通讯类 app——iMessage、WhatsApp、Slack、Google Calendar、Outlook Calendar，诸如此类。而要说清楚你到底习惯怎么用这些系统，其实非常难。所以，「用哪个 app」反倒是容易的部分——比如你可以对你选的任意一个记录系统的 API 发起一次简单的 tool call。可是，客户给你发邮件，那是 Gmail，是工作；家人给你发消息，那就是 iMessage 或者 WhatsApp。光是在个人层面，把这些弄明白就已经相当复杂了。那再想想公司层面——这正好回到你刚才说的那个「问号、问号、问号」。大家对「AI 会解决一切，用 AI 就行了」这种说法特别烦，因为他们知道实际情况相当复杂。想象一家公司，对吧？它过去收购过另外八家公司，所以现在分布在五个不同的地区：圣保罗（São Paulo）、班加罗尔、澳大利亚、悉尼，他们有 23 套不同的记录系统，而且每个地区的做法都不一样。

[9:19] Vas

Chicago is on SAP, Salesforce and Workday. Toronto is on Netswuite, HubSpot and ADP and so on and so forth. So really it is quite complicated and this is exactly why again AI cannot be applied. If you're applying AI over this entire company, which literally spans the entire globe, you're going to end up with just making faster. I hope I can curse on this podcast and bleep that out with you.

芝加哥用的是 SAP、Salesforce 和 Workday，多伦多用的是 NetSuite、HubSpot 和 ADP，如此等等。所以这事确实相当复杂，这也正是为什么我一再说，AI 没法「应用」上去。如果你把 AI 往这样一整家公司上一套——它可是真的横跨全球——最后你只是把（哔）变得更快而已。希望我能在这个播客里爆粗，你们帮我哔掉就行。

[9:43] Greg

Absolutely. [laughter]

当然可以。（笑）

[9:45] Vas

Um, and that's the fundamental issue. So now, again, I don't want to, you know, be too uh beating a dead horse on the problem. The problem is very clear. It's very complicated. So, how do you actually go about doing this? This is our view. And I'm sure there's many different ways of doing this, but our view is process mapping, then re-engineering, then building, deploying, and rolling it out. Um, and there's a few different ways that we do that. So, if you were to go into a company on day zero, here's what I would suggest that you do.

这就是根本问题所在。不过我也不想在问题本身上翻来覆去地说了。问题很清楚：它非常复杂。那到底该怎么着手做呢？这是我们的看法——我相信做法有很多种——但我们的看法是：先做流程测绘（process mapping），再做流程再造，然后是构建、部署、推广铺开。这里面我们有几种不同的做法。所以，假如你在第零天走进一家公司，我建议你这么做。

[10:17] Greg

Mhm.

嗯。

[10:18] Vas

The first is interviews on the human side. So, let's start off with a single department, right? You'll have in finance, you'll talk to the head of AP, AR, reconciliations, billing, banking, FPNA, etc., and you'll work your way down from there. And the reason why these interviews are super helpful is because a lot of information lives in their heads. Very rarely do you have a very clean document source that you can just point your agent at, it'll learn it, and go from there. Very often, it's not written down. It's not documented. And we hear this all the time when we talk to companies. uh they tell us saying, "Oh yeah, this person's been at the firm for 20 years." And they just handle it. It's so common. It doesn't matter the size of the company. Could be a massive Fortune50. It could be a small SMB. They all have these critical key people where information lives in their head.

第一件事是「人」这一侧的访谈。我们先从单个部门开始，对吧？比如财务部，你要去跟 AP（应付账款）、AR（应收账款）、对账、开票、银行业务、FP&A（财务规划与分析）等等的负责人聊，然后从那儿一层层往下。这些访谈之所以特别有用，是因为大量信息都活在他们的脑子里。你很少能碰到一份干干净净的文档，让你把 agent 往上一指，它学完就能接着干。很多时候，这些东西根本没写下来，没有文档。我们跟公司聊的时候老听到这种话，他们会说：「哦对，这个人在公司待了 20 年了。」这事就归他们处理。这太常见了，跟公司规模无关——可以是巨型的 Fortune 50，也可以是一家小小的 SMB——他们都有这么几个关键人物，信息就装在这些人脑子里。

[11:06] Greg

Yeah. I mean, people, if you think that you're going to walk into a company and they're going to have like an obsidian second brain hooked up to AI agents, you are just mistaken. my friend. You know what I mean? Like

对。我是说，各位，如果你以为走进一家公司，他们会有一套接好了 AI agent 的 Obsidian 第二大脑，那你可就大错特错了，朋友。你懂我意思吧？就是

[11:21] Vas

100%.

100%。

[11:22] Greg

There is no second brain happening here. [laughter]

这儿可没有什么第二大脑。（笑）

[11:28] Vas

Yeah. There's nothing you can just plug into call an API and all of a sudden you have your knowledge store.

对。没有哪个东西是你接上去、调一下 API，一下子就有了知识库的。

[11:34] Vas

It's kind of on you to create that and it's per department and it's cross department and it's very very involved.

这东西得靠你自己去建，既要按部门来，又要跨部门，非常非常费工夫。

[11:41] Greg

Okay. So step one is basically [sighs] in a sense it's creating like a human API. It's getting all the uh you know uh data systems SOPs you know all that stuff into modern digital systems.

好。所以第一步基本上就是——某种意义上，是在造一个「human API」。把所有的数据、系统、SOP 这些东西，全都搬进现代的数字化系统里。

[12:02] Vas

Absolutely. You always have a good way of articulating I like that it's a human API. Um that's what it is and you have to understand you know why do they do things the way they do them today? Who really decides them? Which step is theater? Which step is legitimate? What happens when exceptions take place? how often do exceptions take place? Um, all of these things are not written down. From there, we move on to then mining the systems of record and mining everything else. Right? So, really there's a lot of data that exists in their Salesforce, their Netswuite, their dynamics, their workday, etc. And that doesn't mean it's written down, right? It's not in a document which has like once step one, step two, etc. But you can sort of create that SOP by living on top of their systems of record. So for example, if I have real- time access to a Salesforce for example over the course of you know 3 to 4 weeks I have a pretty good understanding of what sort of data enters Salesforce, how often is it getting corrected etc. And this runs constantly and you can use AI to analyze what the hidden meaning is behind each one of those actions. Right? So if a certain record comes in and it's from a certain company, it's of a certain size, we see these actions taking place and now we can create sort of a graph of okay once something happens, once something enters of a certain category, A goes to X, B goes to Y, C goes to Z, etc. But obviously if you just do that without the interviews, you're missing half the picture.

完全正确。你总是很会概括，我喜欢这个说法——human API，就是这么回事。你得弄明白：他们今天为什么这么做事？到底是谁在拍板？哪一步是走过场，哪一步是真有必要？出现例外时会怎样？例外多久出现一次？这些统统没有写下来。在这之后，我们再去挖掘记录系统，以及其他所有东西。他们的 Salesforce、NetSuite、Dynamics、Workday 等等里面其实有大量数据。但这不等于它被写下来了，对吧？它并不在一份写着第一步、第二步……的文档里。不过你可以靠「住」在他们的记录系统上面，把那份 SOP 差不多还原出来。举个例子，如果我能实时访问某个 Salesforce，三到四周下来，我就能相当清楚：进 Salesforce 的是什么样的数据、多久被修正一次，等等。这个过程是持续在跑的，你可以用 AI 去分析每一个动作背后隐藏的含义。比如某条记录进来了，来自某家公司、属于某个规模，我们看到随之发生了这些动作，于是就能画出一张类似图谱的东西：一旦某件事发生、一旦某个类别的东西进来，A 流向 X，B 流向 Y，C 流向 Z，等等。但显然，如果你只做这个而不做访谈，你就少了半幅图景。

[13:29] Vas

The final aspect is what lives outside of them. So then you actually go into the existing documentation. Sometimes outdated, sometimes it's pretty good. This is in SharePoint. This is in Drive. This is in notion. This is in Slack. This is in Teams. This is in Gmail. This is in spreadsheets. And again, through these three steps, you have literally the entire company's picture. And it varies the split amongst companies, right? For a very large company, they've had 10 years of Salesforce historical data for you to go off of. For an SMB, they probably don't. They probably have mostly in people's heads in interviews, um, and no processes, no system of record, no software. So it's up to you to determine what angle you need to take per company and it varies

最后一块，是存在于它们之外的东西。这时候你就真的要去翻现有的文档了，有的已经过时，有的还挺不错。这些东西在 SharePoint 里，在 Drive 里，在 Notion 里，在 Slack 里，在 Teams 里，在 Gmail 里，在各种电子表格里。同样，走完这三步，你就真的拿到了整家公司的全貌。而这三块的比重，各家公司不一样，对吧？一家很大的公司，有 10 年的 Salesforce 历史数据可以给你当依据。SMB 多半就没有，他们的东西大概主要在人的脑子里、得靠访谈挖出来，没有流程，没有记录系统，没有软件。所以每家公司该从哪个角度切入，得由你自己判断，而且各家情况不一样

[14:10] Greg

crystal clear.

