S2E4 | Agentic Systems & Why Outbound Isn't Dead with Laurens Nys & Spencer Tahil

Hey guys, today I have somebody that's probably more technical than me and probably you've probably seen his name on LinkedIn. His name is Laurens. I'm going to let him introduce himself, but welcome to the show. And we're going to talk about probably the state of technology and go-to-market today, I would imagine, but we'll see where the talk goes. But Laurens, welcome to the show and yeah, let's kick off.

Yeah, man. Thanks for having me. Been a while since I've done an intro. So I guess the short intro is I've been a GTM engineer. It's funny too, like I've I own the domains for GTMEngineer.io as well, because I was used to have recruitment agency, then started doing GTM engineering, but before Clay, like, announced that they were, like, claiming the term GTM engineer. So it's kind of early on. That's why I, you know, was able to buy those domains still. But yeah, anyway, so been doing this for 3 years now, I think, which one of it, one year of it is like full-time building out a GTM engineering company. It's been a fun ride so far, mainly working with SaaS companies, only B2B, I mean, only B2B, mainly focused on like enterprise client stuff. Yeah, I mean, I guess that's a short intro, a bit of a rambled one, but.

Well, I just talked to Frank and, uh, yeah, he, he loved to talk about like the application and everything like this. So I, I think where I want to continue our conversation, I know we're going back and forth on LinkedIn about everything, is kind of the state of technology right now. I think you're one of the people that I look up to in the industry, actually, for a long time before I got into this like go-to-market engineering, you know, terminology, whatever it was. I was like, ah, this guy knows some shit, you know? And I was like, all right, let's dive into it. So like, you know, and I think this is bringing back to like people used Apollo in 2018, nothing against them. People still use ZoomInfo and all of these different tools. And I think in 2024, it was when Clay came out. And then in 2025, last year, we were looking at like MCPs. And now we're in like Q1 '26 and now we've got Claude Code. I would love, because you're close to the action, I'd love you to break down kind of like where you're pivoting and we could just chat about that. I feel like that's super interesting.

Yeah. Um, I mean, it's interesting because it's one, there's so much and two, it's going so fast. Like it's going, like it was going fast last year, but it seems like it's, it's only going faster. Like it's not really slowing down, but yeah, I mean, so we are now pivoting to just writing out code and then the exec we are mainly using is Claude. Agent SDK. Cloud Code is like, build the actual stuff and then kind of build the MVPs in Cloud Code because the Agent SDK is basically just a wrapper around it to build workflows instead of you having to interact with it. And I'm mainly using Prefect. Like, some are using Trigger, but my Python is better, so I like Prefect better, but it's basically the same thing. It's just a wrapper to like, build out the automations like cron jobs. Like let it respond to webhooks, et cetera. So that's what we're pivoting to right now. Like, and it's very interesting because 6 months ago, maybe 12 months ago, or between those two, we pivoted to NaDan. And back then it was like, okay, NaDan is now the advanced shit. Like for TTM engineers, if you're on NaDan and not on Clay anymore, you're one of the advanced ones. And if you would have asked me, like, I don't know, even 3 months ago, if we were going to pivot away from NaDan, I would be like, no, that's like, it's not necessary. Like NaDan is good enough. And then Anthropic came out, they made Claude codes, they made The agent SDK, which is to me the main difference as well as so an agent is very linear in terms of the flows you can build. There's some like looping, et cetera, you can build in, but linear in the sense of it's predefined what it does. It doesn't really update its own prompts. It doesn't really compound, let's say. The agent SDK, the kind of unlock for me or why I'm like, okay, I need to get on this is that it lives inside. Your knowledge base or inside of your repo, and it can update itself. Like, for example, you, I don't know, Outbound is a common use case is you can have some sort of a nighttime computer or whatever to where you're sending emails all day. And in the evening, the agent kind of goes through who's answering, who's not. From those who are answering, like, what emails are we sending? What type of persona is this? What type of segment is this? And it just updates its context on each of them. And then the next day when it's sending emails, it's updated based on the previous day's results. And it kind of keeps learning, which is isn't possible in 8N or, I mean, it might be in theory, but it will be a pain in the ass to set up.

So got it. So we're moving into the age of like continuous looping architecture in marketing right now.

Yeah. I mean, agentic is like the main term I would use, which was, it's used a lot, but to me, agentic kind of means it's less predetermined how a flow works, but it's more on like, you have an end result you want to achieve and then you give the agent context and you give it the tools. To be able to achieve it and like guardrails, et cetera, to make able, like to make sure it's doing what you want it to do. But it's not like NaIN or Clay where you have predefined, like now do this, now do this, now do this. I think that's, that's for me, like the main switch. And I know that the, the word agentic was used a lot like last year, but I don't feel like that was like really agentic. So like you were using AI.

Do you feel like this word like agentic is just like more is just being like thrown around like crazily because at least on my side of the like planet or on my side of like the coin, it's just like, oh, we want AI, we want this, we want that. And then they're like, oh, what do you use? Zapier. What do you use? Clay. Okay, awesome. But like, have you heard of n8n? Have you heard of Cargo? Like, have you heard of QuadCode? And then you have that bucket. And then on the other side of the bucket, you've like I just built an AI to replace my whole 50-person marketing team. Comment reach for reach. And I'm like, okay, fucking bullshit, man. And I think like there's a completely misunderstanding. I don't know if you get that sentiment as well, because I think a lot of people don't know what's real. And I think this word also with like go-to-market engineers, the line is so blurred that it's hard for me, at least as a single person on LinkedIn, to be able to like educate on what's real.

Yeah, I mean, it's hard because a lot of it is also just marketing from those AI companies. I haven't checked Clay's website recently, but I would bet there's some word like agentic is on there somewhere, maybe on their Create and stuff. But yeah, I mean, it's hard because there's not like, there's not like one company that's doing it and defining terms and it's changing so fast that like, even if there is, might be like this weekend and next week it's another one. Like OpenAI is now, I mean, Anthropic is overtaking them in enterprise clients, which like I don't know, a year ago, like, I don't know, maybe a little more than a year ago when Tropic, like, really didn't exist. Like, they were there and they were kind of good, but OpenAI was like the one and it seems to have switched like a lot recently. So yeah, and then there's also like just lack of understanding in terms of like what AI is, how it works. And I mean, I'm doing posts on Cloud Code a lot right now. Like, I think yesterday or the day before, someone commented on something they were doing with ChatGPT. And I was like, dude, that's not, that's not, that's not Cardcode. One. And two, that's like, that's the chat. It's like, it's not the model. It's not the application here. I feel like a lot of people don't really lack understanding of what AI is.

I try to explain to people, random people in the real world, not on the LinkedIn world, like, what do you do for work? And I'm like, I like, I play with artificial intelligence and I make people money. And that's like, they're like, what? And I was like, ah, you know, like You know, imagine, imagine sending 1,000 emails. They're like, uh-huh. And they're like, and then what else do you do with artificial intelligence? I was like, I kind of train it. And they're like, uh-huh. And then what? And I was like, well, that's pretty much it. It's just pattern recognition. But then I let it make its own decisions. Right. And they're like, does that work? And if they're interested, we talk about it. But like, yeah, I would say like this, I think going in a linear-based flow, which is kind of like what you have behind you where it's like, uh, point A, webhook, API, do something. It's very linear. I think a lot of people that are not super educated are going to just, they're going to call it agentic AI. And then artificial intelligence, most of the time what I'm seeing, and again, I think that I would assume that you work on the same type of projects with the same level of intricacies that I do. A lot of people use AI for very complex formulas and I don't think it's needed. I think if you can write a formula, and a very complex one at that, a lot of the times you don't need an agent to do it. Agent and AI, quite different terminology, terms that I think get confused a lot. In my, in my head, how I define an agent is an agent is like a human without its body. It's just the brain. It's capable of making its own decisions based on rationale that you provided to it beforehand. And if it says, hey, arm 1, go move up and down, or agent number, sub-agent number 1 is going to move their arm up and down. but now the complexity is like, oh, move the fingers and like write the email, use the context, whatever. That's how I would describe agents. But if you're like looking at just like AI-powered workflows, then it's just a linear, you know, line, like a linear workflow. I think there's a lot of miscommunication in the industry around this, especially with marketing now, like with, with PlayChat and everything.

Yeah. Yeah. Yeah, I agree. I mean, there's different levels. Like one, you have the chat interface, a lot of companies are using as well. And then it's like the, I don't know, I don't want to put the custom GPTs in there, but you do see it a lot because I get comments on my post like once in a while and somebody's like, oh yeah, I built this custom GPT and I'm always like, yeah, you wrote a prompt. That's what you did. I don't know if it's like updated or whatever now to where maybe you can do more, but initially it was just, you just iterate on a prompt and then you have some sort of system prompt and okay, that's not really an agent or anything, but I feel like probably 90% of people are still stuck there on just like Yeah, I use AI for marketing and it just means they open ChatGPT, they text with it and then they have a post and then they complain about like, like AI isn't that advanced. It's like, well, yeah, like it's the application of it.

I think, I think the reason why Claude, and I'm a very big user of Claude and I know that you are too, but I don't know if you'll believe me if I say this, but I don't use Claude Code.

Why not?

I don't, I don't know. I just never got into it and maybe I'm missing out and everybody's going to tell me you're missing out, you're missing out, but I I don't need it because I'm going to design the models and then I'm going to port them into the systems that support them anyways, right? I use my Claude account to be able to draft me, you know, the architecture, run me through the workflow. I have the.md files as well, but I don't feel like I'm missing a lot. A lot of people think that I do everything with Claude Code. No, I do everything by hand because everything slows me down. Of course, I could become way more efficient. I just I don't know. I just never got into it. I talk to it. I don't even use Whisper Flow like a lot of people do.

I use Google.

I literally use Google Speech to Type, or I open up my phone on Claude and I speak to it. I give it my idea and I was like, hey, use APO, automatic prompt optimization, help me build a better prompt, plug that prompt in. Hey, here's the, you know, high-level context, decompose the context so everything that you build contextually makes sense. Chain it all together. Okay, give me each element by itself and then I port them in. I mean, it all comes down, in my opinion, it all comes down to architecture and then individual system prompts, individual user prompts, and then chaining. And you just have to reach the desired result. I don't need Claude code to build that for me. I mean, it'd be great, but I'm not gonna port it to N8N. Probably gonna build that somewhere else. I mean, N8N is great, but I don't use it.

I mean, to me, the main unlock it gives, and we're using Cloud Codeur, it's, it's, I mean, the agent SDK is based on it. So like agent SDK is same thing, but to me, the main unlock is most AI sort of systems or agents are, you have this AI that is able to use tools and it just has a directory of tools and you can connect more tools with different tools to it, et cetera. And it can like kind of go back and forth between whatever prompt you gave it to doing something with tool, tool gives it back and it kind of iterates until it's like, I don't know, does task. With the agentless scale with Cloud Code, it kind of lives inside of a codebase. That's what it's built for. You can like throw away the codebase and just give it Markdown files. So the way I'm using it, sort of like we're using it for clients as well, is we basically build out a knowledge base in Markdown files. And there's like multiple ways to think about it. Like one is you have the sort of compounding knowledge, like every call they have, Agent runs on it, kind of extracts whatever is important from that call, as in this was a call with segment X, here's a file on segment X, we update the file on segment X. This persona was in the call, here's a file on that persona, we update file with persona, and you just automate that using the agent SDK. And since Cloud Coach kind of lives inside of it, you can build workflows on top of it. But then, and workflows are more agentic, as in how we are building workflows is basically called skills. But then via the agent SDK, so you can like just, because skills and commands, they merged since a few days ago, so you can just invoke them and then it just goes off and does it. But since it lives inside of the codebase, it can then like reference all those files. You can just like, for example, you tell it to write an email to this persona and whatever, it will get the file on persona, get the file on the segment, try to email it, kind of go on. And then every night when you then updates files based on new calls, it updates that file. And then the next time the workflow runs, it has like updated information on those files. That's like one use case for it, but it's the, does the agent live inside of an environment or does it have tools to talk to an environment? It's like the main difference, I would say.