非常清楚。

[14:12] Vas

So here's a very concrete example. You wanted to go away from high level. This is what we're trying to do. Um this is a public software company. This is actually an engagement that we that we completed. Um so it was a $5 billion in revenue company. They have over 150 products, tons of solution consultants and we were brought in by the CRO and they're public. So this actually makes them quite complicated. you have to deal with regulation uh certain laws apply etc. So the process document which they actually had and they had completed this I think it was engagement with deote that mapped this out for them um was you build a quote then you submit it then it goes to deal desk then you approve it then you send it you negotiate then you sign it very cut and dry and this is always what happens if you just look at one angle even just one person right they might tell you the wrong story so you have to interview other people as well um but what the reality was and this was as a result of process mining agents that we had deployed over their CRM which I believe was Salesforce. Um it's actually a 20step process with seven different loops. So from step one to step, you know, four or five, then you loop back to one if there's an issue. Um and that's 61% of requests actually follow that loop. Then later down the chain, legal sends it back 12% of the time.

来看一个非常具体的例子。你刚才说想别停在高层面，那我们现在就来做这件事。这是一家上市软件公司，是我们实际做完了的一个项目。这家公司营收 50 亿美元，有 150 多款产品，一大堆解决方案顾问，是 CRO 把我们请进去的。而且他们是上市公司，所以情况其实相当复杂——要应对监管，有些法律要遵守，等等。他们其实是有流程文档的，我记得是之前跟 Deloitte 做过一个项目，帮他们把流程测绘了出来。文档上写的是：先做报价，然后提交，然后到 deal desk（交易审批台），然后审批，然后发出去，谈判，最后签约。干净利落。如果你只从一个角度看，甚至只问一个人，结果总是这样，对吧？他们讲给你的可能是个不对的版本，所以你还得去访谈其他人。但实际情况是——这是我们在他们的 CRM 上部署了流程挖掘（process mining）agent 之后得出的结果，我记得那套 CRM 是 Salesforce——这其实是一个 20 步的流程，里面有七个不同的回路。比如从第一步走到第四、第五步，一旦有问题，就绕回第一步。实际上有 61% 的请求都会走这个回路。再往后，法务有 12% 的情况会把它打回来。

[15:26] Vas

Later down the road, 30% of the time a new quote has to repeat uh getting approval, etc., etc. So none of this was really documented and it was up to us and actually my team of four deployed engineers to go in and create this process mapping.

再往后，有 30% 的情况新报价得重新走一遍审批，诸如此类。这些其实都没有成文记录，所以得靠我们——其实就是我那支 FDE 团队——进去把这份流程图梳理出来。

[15:39] Greg

It's interesting like the way I'm thinking about it is every business is sort of like a factory and a factory if you think about imagine you're looking down at a factory floor. You know there's an assembly line. There's different parts of the assembly line. And there's probably maybe some offices where people are doing some accounting or you know and they all kind of work together to create a product. And what's really cool about what you're doing and just for deploy engineering as a service in general is you're basically saying like how do I distill every business down to these systems or set of systems uh per department and then you're you're basically what's really cool is like now well we have everything we need to actually get the you you usually the work done. There's the dig step one is like the digital tools like the sales forces, the net suites, all those products. Then there's the agents that actually do the work and then there's the human beings. There's often times like a human being step to it, right? Like not everything could be fulfilled by agents. So what you're doing here is you're kind of like exposing the whole system. um you're acknowledging that a lot of people think that they have a full system but it's actually usually just the tip of the iceberg and you're basically saying how can I open up this system optimize it and then I would imagine like have eval or um which you can talk about but basically make sure that you know it's doing its job at be uh you know as good as as possible.

有意思。我是这么想的：每家公司都有点像一座工厂。你想象自己从上往下俯视工厂车间，有一条装配线，装配线上有不同的环节，旁边可能还有几间办公室，有人在里面做账之类的，所有这些配合在一起，造出一个产品。你做的这件事——或者说整个「forward deployed engineering 即服务」——特别酷的地方在于，你其实是在问：我怎么把每家公司都提炼成这样一套系统，或者说每个部门的一组系统？然后更酷的是，现在我们真的把干成这些活所需要的东西都凑齐了。第一层是数字化工具，像 Salesforce、NetSuite 这类产品；然后是真正干活的 agent；再然后是人。很多时候里面总有一步得由人来做，对吧？不是什么事都能交给 agent 完成。所以你在这里做的，相当于把整个系统摊开来。你点破了一件事：很多人以为自己有一套完整的系统，但那其实往往只是冰山一角。你其实是在问：我怎么把这套系统打开、优化它，然后——我猜——还要有 eval（评测）之类的东西，这个你可以展开讲讲，总之就是确保它把活干得尽可能好。

[17:37] Vas

Yeah, 100%. And it's funny that you mentioned that like the mapping of the processes let's say across a sales department as we have here. Um it's it's actually never been shown before to anybody in the department

对，100%。你提到这个挺有意思的——像这样把流程测绘出来，比如我们这里展示的整个销售部门的流程，其实从来没有人给这个部门里的任何人展示过

[17:50]

so cleanly

这么清晰。

[17:52] Vas

to the point where you know the CRO and CFOs are telling us like I I feel like you understand our department better than we do. And it's true because no one has done that yet. And that's the real issue, right? If you go and say, "Hey, we want to use AI." Well, on what? What are we doing? What what is the broken process? What's the problem we're trying to solve? And that's why you need this step. And and then to your point of, you know, not everything can be agentic, some things should be deterministic, etc. These are the 20 steps, and this is what we bucket them into, right? There's four buckets. One is delete the step. This shouldn't exist in a post AAI world. Um, some steps are plain code where you have, you know, rules. There's no judgment required. For example, it's a simple API call. If this happens, then this happens. Five are agentic, right? You have building a quote that requires some level of judgment with who's the customer, what's it worth to us, how much do we need to spend, what resources do we need to allocate, etc. And then finally to your point, humans in the loop, human decision makers who are going to be handling the most risky tasks, things that you really cannot afford to, you know, get wrong and and it's approval, it's negotiation, it's signing, it's submitting payment, etc.

以至于 CRO 和 CFO 们会跟我们说：「我感觉你们比我们自己还懂我们部门。」这是真的，因为之前没人做过这件事。这才是真正的问题所在，对吧？你跑去说「嘿，我们想用 AI」——好，用在哪儿？我们到底在做什么？哪个流程是坏的？我们要解决的问题是什么？所以才需要这一步。再接你刚才说的那点——不是什么都能交给 agent，有些东西就该是确定性的，等等。这就是那 20 个步骤，我们把它们分到这几个桶里，对吧？一共四个桶。第一个是把这一步删掉——在 AI 之后的世界里它就不该存在。有些步骤就是纯代码，有规则可循，不需要判断。比如就是一次简单的 API 调用：发生了这个，就执行那个。有五步是 agentic 的，对吧？比如做报价，需要一定程度的判断：客户是谁、这单对我们值多少、我们需要花多少、要调配哪些资源，等等。最后就是你说的，human in the loop（人在回路）——由人来做决策，处理风险最高的任务，那些你真的出不起错的事：审批、谈判、签约、提交付款，等等。

[19:04] Greg

Yep.

嗯。

[19:04] Vas

So this is also how we do it. And this is one thing that I really want to call out, right? A lot of people think of AI agents as a new surface and we take the opposite approach and this is what I've seen really resonate with a lot of executive leaders and it's what I strongly recommend to anybody who wants to be an FTE is pitch yourself as building these agents inside their systems of record. So one thing that we do is for example we'll have agents that mine your Salesforce and take action in your Salesforce. Uh, and there's no new surface that you need to go into to interact with this. We're not asking you to replace your Salesforce, your CRM. That's impossible. Real quotes from our customers. They spent several years and several million dollars, I think $10 million one time on migrating from one ERP to the next. Same thing for, you know, one CRM to the next, etc. If your pitch for AI is, hey, we're going to move you from Salesforce to this AI native CRM, you've lost them. And you know, maybe in SMB where they don't have a CRM, you can put them on one. That's great. But otherwise, try to be inside of systems of record. Even here when you have human in the loop, it's a message in Slack. And that's what we've really seen resonate with our clients as well. Um, so I strongly recommend this being the case. And this is also how you don't have to retrain staff, right? they are already used to this. You're working inside their their records and you can just hit the ground running with a much faster um rate of you know utilization, much higher efficiency, etc.