Okay. Yeah. So this brings me back to kind of the progression over time because like with our go-to-market, like we sit in the same sphere, it's like Clay. I mean, Apollo, like old, old tech, right? Like Apollo, ZoomInfo, Cognizant, kind of like database, database-only vendors, right? Then you have those databases with all of that paid data. And then you have— now we have market data using like Boolean-level searches and everything like that. And then now we have like 2 or 3 years ago, Clay debuted, and now we have one place to unify all the contact and company-level information in one place, up to 50,000 rows. Last year we got MCP with, with Claude. That was huge. Now we see like the HubSpot Ventures backing like active pieces with all the MCP libraries and everything like this. That was huge. Now we're in this like kind of like vibe coding state where it's like, where do we actually build this access layer? And this is like what I'm doing today. I show you behind the scenes with some pretty large, like mid-market companies. They are targeting the enterprise-level accounts. So what they need is they need a market intelligence agent. They need a context agent. They need something that hooks into their vector DB, but they need to be able to query it via Slack. So now how do we build like a system that could do that? I mean, it's just a matter of like connecting the pieces, but in my opinion, it's getting a lot easier to be able to do that via MCP. The real question is like, when do you restrict data to those specific, the specific data, like those databases, either like Supabase or something like this, like sales doesn't need to see everything. HR doesn't need to see everything. This is where I'm struggling to actually build things based on like role-based access control for these larger companies. Because did you see that new Claude bot thing that came out?

Yeah. Yeah.

Did you see that? That it's like you give it access to literally everything in your system and there's like no regard for security. I will not put that on my computer. It's not going to happen.

Yeah. Yeah. No, I've looked into it because, um, the, uh, like Liorin, the employee we have, like he was like, yeah, I'm trying to set this up, not on our systems, but he was just playing around with it. And I was looking into it and I was like, dude, that's, that's just cloud code, but you just give it everything and then you connect via SSH, which like isn't very secure at all. And then people are just putting it on, it was like virtual machines and they're just putting it up and like zero security. They're just like, and, but 5Code is, I don't know, this is an older rant on like, if you know, just in, in GitHub repos, look for API keys, you just find so many.

God. Yeah. I heard about this. The security is bad. I mean, it's, I feel like nobody like really talks about it. Because they don't really care. They're just a matter of like building like Clay system. They're building Cargo. They're loading NNN and it's like, great. Where's the repo? Awesome.

Free keys.

How many times have I duped a Clay table and I'm like, shit, I got to change the API key that's attached to it. And then I noticed 2 months later that I was like duping a table for something and I don't know what's hitting my API key. I was like, ah, shit. And I got to go reset the whole key. That has happened to me more times than I can count when I was like starting, when I was like using like very specific templates. Before they introduced kind of, you remember when you could, remember when they introduced it where you could group the columns together and then you could use those columns as groups? Before you could do that, I was duping the base tables over and then just moving everything. Do you really, do you believe that everything is moving away from things like Clay or are you still pretty like bullish on the platform for go-to-market?

So I haven't been in Clay in a while now. I think it's been like, it's been months since I've like opened up Clay. I still do it here and there. And I mean, I started as some sort of a Clay agency because I was like the first tool that like kind of opened up TTM engineering as a space. I'm not sure, to be honest, like depends on the use case, I guess. I guess it's pretty decent at like building out lists. Enriching lists feel like it depends on who you're talking to because it's still way easier to use a Clay table versus setting up an N8N flow and then versus setting up some Agentic flows. Using Cloud Code, especially like Cloud Code as well. It's like, it's in a terminal, it's in a repo. I think a lot of people are like, I don't know, probably won't really use it just because of that, because they don't really understand. They don't understand, like ideally it's in Git, it doesn't have to be, but they don't understand how to operate it. Yeah, so it's a, I don't know, they're in, like they're in ChatGPT, they're in Claude now as well. Haven't tested it out, but if they keep going that direction, I'm assuming people will keep using it because it will just be very easy to— I'm not sure, like, can you build it in ChatGPT, Claude, or like, what's their feature sets?

I think it's a matter of— I was just speaking to Frank from Salesforce about this in our last episode. It was just like the idea from going from like 0 to 1 is very easy nowadays of like going from like idea to prototype using Claude code or using like GPT or even Gemini to be able to create the code on the backend. Super, super easy. Now you have Replit, Lovelace. I never opened any of those tools. I mean, most of my value that I provide is from my head and my ability to think critically, but like the system has to match the, the, the, the solution. The solution has to match the situation, right? And I think where we're looking at it going, at least I've thought about it, is where we have a lot of people just being able to click buttons and we have like the adoption, like people being to adopt new tools and new technologies like MCT. In Claude very, very quickly with no training. And unfortunately, we're going to get 100 times the volume, but not 100 times the quality. And you see this all over the place on LinkedIn. I know I see it on my feed. It's like, get the template, get the template. And then I comment on it and it's like, waiting for your auto reply. Sent to you soon. Saw one today. So I like was messing with automation, but I think a lot of people are just implementing and adopting new technology and they have no regards to how it fits into the bigger environment. This is the, this is the, like the larger issue that I see.