这也是我们的做法。有一点我特别想强调，对吧？很多人把 AI agent 当成一个新的界面（surface），我们的做法正好相反。我发现这一点特别能打动企业高管，也是我强烈建议所有想做 FDE 的人去做的：把自己定位成「在他们的记录系统里面搭 agent」。举个例子，我们会让 agent 去挖掘你的 Salesforce，并且直接在你的 Salesforce 里执行操作。你不需要进到任何新的界面去跟它交互。我们不会要求你把 Salesforce、把你的 CRM 换掉，那是不可能的。这是我们客户的原话：他们花了好几年、好几百万美元——我记得有一次是 1000 万美元——就为了从一套 ERP 迁到另一套。从一套 CRM 换到另一套也是一样，等等。如果你推销 AI 的说法是「嘿，我们要把你从 Salesforce 搬到这个 AI 原生的 CRM 上」，那你已经失去他们了。当然，如果是还没有 CRM 的 SMB，你可以给他们上一套，那挺好。但除此之外，尽量待在记录系统里面。就连这里需要 human in the loop 的环节，也只是 Slack 里的一条消息。我们也确实看到这一点很能引起客户的共鸣。所以我强烈建议这么做。而且这样你也不用重新培训员工，对吧？他们本来就习惯这些了。你就在他们自己的记录系统里干活，一上来就能跑起来，使用率上得快得多，效率也高得多，等等。

[20:38] Greg

Yeah. I mean in general like you're trying to sell anything to anyone like you want the path of least resistance, right? So, um I think that if you're going to ask them to completely move softwares and then kind of like introduce all these new concepts to them because they probably haven't heard of a lot of these I mean maybe some of them maybe have but some of them probably haven't around this like whole agentic world that we're living in. So, uh, yeah, my point my take here is like or I'm just agreeing with you like obviously it makes sense like if especially if someone's listening to this and like wants to be a forward deployed engineer or wants to like start a forward deploy engineering business like path of least resistance.

对。我是说，一般来讲，不管你要把什么东西卖给谁，都要走阻力最小的那条路，对吧？所以我觉得，如果你要他们把软件整个换掉，还要给他们灌输一大堆新概念——因为这里面很多东西他们大概听都没听过，可能有些人听过，但有些人多半没听过，就是我们现在身处的这一整个 agentic 世界。所以，嗯，我的观点是——或者说我就是在赞同你——这显然说得通。尤其是如果有人正在听这期节目，想当 FDE，或者想做一门 FDE 服务的生意，那就是：走阻力最小的路。

[21:29] Vas

Yeah, 100%. They haven't heard of Jev or Muse or any of his stuff, right? So, just, you know, keep it simple.

对，100%。他们没听说过 Jev、Muse 这些东西，对吧？所以，保持简单就好。

[21:36] Greg

Yeah.

对。

[21:37] Vas

Um, here's another example, right? And we mentioned there's a lot of PE firms that are looking to uh identify their their stack. For for anyone who's not very familiar, a private equity firm will have ownership um in dozens of companies, right? And their mandate is all right, let's roll out AI across all of them. And it's impossible to do so, right? If you have 26 different companies, that's 26 different engagements. Um and then each one of those companies 26 has 10 different departments. Each department has 10 different workflows. and all of a sudden you'll need to deploy 50,000 Ford deployed engineers across 3 years if you want to even make a dent. Um so the way that we recommend going about this is grouping together portfolio companies. Um often times it's on systems of record. So for example, if we have a port a PE firm with 26 different companies and we're tackling finance for all of them, we'll group together the you know five that are on Netswuite as their ERP and four that are in Dynamics. So if you're an FTE, your job is to make sure you really understand what capabilities already exist in each software. So what does Netswuite offer? What does it not? What does Dynamics offer? What does it not? And then where can you fill in the gaps and how do you make that talk to whatever else they're on which is ramp or bre or tipi or blackline or concur or expensify. On the sales operation side it means some are on Salesforce, some are on HubSpot. Um how do you integrate that with everything else across their stack and their spreadsheets etc. You get the idea. Um, and what happens is if you go about it this way is you only have to tackle similarities, right? Instead of getting pulled in different directions, you just streamline with one, you know, entry point and that's a system of record. That's what we strongly recommend, especially if you're an FDE.

再看一个例子，对吧？我们前面提到，有很多 PE 机构想把自己的技术栈 AI 化。给不太熟悉的朋友解释一下：一家私募股权机构会在几十家公司里持有股权，对吧？他们给自己定的任务是：好，那就在所有公司里把 AI 全面铺开。可这根本做不到，对吧？你有 26 家不同的公司，那就是 26 个不同的项目。这 26 家公司每家又有 10 个不同的部门，每个部门有 10 条不同的工作流。一下子，你哪怕只想做出一点成效，都得在 3 年里派出 50,000 名 FDE。所以我们建议的做法是把投资组合公司分组，很多时候是按记录系统来分。比如，一家 PE 机构旗下有 26 家不同的公司，我们要给它们全部做财务这一块，那我们会把其中 5 家用 NetSuite 做 ERP 的归成一组，4 家用 Dynamics 的归成一组。所以如果你是 FDE，你的工作就是要真正搞清楚每款软件里已经有哪些能力。NetSuite 能做什么、不能做什么？Dynamics 能做什么、不能做什么？然后你能在哪儿把缺口补上？又怎么让它跟他们在用的其他东西打通——Ramp、Brex、Tipalti、BlackLine、Concur 或者 Expensify。到了销售运营这边，就是有的用 Salesforce，有的用 HubSpot。你怎么把它跟他们技术栈里的其他东西、跟他们的电子表格等等整合起来。你明白这个意思了。这么做的结果是，你只需要处理共性的部分，对吧？不用被扯向四面八方，而是用一个入口把事情理顺，这个入口就是记录系统。这是我们强烈推荐的做法，尤其如果你是 FDE。

[23:32]

If you're being asked to do 10 different companies, you have to simplify it for yourself. Um, it also simplifies the politics. You know, you're working with CFOs uh with the same buyer over and over again. It's interesting that the CFO is the sponsor. Like it's, you know, you would think that it might be the CTO or or someone just from technology.

Vas： 如果有人要你同时做 10 家不同的公司，你必须给自己把事情简化。这样也能简化内部政治。你打交道的都是 CFO，一遍又一遍面对的是同一类买家。

Greg： 有意思的是，出面牵头的 sponsor 居然是 CFO。你会以为应该是 CTO，或者技术那边的什么人。

[23:57] Vas

Yeah, we we very often get brought in with CIOS um and then that's just from like the we want to identify everything perspective. Um, but also it's like CEOs, CFOs, CRO who are like, well, this is my department and I want something there. Now, caveat being CFOs are self-proclaimedly notoriously skeptical, so good luck selling into CFOs, but um, that's just how it goes.

对，我们经常是被 CIO 请进去的，那是出于「我们想把一切都 AI 化」这个角度。但也有 CEO、CFO、CRO 这些人，他们会说：这是我的部门，我想在这儿上点东西。不过要提醒一句：CFO 自己都承认他们是出了名的多疑，所以想卖给 CFO 的话，祝你好运。但事情就是这样。

[24:25] Greg

Um, yeah. I mean, to me, don't want to offend any CFOs listening to this, but to me, like CFOs just care about optimizing costs.

嗯，对。在我看来——我不想得罪任何正在听这期节目的 CFO——但在我看来，CFO 就只关心优化成本。

[24:37] Greg

So, when they hear, you know, agentification or agents, they're just thinking, how do I lower my costs, increase margin? Um, are you finding that's like the best way to get in? like, hey, like we're going to like deploy all this stuff and like you're going to save a bunch of money. Is that how you how you're thinking about it or am I missing it something?

所以他们一听到 agent 化、agent 这些词，想的就是：我怎么把成本降下来、把利润率提上去？你有没有发现这就是最好的切入方式？就是说：嘿，我们会把这一堆东西部署上去，你能省一大笔钱。你是这么考虑的吗？还是我漏了什么？

[25:02] Vas

No, I would say that's that's largely correct. Um, we obviously deliver value on three buckets, right? One is cost savings, but the second is revenue uplift and third is risk mitigation. Um and what we've seen actually resonate with CFOs is yes cost cutting absolutely uh first and foremost but then also like how long does it take you to close your books and we've heard like about 22 days uh 4 weeks 6 weeks well okay that's that's over a month that's it's called month end close what are we talking about here um so if we can bring that down to 4 days or 8 days with higher accuracy at the same cost you're running it today even they've really found that to be useful and then obviously further if we can cost cut um you know down the line that's even better but it's a bit of both it's not just that we want to keep our existing slop but cheaper they actually do want to move towards faster more accurate etc

没有，我觉得你说的基本没错。我们交付的价值显然分三个桶，对吧？第一是节省成本，第二是营收提升，第三是风险缓释。我们看到真正能打动 CFO 的，没错，首先肯定是削减成本，但还有一个：你们关账要花多久？我们听到的回答有大概 22 天、4 周、6 周。好吧，那都超过一个月了——这东西叫「月末关账」（month-end close）啊，这是在说什么呢？所以如果我们能把它压到 4 天或者 8 天，准确率还更高，哪怕成本跟你现在跑的一样，他们也觉得这真的很有用。当然，如果往后还能进一步削减成本，那就更好了。但这是两方面都有：他们并不是只想把现有的那堆 slop（糙活）原样留着、只求更便宜，他们确实想往更快、更准这些方向走。

[26:01] Greg

it's I mean maybe it's just like you speak to who you're selling to so if you're speaking to like a CIO or CTO it's like the efficiency maybe it's the efficiency, it's it's you know the output, it's it's productivity, it's stay up to date, you know, it's all of that, but like then it gets like handed over to the CFO. It's like they might not care as much about the efficiency in terms of like or the output that it's like way cleaner and nicer. They might just care about like the revenue uplift and and and uh lowering cost. So, I mean, obvious to say, but like that's, you know, for people listening, it's like sell to who you're you're speaking to.