Yeah. Yeah. Yeah. It's, it's weird. Cause it's like, let's say you're using NA, you can't really build a flow without understanding how to build a flow. I mean, now you can with their, like, I'm not sure what it's called, but their AI assistant that builds flows. Um, but it didn't used to be possible. So you like, you had to learn, like you had to go in and kind of figure out what are nodes? How do they work? How do they transfer data, troubleshoot, et cetera. Um, Clay, same thing, like used to be true. Like you, you had to figure out like, How do you enrich, like, what's that data and how do you continue with it, etc.? AI also wasn't very good initially. Like, I don't know, probably, I don't know when it started to be like really, really good to where you can just use AI. I don't know, even then, like on your point of like complicated formula versus AI, I would always go complicated formula regardless of how good AI and how much faster it is because it's still AI, it's still generated, it's still, there will still be variance and it will still be wrong and a formula is never wrong. That's, that's why it's a formula. So kind of, kind of got lost in what the question was exactly, by the way.

I was just wondering if you're noticing adoption changes and this noise and everything.

Yeah. I think the point I was trying to make is AI makes it so easy to build stuff that you don't have to understand it to build something. But in order to understand, like in order to build something that's actually good to do, yeah, it's one of the things I, I really like about AI, for example, it's like it also makes it very, very easy to learn stuff because it kind of knows you. It has context on you. It knows like other concepts you understand, etc. So when I'm building something, like I'm also one of those, like I want to understand, like I want to know what it does. I don't know why. And I want to, I want to like kind of have this also like in coding, like when you code, you kind of want to have an internal map of what you're building and how it works and how it ties together, etc. I feel like With a lot of people vibe coding, they don't have that internal map. And then if something is broken, they like, they don't even have a map. So they don't even know where to start to look or point the AI. Because AI still isn't like, especially Cloud Code, like I love the model, but it doesn't like, it doesn't always clean up its code after itself. Like sometimes it like, I don't know, it doesn't use anything anymore. It's still there. It writes a test and then, I don't know, tests keep there and then just gets bloated and bloated. And then after a while you just don't, I mean, I want to say you don't understand your code anymore, but I'm guessing if you're 5 coding, you might never have.

I think it's important for you, what you just said, because I noticed that a lot of people are building things that they don't understand and then they're like, oh, have you seen those memes on like Instagram or TikTok where it's like me when my cloud-coded lovable app crashes with 10 users? Uh, I, I, I get this like shit on my Instagram, but like, uh, it's just like very interesting that people don't have to ask the reason why anymore. And this is like very prevalent. And when they're They're not asking why, they're just consuming like crap content on LinkedIn and, you know, like put me on LinkedIn Lunatics, whatever. But I do think there's a lot of smoke. Like nobody actually knows what's real and what's working anymore. This is why I started this series. It's just like, what is actually working? But it's hard. I would hate to be a buyer right now, honestly. Or like a builder, you can figure it out. Yeah. Like I asked Claude, hey, help me figure this out, but I feel like people are using it as a crutch to try to figure out what they're trying to build. They're like, hey, help me build this and don't ask me questions about it.

Just build it. Yeah. Don't make mistakes or you go to jail, stuff like that.

Yeah. Oh yeah. The Daemon prompts.

Yeah. Yeah. Yeah. But it's weird because like the, I mean, the issue is it works, right? Like it very often works until it doesn't. And then you're like, good luck figuring it out because you have no clue what you just built. So I don't know. There's like, either the model gets so good that it doesn't matter anymore and code just, because code already is an abstraction of like how a computer like zeros and ones, right? There's probably gonna come a point to where we don't even care anymore about the code itself because AI is so good that we can go an abstraction higher. And then I don't know what it would look like, like probably the anti-gravity one from Gemini gives you like decent briefings when you're coding on what it's building, et cetera. I mean, in theory, anything is an abstraction jump off code as well, but you have to like put it there yourself. So I don't know, I think there will always be like understanding the fundamentals of something is always going to be important, even if you're like, you have capability to build something you don't understand. I feel like it's just—

The gap is definitely getting smaller.

Yeah. There's also the like good versus great kind of thing, as in you'll probably be able to build something that's like okay or good or works, but I'm not like, I haven't seen AI come up with something that's like net new, that's revolutionary or whatever. I feel like there's still a human thing to come up with new stuff, but it's, I don't know, it's a weird time to be alive.

What's the most interesting prompt that you've used, like prompting principle? Because the last one that I discovered was something called multi-perspective simulation and chain of verification. I would love these principles. This is like amazing to me. I don't know if you have any. Secret prompts up your sleeve that I can steal from you?

I mean, I don't really, it's gonna be weird, but I don't do like real crazy prompting, to be honest. I mean, I use XML tags, which like increase accuracy a bit, but it's nothing huge. I don't know. I don't have, I'm like, like 90% of my time, like when I'm building an AI system is like kind of reasoning through what context does this agent need and how can I make this context as dense as humanly possible. So that's like the least amount of tokens and the most amount of information and make it so every time this flow goes through or this agent is doing this task, it has just that piece of context that it needs. I'm spending way more time on that front than I am on prompting the agent itself. I don't know if that makes sense.

The making it efficient. Yeah.