我的意思是，也许说到底就是看人说话——你卖给谁，就讲谁听得进去的话。如果对面是 CIO 或者 CTO，讲的可能就是效率，是产出，是生产力，是跟得上最新进展，这一整套。可这事接着会被交到 CFO 手里，他们可能就没那么在乎效率，或者产出是不是干净漂亮得多，他们在乎的可能就是营收提升和降成本。这话说出来有点像废话，但对正在听的人来说，就是：对面坐的是谁，你就卖给谁。

[26:48] Vas

Yeah. 100%. Sell the outcome, right? If we're talking to a CHRO or a chief people officer,

对，100%。卖的是结果，对吧？如果我们面对的是 CHRO 或者首席人才官（chief people officer），

[26:53] Vas

it's you can hire better people faster and train them quicker on day one. It's not about the cost. They don't want to save money here. They want to get way better output. Um, so it it there is

那要讲的就是：你能更快招到更好的人，从入职第一天起就更快把他们带出来。重点不在成本。他们在这儿不是想省钱，他们要的是好得多的产出。嗯，所以这里面确实有

[27:06] Vas

cool. So here's again deep dive into a concrete example. Um five portfolio companies all on Netswuite but even then they have very different ways of running things. Uh you know 12 steps, 9 steps, 15 steps, 18 steps, 13 steps and then imagine you know they have regional differences, regional variances at each portfolio company level. So again this is why mapping it out is so important and this is useful not just for each portfolio company, right? Each CFO has the same, you know, uh, investment that we talked about earlier where they see this and you understand their department better than they do, but the PE firm does as well where they can see things get mapped out and streamlined. And what we've seen is a lot of PE firms ask us for okay, what's the playbook, right? Tomorrow when they acquire a new company, what process should they follow? Um, how should they go about it? What software should they adopt? Um, and this mapping of saying, "Hey, look, if you're on Netswuite, this is the concrete way of doing things. Here you go. You turn this into six steps, agents in certain locations, etc." Now, this is obviously a dramatization. You will very rarely get to as clean as like, hey, everyone's on six steps and we did it, guys, perfect efficiency. But you'll actually come quite close to this and you'll get that by being very deeply involved mapping it out working with the stakeholders then re-engineering it with um a lot of foresight and a lot of uh thought thoughtfulness. So again this is actually and this is anonymized obviously we can't share you know details about our clients actual workflows but

好。那这里还是深入看一个具体例子。嗯，五家投资组合公司，全都在用 NetSuite，但即便如此，它们各自的运转方式也大不一样。你看，12 步、9 步、15 步、18 步、13 步，然后再想想，每家投资组合公司这一层还有地区差异，各地区的做法又不一样。所以还是那句话，这就是为什么把流程测绘出来这么重要。而且它不光对每家投资组合公司有用，对吧？每位 CFO 都会有我们前面聊过的那种投入感——他们看到这张图，发现你比他们自己还懂他们的部门；PE 机构也一样，他们能看到事情被梳理出来、被理顺。我们看到很多 PE 机构会来问我们：好，那打法手册（playbook）是什么，对吧？明天他们再收购一家新公司，该照什么流程走？嗯，该怎么着手？该采用什么软件？而这份测绘就是在说：「你看，如果你用的是 NetSuite，这就是具体的做法，给你。把它变成六步，在某些位置放上 agent，等等。」当然，这显然是夸张化了的。你很少能做到这么干净，像是「嘿，所有人都变成六步了，伙计们我们做到了，效率完美」。但其实你能做到相当接近，而靠的是：非常深地参与进去，把流程测绘出来，和各方利益相关者一起做，然后带着大量的预见和深思熟虑去做流程再造。所以还是那句话，这其实是——当然这是做过匿名处理的，我们没法分享客户实际工作流的细节，但

[28:46] Vas

this is a real process mapping for I think it was accounts it was accounts payable yeah

这是一份真实的流程测绘，我记得是应付……对，是应付账款（accounts payable）。

[28:50] Vas

um and it was 17 different steps with exceptions being handled right so these are all like exceptions and these are all the steps etc um and this we uncovered over the course of I believe 2 3 weeks with interviews process mining etc um and this is actually a less complicated you know workflow for them and in general. We've seen workflows that are 40 steps or 200 steps. Um, it gets crazy. But your job as an FTE is to one, map this out. Don't skip this step. Don't skip educating the client on on what their, you know, mess is. They want to see this. It breaks their heart, but they need it. Um, and then you turn it into here's the agentic future. It one visually looks much cleaner so they can take a deep sigh of relief and two it actually runs much smoother where you have agents that are handling what they need to handle and you have humans in the loop where needed etc. Um these green boxes are you know steps after I think human decisions um and this is still like an exception where like you know determines the code etc. So, this is your job as an FTE.

嗯，一共是 17 个不同的步骤，还带着例外处理，对吧？这些全是例外，这些全是步骤，等等。嗯，这是我们花了，我记得是两三周，靠访谈、流程挖掘（process mining）这些手段摸出来的。嗯，而且这在他们那里、乃至总体来看，其实还算是不那么复杂的工作流。我们见过 40 步的、200 步的工作流。嗯，会变得很离谱。但你作为 FDE 的工作，第一是把它测绘出来。别跳过这一步。别跳过让客户看清自己这摊乱局到底是什么样的这一步。他们想看到这个。看了会心碎，但他们需要。嗯，然后你再把它变成：这是 agent 化之后的未来。第一，它看上去清爽得多，他们可以长长地松一口气；第二，它实际跑起来也顺畅得多，该 agent 处理的由 agent 处理，需要的地方有人留在回路里把关，等等。嗯，这些绿色方框，我记得是人工决策之后的步骤，嗯，而这里仍然算一个例外，比如要确定编码之类的。所以，这就是你作为 FDE 的工作。

[30:00] Greg

By the way, if you're selling this to a CFO, the way to do this is like you have the agents there and then you put the estimated cost per month of the agents because it's going to be like shockingly low, right?

顺便说一句，如果你是把这个卖给 CFO，做法就是：把 agent 摆在图上，然后标上这些 agent 每个月的预估成本，因为那个数字会低得吓人，对吧？

[30:13] Greg

Relative to human beings

跟人比起来的话。

[30:15] Vas

100%. And you have the same thing with like the accuracy and all the KPIs that you can throw at them. They love that. Any sea suite, right? Um, and part of your job is based on those KPIs. here's how bad it runs now and here's how it's going to run in the future and then you hold yourself to that standard. So 6 months from now you can say I did this.

100%。准确率也是一样，凡是你能甩给他们的 KPI 都一样。他们特别吃这一套。哪个 C-suite 高管都是，对吧？嗯，你的工作有一部分就是拿这些 KPI 来说话：现在跑得有多糟，将来会跑成什么样，然后你就用这个标准来要求自己。这样 6 个月之后你就可以说：这是我做到的。

[30:35] Greg

Love it.

太棒了。

[30:36] Vas

And you you keep you keep stealing my thunder, Greg. That's exactly what we're showing here. Um so you know you show them you know from 17 process steps to seven cycle time from 24 days to six exceptional loops from 6 to 1. This is a really good one. The next two, the straight through rate of an invoice, we drove that from 18% to 87% for this client. And that was a gamecher for them because all of a sudden, literally a majority of their invoices were going through exception routes. Like that's terrible. That means you have really bad processes. And we fixed that process. That's a process re-engineering flow, by the way. That's not even um all about agents. And then finally, we drove the cost of handling a single invoice down from $31 to $6. It's an 80% reduction. Um, so to your point, right, you show them that in this slide,

Greg，你老是抢我的风头。我们这一页展示的正是这个。嗯，所以你要给他们看：流程步骤从 17 步降到 7 步，周期时间（cycle time）从 24 天降到 6 天，例外回路从 6 个降到 1 个。这一项真的很漂亮。再看接下来两项：发票的直通率（straight-through rate），我们帮这家客户从 18% 拉到了 87%。这对他们来说是颠覆性的，因为说白了，他们大多数发票真的都是在走例外路径。这太糟了，这说明你的流程非常差。而我们把这个流程修好了。顺便说一句，这属于流程再造的活儿，甚至都不全是 agent 的事。最后，我们把处理单张发票的成本从 31 美元降到了 6 美元，降幅 80%。嗯，所以回到你说的，对吧，你就在这页幻灯片上把这些给他们看，

[31:25] Greg

right? And to the people who are like, Voss is just replacing human beings with agents. The other piece of this is if you're able to optimize a company such that their cost per invoice is going from 31 to $6. Now, all of a sudden, that company has more margin. Yes, they might take some of that margin, but they also might give back some of that margin to customers.

对吧？至于那些会说「Vas 不过是在拿 agent 替换人」的人——这件事还有另一面：如果你能把一家公司优化到单张发票成本从 31 降到 6 美元，那这家公司一下子就有了更多利润空间。没错，他们可能会把其中一部分留下，但也可能把一部分利润让回给客户。

[31:51] Vas

Absolutely. And also use that margin to hire because

完全没错。而且还会拿这部分利润去招人，因为

[31:56] Greg

right

对。

[31:56] Vas

the truth be told, right, the the clients that we work with are very often like Fortune 1000, Fortune 500. Um they're they want to win, they want to grow. Uh and we've very often seen that it's reallocation of resources. It's not about doing mass layoffs, right? They would rather have their best people in finance, not spend their time doing manual invoice routing and approvals and parsing of an invoice. That's ridiculous. They would rather have those people on higher leverage tasks, right? Planning out FPNA certain aspects of that um or migrating them across cross functional or building their own FTE teams. Um, so I hope that no one is under the impression where these are, you know, going to replace everyone's job. Yes, there will be migration of job. Um, but I I do think that the companies who want to win are reallocating. They're not they're not, you know, doing 50% layoffs. That's crazy.

说实话，对吧，我们合作的客户很多时候都是 Fortune 1000、Fortune 500 这个级别的。嗯，他们想赢，想增长。而我们最常看到的是资源的重新配置，不是搞大规模裁员，对吧？他们宁可让财务部门最好的那批人别把时间花在手工分派发票、审批、解析发票上。那太荒唐了。他们更愿意把这些人放到杠杆更高的事情上，对吧？比如做 FP&A 的规划、其中的某些部分，嗯，或者把他们调到跨职能的岗位上，或者组建他们自己的 FDE 团队。嗯，所以我希望没有人会以为这些东西要把所有人的工作都替代掉。没错，岗位会发生迁移。嗯，但我确实认为，想赢的公司是在重新配置人手，而不是，你懂的，裁掉 50% 的人。那太疯狂了。

[32:54] Greg

Cool.

好。

[32:54] Vas

Another example, 60 person accounting firm. This is more of the SMB side, right? And this is for, you know, if this is your first FDU project, you should probably start here. Um this was uh not actually one of our clients but someone else in the industry that I had chatted with. Um they had $12 million in revenue, 400 clients, four systems. Uh and what made it hard is you know every client sends it books in a different way. I I actually heard that some invoices were sent as a picture

再看一个例子：一家 60 人的会计师事务所。这更偏 SMB 这一侧，对吧？而且如果这是你的第一个 FDE 项目，你大概就该从这里起步。嗯，这家其实不是我们的客户，是我跟业内另一个人聊天时听来的。嗯，他们营收 1200 万美元，400 个客户，四套系统。难就难在，每个客户交账的方式都不一样。我甚至听说，有些发票发过来的是一张照片，

[33:24] Vas

of someone scribbling in a notebook. And that was when I knew like, okay, you can't just apply AI. You got to really get in there and do it deeply. I won't beat a dead horse here because, you know, one, this link will be in the description, and two, it's more of the same. But reality is, they said they had six steps. The reality is they had 14. They have a lot of loops. It's on you to go in and figure this out. Previously, we talked about it in a sales perspective. Now, we're talking about it in a finance perspective. So, collections, then you reask 70% of the time that happens. Then a partner sending it back later. You see step four here. that happens 35% of the time. To be very clear, if you're an FTE, your job is yes to map this out, but also to educate them on what is the cost of this happening. So if if if 35% of the time you have to do this loop, what does that cost? Not just in terms of money, but in terms of time. What is the cycle time of this one person on the team to the next person on the team? That's where a lot of the time goes. Um in the article that I put out uh Michael Hammer who did this study of you know digital transformation said that you might have 20 steps and if you speed up each step you might not make the process any faster because it's the cycle time between steps that makes all the difference. That's where the 20 days of of time comes out to be and that's seen time and time again. That was 30 years ago that he said this. So we're seeing the same thing today again same five the same four buckets sorry uh deletion plain code three agents two human decisions it's on you to figure this out if you want to know how to go about this it's very simple plain code is to be used when it's a simple if x then y there's no judgment there's no variance and if there is variance it's a switch case right if x then y if z then a whatever I random letters you get the idea um on the agent side it's where you

拍的是有人在笔记本上随手写下的字。那一刻我就明白了：行，这不是把 AI 往上一套就完事的，你得真正钻进去、做深。这里我就不翻来覆去讲了，一来这个链接会放在简介里，二来道理都差不多。但现实是：他们说自己是六步，实际上是 14 步。里面有一大堆回路。得靠你自己进去把这些摸清楚。前面我们是从销售的角度讲的，现在是从财务的角度讲。先是收集资料（collections），然后要回头再问一遍，这种情况 70% 的时候都会出现。接着是合伙人过后又把东西打回来，你看这里的第四步，这种情况 35% 的时候会出现。说清楚一点：如果你是 FDE，你的工作没错是要把这些测绘出来，但还要让他们明白，这种事发生一次的代价是什么。如果有 35% 的时候得走一遍这个回路，那代价是多少？不光是钱，还有时间。从团队里这一个人到下一个人，中间的周期时间是多少？大量的时间就耗在这儿。嗯，在我发的那篇文章里提到，研究过数字化转型的 Michael Hammer 说过：你可能有 20 个步骤，就算把每一步都提速，整个流程也未必会变快，因为真正起决定作用的是步骤与步骤之间的周期时间。那 20 天的时间就是这么来的，而且这一点被一再印证。他说这话是在 30 年前。所以我们今天看到的还是同一回事。还是那五个——抱歉，是那四个桶：删除、纯代码、三个 agent、两个人工决策。得靠你自己把这些想清楚。要是想知道具体怎么分，其实很简单：纯代码用在简单的「如果 x 就 y」上，不需要判断，没有变数；就算有变数，也就是个 switch case，对吧？如果 x 就 y，如果 z 就 a，随便什么，字母是我随口说的，你懂意思就行。嗯，至于 agent 这一边，它适用于你

[35:19] Vas

have enough historical data and judgment is required that you can be pretty concrete about hey for example we have an invoice a line item shows monitors that's very likely going to be office supplies but there are exceptions if it's from a certain vendor then we know that it's actually not office supplies it's you know some other thing I don't know you get the idea uh and then finally human decisions right if we have to send out payment there should probably be a human on that we don't want an agent to go end to end because then You have fishing scams, right? You have an invoice that comes in, agent says, "This looks legit. We're going to go ahead and pay it." Human in the loop is always super helpful for reviewing and for delivering approval, etc.

有足够的历史数据、又需要做判断的场景，而且你能把规则说得相当具体，比如：我们有一张发票，某个明细行写的是显示器，那很可能归到办公用品；但也有例外，如果它来自某个特定的供应商，我们就知道它其实不是办公用品，而是，嗯，别的什么东西，我也说不好，你懂意思就行。嗯，最后是人工决策，对吧？如果要往外付款，那上面多半应该有个人盯着。我们不想让 agent 从头跑到尾，因为那样你就会碰上钓鱼诈骗，对吧？来了一张发票，agent 说：「这看着挺正规的，那我们就直接付了吧。」让人留在回路里（human in the loop），对复核、对给出审批这类事永远都特别有帮助。

[36:01] Greg

How should people think about frontier models versus open- source models, Chinese models versus America models? like in in like we've just talked about agents as agents but if you're actually going to deploy these agents how how should people think about these ecosystems?