Yeah. And just making sure like, does it have the data? Like for example, when, um, when like Irvine's, cause it's, it's so such a common use case. I'm spending 90% of my time on figuring out like, okay, we're working with this company. What problem do they solve? Okay. Like this problem, like what are sort of the vectors here? As in, if I'm a company, what are the different data points that make my problem worse or like less worse? And how like, can I figure it out at scale and give my agent context on it? Like, I don't know if it's, what's an easy example, because I usually tend to work like more vertical companies, so there's also more of a, there's a depth to problems when you're solving a vertical problem and you can like figure stuff out. Like for example, I don't know if you're, one company we're working for, they do AI within field service management and like the algorithm that kind of like plans and schedules. So then we can kind of figure out, okay, like what's the feature set they have right now? Like how mature is the company? Do they like, you have the company and you have the market, like how mature are the competitors? Because like if you're not reaching it for everybody, that's just data you have already. So that's, then it's just using it. And it's kind of figuring out like, who are they selling to? Like, who are companies selling to them? Like, is there a chance they're losing deals because competitors are more advanced versus them? And then it's like, what's their roadmap? Like in their documentation, are they like saying anything about like, are they updating the algorithm often, et cetera, to like looking at the team, like does the current team have the capability to build this thing, et cetera. And then it's figuring out, okay, how can I make this information very dense? So an agent is writing the email. It knows company, knows the feature set, knows, knows everything basically.

What I'm doing is I'm doing chaining, like I'm chaining the context together. So if the prompt, if the context window is just, I don't know, like with GPT, if I'm running something through like a GPT, uh, not a GPT key through a Gemini key, I know that my context window is like 2 million tokens or something like this. I think it's like a million tokens if you're using like FlashTube or something. I have to check what the limitations are. What I feel that I struggle with the most right now is actually token optimization. It's like making sure that the context doesn't get diluted so bad from the raw information and that it's crunched properly. What I have to do now is I have to chain the subagents together to be like, hey, you process just this, you process just this. And I think this is where the skill comes in. I don't know that inherently Claude-coded multi-subagent system is going to outperform our intuition. At least I don't feel like it does. I still have to design it. I still have to go in and tweak everything. I'm never gonna get it on the first try. I think like that's where the build time comes in. It's like, I'll quote a project for a month. It'll take me 3 days to make, but it'll take me 2 weeks to test and iterate and like to re-optimize the prompts and the tokens. This is where I spend probably 80% of my time.

Yeah. Yeah.

So even with Clay, I mean, the Clay tokens, the Clay credits, yeah, Clay credits, uh, for, for GTM builds is just crazy. I mean, there's a ton of different ways you can optimize that.

Yeah. Yeah.

Yeah.

That's, I mean, so a few things in, in the agent SDK, you also have the ability to like use subagents and then chain them, et cetera. I think I, I started doing it after like I don't know if you read the like the Cognition article, they were like one of the first that were like, don't use, because I don't even know how long ago it is, but a few months ago, like everybody was basically building agents, but then it was like orchestrator and then subagents and they just kept going back and forward. And then they were like, hey, that's stupid. Just chain them. Anyway, I don't remember when they, uh, when they published that, but Cloud Code is implementing that a lot as well. It's still, still sort of the orchestrator and then the subagents. But the cool thing is it doesn't have to be as in you can just get subagent to write to a file and then next agent read file, update file, read file, update file. And that way you can just chain them.

This is what I'm doing a lot of as far as I'm actually just using for like our personal stuff. We don't, we don't really do anything crazy. Unfortunately, we do all the cool stuff for our clients, but same, by the way, it's, this is such like the, what is it called? The cobbler's son has no shoes, something like this dilemma or whatever it's called. We Dude, I run everything on Google Sheets and just like, hey, update row, create, create row. Okay. Once a week on a cron job, read all the cells in this column, collate, find all, do a temporal analysis, like spit something out for me. And that's how I'm running everything right now. Yeah. I don't have like a VectorDB. I don't have like, I mean, I do have Snowflake for our internal stuff, but yeah, I've run everything that is low budget, low needs in Google Sheets.

Yeah, we do use a repo with like markdown files because I just love Cloud Code and it just allows me to like, I don't know, look stuff up and use Cloud to find stuff for me very easily and fast. But like automation-wise, like I'm not even doing RBACs, like I haven't for like ever, which is weird because I have companies with it.

So we have a backup of all of the, of all of the data. I mean, we have like a data lake of, I think, I want to say close to a million records. Just sitting in Snowflake right now. And I mean, we query that through Cargo currently, but it's, it's great. Like we don't do anything with it. We just, we have the backups of all the data anyways, you know, in case our clients need it. It's like, okay, it's there.

Yeah.

I get a lot of my leads from LinkedIn, actually.

Yeah, same. Yeah. LinkedIn and word of mouth. And that's really all I do. Um, and it's kind of like we meet in phases. Like you'll see, like I post a lot for a week and then I don't post for like 3 weeks and then like I need to post again and then I post another week.

Where do you get your content ideas from? This is my, this is my real question to you. As somebody that's doing AI and, uh, and using AI, where do you pull your ideas from?

Yeah, it's like, I don't have a system at all for content and all my posts, I mean, there's maybe like 2 or 3 that were AI written. All the other ones is just me, like, it's just me in the morning or in the evening where I'm like, oh, that would be a nice post. And I just open LinkedIn and I just type it out and I'm done. It's just manual.

It's, uh, that's what I do. I mean, I get my ideas. I'm building a, maybe when I build this new system on, on Google Sheets, I'll send it to you. I'm doing, uh, I'm, I'm recreating the call intelligence workflows that I've built for some, for some larger organizations, just like kind of like downgrading it for us. But I'm, I'm using Fireflies and I'm gonna pull out all the raw data and then just run everything through Google Sheets and then turn that into content ideas. And then from there, just have a list of ideas. I could send it to you if you want to like dupe it or something.

Yeah, sure. Yeah. But I mean, I've tried, but then on the one hand, I'm like, I kind of like writing out ideas I have in my head because I feel like it, it kind of like, because then you write something and I look at it and like, ah, stupid. So I need to like think more on it or write something and like, I'm genius. I'm going to post it.

And I felt like I've tested fully written AI content and let me just tell you, people don't hate it, but people don't love it. I've noticed that when I shoot a lot of video, and of course I don't have my camera right here next to me, but when I'm like walking and talking, especially lately, they love that shit. It's like the less I try to make content, the better it performs. Okay, cool. Awesome. Um, yeah.