大家该怎么看前沿模型（frontier model）和开源模型、中国模型和美国模型？就是说，我们刚才一直是把 agent 笼统地当 agent 来聊，但如果真要把这些 agent 部署下去，大家该怎么看这几套生态？

[36:22] Vas

Super good question. So on one hand most companies are on co-pilot, Microsoft copilot. Now behind them is cloud code. Behind that is codecs. Um and a lot of what you want should start especially for an SMB in skill files wherever you are. Now, if you're actually building agents, which requires engineering expertise, I'll be honest, and the big labs don't want you to know this, but most of what you're looking to achieve does not need to be leveraging a frontier map. There's very few cases where we've seen the need to deploy Fable or Astra. Um, now that doesn't mean that we're not using their models. We're using, you know, Opus, you know, 4.8 aid or sonnet more likely or you know GPT you know with lower thinking model I don't even know what their naming convention is anymore um and also open source now the other part is if you're talking to enterprise uh and this is some sauce for for the viewers they have an aversion to Chinese models even though they're floating point numbers not actually how that works uh they don't want to work with models that are out of China they can't or you know there's a there's a strict aversion to it but we have leveraged open source models like you know Muse um and sometimes Chinese models where we're allowed to uh grock has been a great model that we've used as well. Um so don't feel the need to silo yourself into just chatgbt claude frontier models. You can use their non-frontier models. You can use gro. You can use muse. You can use GLM, Kimmy, Quinn. All these different things are are toolkits and and you should benchmark every single workflow against every single model to determine what model is the right use is the right model for your use case.

特别好的问题。一方面，大多数公司用的是 Copilot，Microsoft Copilot。排在它后面的是 Claude Code，再往后是 Codex。嗯，而且你想做的很多东西，尤其对 SMB 来说，都应该从 skill 文件做起，不管你用的是哪一家。不过，如果你是真的在构建 agent——老实说，这需要工程方面的专业能力——那大实验室不想让你知道的是：你想实现的大部分东西，并不需要用到前沿模型。我们见过的真正需要部署 Fable 或 Astra 的情况非常少。嗯，这并不等于我们不用他们家的模型。我们用的是，你知道，Opus 4.8，或者更多时候是 Sonnet，或者 GPT 里思考档位调低一些的模型——我都搞不清他们现在的命名规则了——嗯，另外还有开源模型。还有一点，如果你面对的是大企业，嗯，这算是给观众的一点独家秘方：他们对中国模型很抵触。尽管模型不过是一堆浮点数，实际情况并不是他们以为的那样，嗯，但他们就是不想用出自中国的模型，要么是不能用，要么，你知道，就是有一种很硬的抵触。不过我们用过开源模型，比如 Muse，嗯，在获准的地方有时也用中国模型。嗯，Grok 也是我们用过的一个很不错的模型。嗯，所以别觉得非得把自己圈死在 ChatGPT、Claude 这些前沿模型里。你可以用他们的非前沿模型，可以用 Grok，可以用 Muse，可以用 GLM、Kimi、Qwen。这些全都是工具箱里的工具，你应该拿每一个工作流去对每一个模型做基准测试，来确定哪个模型才是适合你这个用例的模型。

[38:16] Greg

We've seen personal agent platforms start to get big. So we have Grockbot now. We have Muse. Muse actually has Muse for small business. They just announced that. Um, Instinct, which is more on the consumer side. Um, you got to think the other big players are going to come into that space too. Um, how are you thinking about using the Grock bots of the world to to, you know, deploy into these enterprises? Are you thinking about it?

我们看到个人 agent 平台开始做大了。现在有 Grok Bot，有 Muse。Muse 其实还有面向小企业的 Muse for small business，他们刚刚宣布的。嗯，还有 Instinct，它更偏消费者那一侧。嗯，可以想见其他几家大玩家也会进到这个领域来。嗯，你是怎么考虑用 Grok Bot 这一类东西，去，你知道，部署进这些企业的？你有在考虑这件事吗？

[38:51] Vas

Yeah. and then open AI yesterday with dots, right? Um I think there's an unlock for that, but it's hard to see governance for those, you know, agents, personal agents in the enterprise use case, right? Um I think that is an extension. I think my my philosophy and the philosophy that we follow here at Veric is that there's two streams of agents for for any business or whatever. There's the sidekick agent, which is your co-pilot. You chat with it, you get stuff done. And Instinct and Dots and Grobbot and Muse are extensions of that where you have to chat with it and they get stuff done. The other angle is background agents that truly do work in the background. They don't bother you. They just do the same thing and they ping you when they need to. They know what they have to do already. And that's why this deep dive process mapping process engineering is so valuable. Now I can see them connecting at some point where you can use a muse or a Grockbot to set up these background agents that just take work for you all the time, but that governance isn't there yet. Uh there's still a massive gap in how much involved you have to be and how involved you have to be from a software engineering perspective. Um, so so far the use cases to answer your question are limited and we would rather take work off of their plate rather than make them move faster because that's the difference in ROI.

有。还有 OpenAI 昨天发的 Dots，对吧？嗯，我觉得这里面是有东西可以被打开的，但在企业场景里，很难看到针对这类 agent、这类个人 agent 的治理机制，对吧？嗯，我认为那是一种延伸。我的理念，也是我们在 Varick 遵循的理念是：对任何一家企业来说，agent 都有两条线。一条是 sidekick agent，也就是你的 copilot（副驾）：你跟它聊，把事办了。Instinct、Dots、Grok Bot、Muse 都是这条线的延伸——你得跟它聊，它才把事办了。另一个方向是 background agent（后台 agent），它们是真的在后台干活。它们不打扰你，就一直做同样的事，需要的时候才来叫你一声。它们本来就知道自己该做什么。这也正是这种深入的流程测绘、流程工程为什么这么有价值。我能想象这两条线在某个时候接上——你可以用一个 Muse 或者一个 Grok Bot 来搭建这些后台 agent，让它们一直替你把活儿接走，但那套治理现在还没到位。嗯，在你需要介入多深、需要从软件工程的角度介入多深这件事上，仍然有一道巨大的鸿沟。嗯，所以到目前为止，回答你的问题：用例是有限的，而且我们宁可把工作从他们手上拿走，而不是让他们干得更快，因为 ROI 的差别就在这里。

[40:19] Vas

This gives them 10 20% faster output. This gives them 70 80% faster output with higher accuracy etc.

这个能让他们的产出快 10%、20%；这个能让产出快 70%、80%，准确率还更高，等等。

[40:27] Greg

Mhm. Cool.

嗯，不错。

[40:30] Vas

So again selling the idea of an FTE you need to be three people in one. uh one you need to understand how the work actually gets done. Uh ideally that means you go off on your own and you really study these systems of record. You study salesforce net dynamics as we talked about earlier. What do they offer? What do they not? And then finally and then and then part of that is you know how should accounts pay table function and you learn that either on your own in combination with going into a company. Uh the second is shipping production code. You have to be able to do engineering work. Now doesn't mean you need to be a undergrad in computer science and you know software engineer for 10 years especially with the advances in AI engineering. Uh but you do need to be able to ship production code agents that call into these different systems of record and do so with auditability, governance, um security in place. And then third is the AI layer, right? Knowing what model to use, knowing what you can trust a model with versus what you can't, knowing how to test each model through eval, you know, optimizing your harness. And then finally, you know, how to handle agents taking incorrect actions, right? Roll backs on agents hallucinating, etc. And if you can do all three of those, you are the best FTE. You're a very cabley. And this is actually very very rare. Usually you just have one or two of these. Um or you're even kind of mediocre at all three. You have to be exceptional at all three. Plus the communication of it all.

所以再给大家推销一下 FDE 这个概念：你得一个人顶三个人。第一，你得懂活儿实际上是怎么干成的。理想情况下，这意味着你自己下功夫，把这些记录系统真正研究透——就像我们前面聊过的 Salesforce、NetSuite、Dynamics，它们提供什么、不提供什么。这里面还有一块是：应付账款这个职能到底应该怎么运转——这个你要么自学，要么结合着进到一家公司里去学。第二是交付生产级代码。你得能干工程的活儿。这不是说你非得是计算机科学本科出身、当过十年软件工程师，尤其是 AI 工程已经进步到今天这个程度。但你确实得能把生产级的代码上线——做出能调用这些不同记录系统的 agent，而且可审计性、治理、安全都得到位。第三是 AI 这一层，对吧？知道该用哪个模型，知道哪些事可以放心交给模型、哪些不行，知道怎么用 eval 去测每个模型，怎么优化你的 harness。最后还有，agent 做了错误的动作该怎么处理，对吧？agent 出现幻觉时怎么回滚，等等。这三样你要是都能做到，你就是最顶尖的 FDE，你就是非常有能力的那种人。而这其实非常非常少见。通常你只占其中一两样，或者干脆三样都只是平平。可你得三样都出类拔萃。再加上把这一切讲清楚的沟通能力。

[42:09]

Right? Being able to speak to senior leadership uh and and convince them this is the right way to go. And if you are this person, we really need to talk. We want to hire you. Um and that being said, you'll also have ample opportunity everywhere else, right? Everyone's looking for top FTEES. Totally [gasps] 100%. Yeah. I mean, she I like that you shot your shot there, you know, respect. Um I think uh this person is like your NBA player, right? It it's it's it is top 0 1%. But if you can figure this out and the cool thing is you can figure this out in like you you know like you said like you don't need to be a have a CS degree. You just have to dedicate yourself to learning the craft. You you need experience deploying the craft and you also need a lot of uh reps around just all the different ecosystems open source versus closed source like a lot of the different um I mean even like the the Microsoft ecosystem and the Salesforce ecosystem you have to understand all these words bring it together communicate it in a way that you know sells to execs so it is hard, but like that's why these people get paid what they get paid. You know what I mean? And that's why the value is so huge, right?