Interesting. Yeah.

Yeah. I mean, I still write all of my posts. What I struggle with is making the graphics. So I'm like experimenting with Nanobanana. But it's just gonna get a designer or something because I don't know, just, it just shows that we're not, we're not our own customers, which sucks.

Yeah. I do use Figma for designs because I've like, I don't know, it's funny. I've always been like some sort of an early adopter. So I was like, I was using Figma in like, I don't know, 2018, 2016. I don't know, somewhere along those lines. And I just got good at it and then it became very popular. Same with like Apollo, like my first sales job. They didn't want to use any tools, so I was just paying for them myself and using it and setting it up, etc. But yeah, I mean, in Figma you can just, or Canva, Canva is like very easy to use as well. But I don't know, I post a lot without visuals and it seems like for me, like my account, there doesn't seem to be an obvious trend between having a visual versus not having a visual. I'm also not like a real influencer, so I don't know.

Yeah, I don't know. I feel like I don't know how many followers you have. I feel like you have more influence. I knew your name way before you knew my name, so that should say something. Are you checking how many followers you have right now?

Yeah, I think it's, yeah, it's 13.6.

Damn, I have like half of that. I'm at like 5,100, man.

Yeah, but it's, it's, I don't know, when I, um, when I started my first company, it was like a recruitment agency. I was like, I don't hate doing sales myself, but I was like, if I want to scale this to the extreme, it's probably going to be way more efficient if you have of inbound coming in versus calling out because like calling out is like, I either gonna, I'm either gonna have to hire people, but I can't like keep scaling the amount of people I can call, but I can keep scaling my audience. And I just started posting every day for, I don't know, 1 year, close to 2 years. And it just went from okay initially to a little above 10. And then after 10, I was like, because 10 was the goal, right? Like this whole time. And then I got it and I was like, yeah, now I'll just post like When I feel like it.

I mean, I, dude, I looked and I was like, in about a year I gained like 3,500 followers and I have like 700 people waiting to be my friend or something. And I was like, yeah, this is a, this is not a qualified person that I need to put my content out to follow. I'm not gonna like accept the connection. I do that a lot actually.

I don't, I'm like, if you send me a request, I'm gonna accept it cuz I don't like, I have like 702 people waiting, dude.

So what I do, what I did before I launched Growth Alliance, I had like like 400 people waiting and I think I had just crossed like the 3,000 threshold and I was like, all right, the day right before I'm gonna accept all of them and then I'm gonna make the announcement post and it worked and it had like 10,000 impressions. I hacked the system.

Nice.

So I'm waiting for the next big announcement and then I'm just gonna make click, click, click, click, click, click, click, click.

Yeah. Yeah.

Yeah. Yeah.

It's interesting. I do still feel like you, you should try class coaches, But do you, do you have some sort of like, in terms of context, like for me, like the main use case is just to have these files compound. Do you just update your prompts or, because you can also use like LangFuse, et cetera, to kind of update them.

But I mean, I don't want, if I'm going to see like, this is the thing, like I don't really do a lot of like, I don't know, quite, quite engineering, like, okay, fine. I put the tools together, I connect the APIs and then I build the initial models but I'm not usually building a system that takes a lot of maintenance. Yes, in the long run, yes, they need to be adjusted and everything, but I would, I would be too afraid to develop something for an enterprise-level company that updates itself because usually they want to have somebody like, you know, like really on the trigger of like, hey, we have an AI implementation specialist on our team. Go talk to him. He's going to help you connect to their VectorDB. If you guys need to change the context layer, you know, the access to the context layer, like he's the one in charge of role-based access control. So talk to him. Nothing changes unless it goes through him. So like, and this has happened 3 or 4 times when I, when I work with these teams and they're like, we don't want it to self-update. So I've just stopped trying to do that. So I never have something like automatically optimize itself. What I do is I just build it once. I build the system prompt, I build the user prompt, chain them together, test the first model, deploy it, make iterations 2 weeks later, do that for 4 cycles. Basically by the 4th or 5th cycle, it doesn't need any updates. Because the output is so optimized at that point. I, on, on my side, I just use Claude skills. If I do something more than 2 times, I just say, hey, we were talking about this before, create a Claude skill.

Yeah.

Ask me 10 questions about optimizing it and then use APO to optimize the skill to be read and used with the triggers relevant to the conversation that we were having before. That's all I do. That's like my actual workflow. It's, I don't consider it very crazy. I feel like simple is better sometimes. I just feel like if I dive into cloud code, I'm just going to get lost in the sauce.

Yeah. No, it's interesting because to me, one of the main use cases for AI is to have itself evolve, to have it learning. That's to me like the biggest unlock because I'm like the company that's like learning the fastest is going to have that edge, right? And you, like there's multiple things, right? If you have multiple salespeople, you'll, there will probably be quite a big of a lag between SDR one on like the other side of the room hears company X say something and then SDR other one on the other side of the room hears the same thing, but they don't connect it with each other because like, I don't know, they have a bunch of calls, et cetera.

This is, and I think this is like the data layer. And I know I try to talk about this is like the different types of input to a system is like data layer, market data layer. Contextual input, like data layer. Okay. Factory data layer, CRM data layer. Where, what layer are you querying when, when you're attaching the LLM to these databases, right? Like, not all the time I need to pull in the full context of the CRM, but definitely if we have like call intelligence from like Gong or Fireflies or whatever, like meeting recorder you're going to choose, I do have a database on their backend. With like Markdown files of all the transcripts and everything. And then that layer gets queried over time. Yes. But am I, am I like going to update the prompt to read those transcripts every time? No. What I would do is I run a cron job, you know, either through Cargo, usually through Cargo actually, say, hey, every 7 days, look at this subset of meetings with sales in the title or with any of these users, pull out all the keywords, give a report to Hey, whatever. We've built that before. Updating the actual prompt to read it though.