Vas： 对吧？能跟高层管理者对话，说服他们这条路是对的。如果你就是这样的人，我们真得聊聊，我们想招你。话说回来，你在别的地方也会有大把机会，对吧？人人都在找顶尖的 FDE。

Greg： 没错，100%。对。我喜欢你刚才抓住机会直接开口招人，佩服。我觉得这种人就相当于 NBA 球员，对吧？就是最顶尖的 0.1%。但如果你能把这个搞明白——而且妙就妙在你是能搞明白的，就像你说的，你不需要有 CS 学位。你只要一门心思去学这门手艺，你需要实际把这门手艺部署出去的经验，还需要在各种不同的生态里反复练手——开源对闭源，各种各样的……甚至像 Microsoft 的生态、Salesforce 的生态。你得听得懂所有这些词，把它们融会贯通，再用一种能打动高管的方式讲出来。所以这确实难，但这也正是这些人能拿到那个薪水的原因，你懂我意思吧？这也是为什么这里面的价值这么大，对吧？

[43:38] Greg

Like the problems that you're solving with deploying FTEES like I mean, as we've seen in this episode, like is multi-million dollars of savings and efficiency per year easily. So, like someone once gave me advice, well-known well-known person, well-known founder, uh several multi-billion dollar exits when I was, you know, young. And he he he he always said like, you know, if you're if you're finding a job, the best job to find is the one closest to the money. um the one that could show that you can optimize that you um you know revenue profit um because the people that do that are the ones that are naturally going to get paid the most because there it generating as much value for that as we talked about earlier like that factory system. Um, so it's like, yeah, you you know, you you gave $10 million of value to this company. Do I could I pay you 10 10% of that? A million dollars? Like maybe, you know, like that might be a trade. And that's why I think FDEEs are so in demand one, and two, getting paid so well.

你派 FDE 进去解决的那些问题——就像我们这期里看到的——轻轻松松就是每年几百万美元的节省和效率提升。我年轻的时候，有人给过我一个建议，是个很有名的人、很有名的创始人，有过好几次数十亿美元级别的退出。他总是说，你要是在找工作，最好的工作就是离钱最近的那个——就是能证明你可以优化收入、利润的那个。因为做这种事的人自然就是拿钱最多的人，因为他们在给那套东西创造那么大的价值，就像我们前面聊过的那个工厂系统。所以就变成：行，你给这家公司带来了 1000 万美元的价值，那我能不能把其中的 10% 付给你？100 万美元？也许可以吧，你懂的，这可能是笔做得成的交易。所以我觉得，这就是为什么 FDE 第一这么抢手，第二薪水这么高。

[44:55] Vas

Yeah. In a heartbeat, you would pay them 10% of what they can deliver for you. And there's even PE firms who are hiring FTEES and giving them a percent ownership in the carry where you know they bought it for a billion and they hope to sell it for 5 billion and you'll get you know.5% of whatever that delta is

对。他们能给你交付多少，你把其中的 10% 付给他们，眼都不会眨一下。甚至有 PE 公司在招 FDE，直接给他们 carry（业绩分成）里一定比例的份额——比如他们花 10 亿买下来，希望卖到 50 亿，那这中间的差额不管是多少，你能拿到其中的 0.5%，

[45:14] Vas

based on the work that you're able to do for them.

依据就是你能为他们干成的那些活儿。

[45:17] Vas

Um like to your point it's it's the NBA players and like you have to be good you have to be great. You have to be the best of the best.

就像你说的，这就是 NBA 球员。你得优秀，你得卓越，你得是顶尖里的顶尖。

[45:25] Vas

Yeah. So action items, if you want to be, you know, deeper in the FTE space, if you want to kind of get started, maybe dip your toes in, if you're starting from nothing, this is what I would do on a personal level. List every single app that holds your stuff, right? I I talked about mine earlier today, earlier today earlier in the presentation, which was, you know, five different inboxes, three different texting communication channels, etc. Do this for yourself. um write which one wins, which one disagrees, you know, how to route this kind of do a whole process mapping of your own life. And the next step, take 20 things you did last week, right? You paid a bill, you cancelled a subscription. This is a big one. I'm seeing a lot of use cases on Instinct and Muse where they go in and cancel their, you know, stuff they forgot about. I got to do that with Adobe, by the way. Adobe, if you're listening, charge me 40 bucks a month for two years. Um on Wednesday you write one process down step by step. You know how you pay a bill end to end or how you submit an invoice if you're a freelancer for work. You know map this out and try to get as detailed as possible. So you know there's five different you know ways of doing it. There's 10 different exceptions etc. And Thursday sort every single step like we talked about the four buckets. What gets deleted? What's deterministic? What's agentic? What do you still need to be there for? And then Friday, figure out who you want to reach out to and reach out to a bunch of SMBs that you can either get connected with or you can do cold outbound to to do this for them and offer this in a single process.

好，那说说行动清单。如果你想在 FDE 这个领域走得更深，如果你想起个步、先试试水，如果你是从零开始，下面是我个人会怎么做。把存着你东西的每一个 app 都列出来，对吧？我今天前面——就是这场演示前面——讲过我自己的情况，五个不同的收件箱、三个不同的短信沟通渠道，等等。你也给自己做一遍。写下来哪个说了算、哪个会对不上、该怎么路由，相当于给你自己的生活做一整套流程测绘。下一步，拿出你上周做过的 20 件事，对吧？你付了一笔账单，你取消了一个订阅——这个是大头。我在 Instinct 和 Muse 上看到很多这样的用例：人们让它进去把那些自己早就忘了的东西取消掉。顺便说一句，我也得这么对付 Adobe。Adobe，你要是在听——你们每月扣我 40 美元，扣了两年。到了周三，你把一个流程一步一步写下来。比如你付一笔账单从头到尾是怎么走的，或者如果你是自由职业者，你是怎么为工作提交发票的。把它画出来，尽量写细。比如有五种不同的做法，有十种不同的例外情况，等等。周四，把每一个步骤归类，就像我们讲过的那四个桶：哪些直接删掉？哪些是确定性的？哪些是 agentic 的？哪些还得你本人在场？然后周五，想清楚你要联系谁，去联系一批 SMB——要么找人引荐，要么直接做冷外联（cold outbound）——帮他们做这件事，而且只就一个流程来提供这项服务。

[47:09]

Start with one workflow. Make it super simple and do everything for them. do the skill files, do the uh, you know, personal assistant, instinct, muse, crockbot, dots, whatever it is, and build the agents and give them a timeline. Do it for free if you have to, if you're getting started. Trust me, the experience is worth more and your next one you can charge that five figures, that six figures. But if you haven't done this ever, get some experience in. Um, and then you you go from there and take what you learned from this from this presentation. How important is it is it to know how to deploy hardware with agents at these enterprises? Is that something a lot of people are asking for

Vas： 从一个工作流开始。弄得超级简单，然后替他们把所有事都做了：skill 文件做好，个人助理也搭好——Instinct、Muse、Grok Bot、Dots，不管是哪个——把 agent 建起来，再给他们一个时间表。刚起步的话，必要时免费做都行。相信我，这份经验更值钱，下一单你就可以收那个五位数、六位数了。但如果你从来没做过，先去攒点经验。然后从那儿往下走，把你从这场演示里学到的东西用上。

Greg： 在这些大企业里，懂得怎么把硬件和 agent 一起部署有多重要？是不是很多人都在提这个需求？

[47:52]

hardware in terms of like GPUs? Yeah, like uh you know they have sensitive data and so they you know they want to you know they want open source models on you know at on premise basically versus versus cloud agents.

Vas： 你说的硬件，是指 GPU 这类吗？

Greg： 对，就是，他们有敏感数据，所以他们想要开源模型跑在——基本上就是本地部署（on-premise），而不是云端的 agent。

[48:11] Vas

Yeah, truthfully we've seen zero of that. Um I know there are companies who do that where they kind of like either rent or sell GPU clusters to very large maybe heavily regulated companies. We've worked with some of the most heavily regulated companies on the planet. Um, like banks, financial services, healthcare, pharma. Um, and I don't think that they're there yet. Maybe eventually they might be, but not right now.

嗯，说实话，这种需求我们一例都没碰到过。我知道有些公司是做这个的，把 GPU 集群租给或者卖给那些体量很大、可能受到严格监管的公司。我们合作过的客户里，有一些是这个星球上监管最严的公司，像银行、金融服务、医疗、制药。我觉得他们还没走到那一步。也许最终会，但现在还没有。

[48:37] Greg

Did you see OpenAI launched some security features yesterday at dev day or the other?