Yeah.

Hesitant, given, given my experience and like the people I'm working with.

Yeah. I wouldn't say, by the way, it's the prompt that's updating, but it's the context as in you have the calls and then every night, every week doesn't really matter, but like there's a separate agent goes through calls and then for each call kind of decides what segment is it, what persona is it, goes to segment file X, update segment. And I mean, there's both like one, it's in Git, so you can always go back, you can see what changes it did, et cetera. But then the agent like still has the same prompt, but now when it's, I don't know, doing some outbound for a new company, it's like, hey, this company is this segment, it gets like the file on segment, writes the email. It's prompt says the same thing because you don't really like the, I kind of see prompt as like, that's the SOP. That's like what the agent does and how it does it. And it's like, get this file because this is like segment, like get this tool, et cetera. But the context keeps updating itself.

Context has to be dynamic.

Mm-hmm.

Yes. So the system is static, but the context is dynamic. That's how I'm designing a lot of things right now. Yes. I just had to contextualize that in my head before I said it back to you. But yeah.

Yeah. But do you, um, do you abstract, like, is it, is, because to me it's like the file is like the main, it's not a USP, it's like the main artifact or whatever, because I don't want the agent to query the calls because that's, there's too many tokens, there's too many data there, et cetera. I want it to have one file that's like dense quality context on segment X, persona X, etc. So it doesn't have to go into calls and you can cross-reference and usually like we tend to build everything on the database because agents are very good at SQL, very good at querying databases. I don't like MCPs because they, I mean, I like them, I use them, but I don't want the flows to use them. I just want to like all emails are in database, all calls are in database, all companies are like everything is just in the database, but I do want markdown files that are condensed. So it's like very, very— I think it was Chroma, the like vector database company. They made a post on this as well as like the context rot basically, as in, I think, and they put it very, very low. They were like above 4 or 6K tokens, the output just gets like worse, worse, worse, worse, worse. So I'm always like optimizing for the least amount of tokens possible basically.

I think I work a lot with call intelligence because sales enablement. I'll have to try that. What I, what I do is I usually have full, full context and then I pass the full context into the CRM just for the context and everything like the summary, every, all the keywords, all the competitors that were mentioned based on a list and everything like this. They need full context most of the time, but after every single call, I usually have like a last 5 interaction summary. There's like very high level, be like, hey, we just had a conversation with this person. They, they mentioned that they're talking to linked contact number 1. And it's just a rich text column or a rich text property inside of the CRM. That's like most of like what I'm working on now, because that is like enabling them. I'm actually never having them log in, like the teams. Most of the time they don't usually like actually query their data automatically. They just tell me what flows they want. And it's like, okay, this is where we want to access it. We want to have it in a custom UI inside of Salesforce. Okay. It's going to pop up right here. It's going to be rich text, left side of the screen, right here. Great. Let me go on Salesforce and develop that. Um, or in Slack, they want to, like, a lot of people when I mentioned to them was, yeah, I've built a custom Slack bot for you guys. Okay. Like, we want that. Okay. Now what do we do? Okay. Everybody join this channel that needs access to this type of data and the backend. Great. Now I'll give it access on the backend of that data. Awesome. You need it. You need access to a different type of dataset. Let's put it in here. I feel like that's more where I'm going. You see, again, like, I don't have any use for Claude because a lot of people aren't querying. It's not a human-based query. It's, uh, it's, it's more systematic.

Yeah. I mean, it's query. I mean, agent queries, by the way, not the, not the human. Yeah. I don't know. I would like now, now kind of think like it would be cool to kind of run the experiments as in trying to see the difference in like, it might not be a lot, right? Like it might be, but then it's like, yeah, I don't know how much is the, how much would the edge be of itself evolving? I don't know, is the real answer. Like, if, because I don't know, I feel like, but that's like very much feel like, it's like not data-driven at all. I feel like the models are now good enough to where, because the thing people are most afraid of is like, yeah, but if the AI updates the doc, it will just go crazy and we'll get something completely different. Haven't run it long enough to see, like, I haven't seen it happen. Maybe that's because it's just haven't been updated long enough. Stuff, right? I've been only doing this for like 2-3 months now, this like architecture of it. But yeah, I'm very curious because we do often get the like, when we like run it like for the first time, and then because it's Markdown files also, so just like people can kind of go in and be like, cool, like I didn't know this thing or whatever. We do often get those like, like, is this segment really saying that? And it's like, yeah, like they are. But yeah, I would be, I would be very curious in kind of, and I feel like the larger the sales team probably the higher, like, the, the delta kind is because, like, the more people you have gathering, like, maybe just the more data you have, um, the more it will be, like, distributed across different brains, basically. But yeah, I don't know. It's interesting.

It's definitely, well, the larger the, the larger the organization, what I've noticed is the more friction they have to adopt because it's not just me working with one AE or two AEs and one or two BDRs. It's me working with, like, seven AEs, one in each territory, doing their interviews, understanding What, what custom fields do I need to create inside of their Salesforce or HubSpot? Right. How do you want to access? What do you want it to look like? You know, I think there's an element of querying the data either agentically or, you know, via, you know, human interacted, interact. I don't, I don't know how other to say this other than like the larger the organization, the more friction it comes to adopting and, you know, actually using a new solution. And I think that's like why I haven't barked up that tree. 'Cause like usually the teams I'm working with are like anywhere from like 5 to 25 AEs and they're just like, as long as I get 80% of them to be like, yes, this is what we want. Okay, awesome. That's, that's good for me. It's hard, dude. Like it's definitely an uphill battle to be like, you guys are using AI? What are you guys using AI for? Oh, we just type into Gemini and help us or Claude to help us prep for the meeting. Okay. Okay. What about your security concerns? Uh, legal approved it. Okay. Great. You guys all have access to the same data? No. It's like, where's the data? Uh, we don't know. Great. Awesome. So you guys are all working on different data? Yeah, I guess so. Uh, we don't, we don't really talk to like the SDRs, so we don't know what they're doing. I was like, okay, so now it's a process issue, dude. It's not, it's not, it's not an AI issue. It's a, it's a process issue. So it's, this is like more, I, I, I would run the experiment with you if I can find a willing client, willing client to see like, hey, where are the discrepancies in the data and stuff like this?