你看到 OpenAI 昨天——还是前两天——在 DevDay 上发布的一些安全功能了吗？

[48:43] Vas

No, I missed it. What was that?

没有，我错过了。是什么？

[48:46] Greg

They launched pull it up so I don't butcher it. So they launched private intelligence. Open AI private intelligence helps businesses use frontier AI with greater confidence that their data is protected. So there's a zero data retention with private private safety processing which enables automatic automated safety reviews. Basically you don't have to give open AI uh p personnel access to the underlying content. I think a lot of people were kind of like I want to use some of these models, you know, but I mean when I say people, I mean businesses um businesses are like I want to use some of these models but do I will really want to give the keys to OpenAI? Like maybe not. So they end up launching a feature like that.

他们发布了——我调出来看看，免得我说错。他们发布了 Private Intelligence。OpenAI Private Intelligence 帮助企业更有把握地使用前沿 AI，确信自己的数据是受保护的。也就是零数据留存（zero data retention），加上私密安全处理（private safety processing），这样就能做自动化的安全审查。基本上就是，你不用让 OpenAI 的人员接触到底层内容。我觉得很多人之前的心态是，我想用这些模型，但是——我说的「人」其实是指企业——企业的想法是，我想用这些模型，可我真的愿意把钥匙交给 OpenAI 吗？可能不太愿意。所以他们最后就推出了这么一个功能。

[49:35] Vas

Oh, very cool.

噢，挺酷的。

[49:36] Vas

Yeah. I mean we we route through like Azure Foundry, you know, AWS Bedrock, Google Vertex, and they have, you know, again, they've agreed to not train ZDR, etc. Um, but I can see how this becoming more and more of an issue for a lot of companies. It makes sense why they launched it.

对。我们是通过 Azure Foundry、AWS Bedrock、Google Vertex 这些来路由的，他们也都同意了不拿数据去训练、ZDR（零数据留存）等等。不过我能想见，这对很多公司会越来越成问题。他们推出这个是说得通的。

[49:53] Greg

Cool. And yeah, I like this step-by-step uh process. This is valuable for really not just for a lot of people. Number one, if you want to be an FTE, this is valuable. Number two, uh if you want your company to be more AI native, it's valuable except Friday stuff, like you're not reaching out to people, but Monday to Thursday, you're just like

不错。对了，我喜欢这个一步一步来的流程。这真的对很多人都有价值。第一，你想成为 FDE，这有价值。第二，你想让自己的公司更 AI 原生，它也有价值——除了周五那部分，因为你不用去联系别人；但周一到周四，你做的就是

[50:18] Greg

understanding the system and optimizing the system.

理解这套系统，再优化这套系统。

[50:21] Greg

So, it's it's valuable for a lot of different people.

所以它对很多不同的人都有价值。

[50:24] Vas

Yeah. And one of the questions we get asked is like, how do we make our own FDES internally? And I'd have them follow the same playbook, right? Um so again this is if I if I leave the viewers with nothing else this is this is what it should be. Don't apply AI right

对。我们常被问到的一个问题是：我们怎么在内部培养自己的 FDE？我会让他们照着同一套打法来，对吧？所以，还是那句话，如果我只能给观众留下一样东西，那就应该是这个：别「应用」AI（Don't apply AI），对吧，

[50:43] Vas

too many times and this is the reason why most AI pilots fail is is they try to slap AI on um on on the business and it's it's very handwavy and you know you maybe it's rolling out a license with claw code to everyone maybe it's you know building an agent that doesn't really understand the workflow. it doesn't work. Um when we go into these large companies, this is the exact process that we follow and this is why we successfully have transformed departments with AI. We do these steps. We find the real process. We measure the time baseline all of the KPIs. We pick processes with owners. We sort every single step. We baseline before we build. And then we build it once and we deploy it everywhere. And we measure it constantly. So we'll go to these CFOs over the course of four weeks do an audit um where we you know we understand their systems we build the PC's and all in four weeks the next four weeks we build the agents and then 3 months after that 6 months after that we say hey this is what it used to be this is what it is now and we prove it to them so it's not just oh we built AI we deployed it we're done that's the job of an FD doing everything end to end um so hopefully that was helpful for for all people watching and I'm wishing everyone best of luck.

太多时候——这也是大多数 AI 试点失败的原因——他们就是想把 AI 往业务上硬贴，做得非常虚、非常想当然。可能是给所有人发一份 Claude Code 的许可证，可能是建一个其实并不理解工作流的 agent。这行不通。我们进到这些大公司的时候，走的就是这一套流程，这也是我们能成功用 AI 改造一个个部门的原因。我们就做这几步：找到真实的流程；测出时间基线和所有的 KPI；挑有负责人的流程；把每一个步骤归类；先定基线再动手建；然后建一次、到处部署；并且持续不断地度量。所以我们会去找这些 CFO，花四周时间做一次审计——摸清他们的系统、把 POC 做出来，全在四周之内。接下来四周，我们把 agent 建出来。再过 3 个月、6 个月，我们会说：嘿，以前是这样，现在是这样——我们证明给他们看。所以这不是「噢，我们做了 AI、部署完了、收工」那么简单。这就是 FDE 的工作：端到端把所有事都做了。希望这对所有在看的人有帮助，祝大家好运。

[52:01] Greg

We'll include the link to this in the show notes in the description so people can access it. Uh sometimes people ask me actually they don't even ask me. They go in the comment section. They're like you're involved in this company. No, I'm not involved in this company. You know, like Vos Voss hasn't bought me a beer. He hasn't sent me money. He hasn't given me a coffee. Nothing. I think that he's just really smart when it comes to this stuff. And I think that if you understand this stuff, you have an unfair advantage. And that's why I bring him on here. I'm I'm on here because he he's world class when it comes to this stuff. He's not afraid of sharing the sauce, and that's why he's here.

我们会把这份东西的链接放在节目备注、也就是简介里，方便大家查看。有时候有人会问我——其实他们连问都不问，直接跑到评论区说：你跟这家公司有利益关系吧。没有，我跟这家公司没有任何关系。Vas 没请我喝过一杯啤酒，没给我打过钱，连杯咖啡都没请过，什么都没有。我只是觉得他在这方面真的很聪明。而且我觉得，你要是把这些东西弄懂了，你就有了不公平的优势。这就是我请他来的原因。我请他上节目，是因为他在这方面是世界级的，而且他不怕把看家的秘方分享出来，所以他才在这儿。

[52:39] Vas

I appreciate that. Well, now I feel bad now. I do owe you a beard. Uh but uh yeah, I mean, look, if I get one thing out of this, it's it's I need to hire people. So, if people can apply, that's one thing that, you know, maybe I can I can send grand kickback for. But yeah, we, you know, just really appreciate you having me on.

谢谢你这么说。好吧，现在我倒过意不去了，我确实欠你一杯啤酒。不过，说真的，如果这次我能有什么收获，那就是——我需要招人。所以如果大家能来应聘，这倒是一件，嗯，也许我可以给 Greg 一点回扣的事。总之，真的很感谢你请我来。

[52:58] Greg

No kickback needed at all. Um, I uh I like what you're doing. And it's funny you're saying like don't apply AI, but apply to my company. So that's it's hilarious. [laughter] Uh, and uh, yeah, I wish more people I hope if you've made it this far that you go, no matter if you want to be an FTE or you want to apply these this methodology, like go and do it, get your hands dirty. Uh, if people want me to go deeper uh, on these topics, please let me know in the comment section. I read every single comment. I respond to most. Voss, you're a legend for coming on, sharing sharing the sauce. I appreciate you. Um, please come back again. I'll include links where you can follow uh Voss on the internet and his apply AI article that I saw and I reached out to him and I was like, "Hey, you got to come back. Come come come back on the pod." Um, and uh, I'll see you next time, my friend.

回扣完全不需要。我喜欢你在做的事。而且有意思的是，你嘴上说「别 apply AI」（别『应用』AI），转头又说「来 apply 我的公司」（来应聘），太好笑了。（笑）嗯，我希望更多人——如果你看到了这里，我希望你不管是想当 FDE，还是想把这套方法论用起来，都真的去做，亲自下场动手。如果大家想让我在这些话题上讲得更深，请在评论区告诉我。每一条评论我都会看，大部分我都会回。Vas，你能来、还把秘方分享出来，真是个传奇，谢谢你。请一定再来。我会附上链接，大家可以在网上关注 Vas，还有他那篇讲 apply AI 的文章——我就是看到了那篇文章才去联系他，跟他说：「嘿，你得再来一次，再回播客来一趟。」那我们下次见，我的朋友。

[53:57] Vas

Cheers. Thanks so much for having me.

好嘞。非常感谢你请我来。