Yeah, yeah, that would be cool too. Yeah, I mainly work with like startups as well, by the way. I'm very like selective in sort of the type, like I want to, I want to like tech forward sort of company that's like willing to experiment and willing to like break stuff and like iterate and ideally small, not like it doesn't have to be because sometimes it's just like the go-to-market guy or whatever, but ideally someone that can make decisions fast because I don't know, I've been getting like more requests from bigger companies, but I'm very hesitant. On like doing it because I feel like—

You can trade. I get a lot of inquiries from smaller companies or something. I'll send them to you.

Yeah. Yeah. Yeah. I don't know, but I feel like I don't like, I want to move fast one. And I feel like there might be some bureaucracy. There might be some like, these people want this, this one, like they might like disagree about like what's important, what they want. And I don't know.

That's part of the fun, bro. I mean, I get to, I get it. I, for me, I made the decision with Growth Alliance to be like, we're going to work with bid market and up. But when we decided to do that, or when I decided— there was no we, it was me— when I decided to make that decision, it was like, I would rather serve 5 to 6 to 7 clients only because there's— I can't balance so much context. Like, you know, it's like an agent, like my brain will explode. And I have 3 clients launching their new architecture this week and launching their new agents this week. So it's like, okay, 3 sets of like 5 agents talking to each other. Okay, now we have to go on the backend and like, I have to produce that, but everything was approved. But I, I don't know. I, I don't mind solving a bigger problem. I guess they're all, I mean, size is relative of the problem, right? Like I would rather solve every small little problem and then bucket that together and like put it all together with systems thinking than to just solve like a single issue. Like sometimes people just want to clean up their CRM. Like, yes, that's a, that's a foot in the door for me. It's like, okay, I could do it in a week for you, but what's, what's my question to them is like, you want to clean, you need CRM hygiene. What are you going to do about revenue activation? Oh, we're just going to, you know, pay you like $5,000 or $6,000 to do it and have to clean data. Okay. What's the part of that initiative? Who's in charge of that? Your revenue ops? Yeah. Yeah. Okay. With revenue operations, why is there, you know, why is the CRO asking them to do that? Is it because you guys want more revenue? It's because you guys are losing revenue? Like, what is it? Oh, we're hoping to get more revenue. Ah, okay. So your pipeline is stalled. Oh yeah. Yeah. Okay. So what are you guys doing about pipeline? What are you guys doing lifecycle? And then it becomes a whole question. So I enjoy doing that whole discovery process with the startups. With the startups, it's like usually the same.

Yeah. I don't know. I usually start with like, we do GTM engineering. And like, cause sometimes I get the, like, can you clean our CRM? Our CRM is a mess, et cetera. I tend to just be like, no, like, I mean, I can, but I don't want to. So yeah, we usually do like, but it's a part of it is like the sort of system thinking pitch of like, we map it out and we see where the constraint is and we solve the constraints. And until the constraint is solved, we're not looking at anything else cause that's not the constraints, right? It's not important.

What do you, I mean, I know that we're We've been talking for a long time on this topic, but like, what is, what is the biggest constraint that you're seeing nowadays? Like, because you're speaking to multiple companies, I assume, right? So like, what is the largest constraint that is the most common right now for you as a quote, good go-to-market engineer?

Yeah, it's, um, I mean, it's usually tied to outbound and it's usually a combination of copy isn't good. It's the account selection isn't great. It's not like, I mean, usually they're selecting the right like contacts in there, but it's, that's usually the biggest one. One. It's like the saying, the bottleneck's usually at the top of the bottle, unless they have like great marketing and amazing inbound, which is so rare. If they do, they're like, I don't know, I'm not sure if you're a startup and your inbound is amazing, like maybe talking like venture-backed ones as well, it's like, you'll probably just be hiring and very happy, I assume.

Yeah.

What's one piece of advice that you would give to anybody in the market right Anybody like in the market as in, like, are we talking to GTM engineers, RevOps? Who are we talking to?

Uh, let's, uh, maybe buyers. Yeah, buyers.

This, I mean, this, this is a deeper question in a sense, because like, it has to be like somewhat fundamental so it doesn't change in 2 weeks. Probably something on like, don't outsource your thinking, if that makes sense. Like 5-copying is fine if it's a small thing, but like, reason through stuff. Yourself, like don't outsource it to AI. And if you're like thinking about implementing a system, like reason through like, what does it do? How does it work? And if you're talking to a GTM engineer that can't explain how it works and how it's connected, etc., talk to another one. I think that's the best piece of advice I have.

Cool. Yeah. Awesome. Where can people find you, Laurens?

Yeah, I mean, LinkedIn, since I post there like once in a while. So Laurens, like my name is always like Ness in Dutch. I don't even know how you, how would you pronounce it in like English?

Laurens.

Yeah, but my last name.

Nys.

Yeah. Yeah.

Anyways, I'll put the link. You're on LinkedIn. I know where to find you.

Yeah, perfect. Yeah. So, I'll meet you in—

You have YouTube yet?

Sorry, what?

You have YouTube yet?

Not yet. Not yet. I will though. I'm kind of considering making some sort of a cloud code, like /ssk/ somewhere. I'm not sure how far I want to go because I feel like if I'm going into the data layer, etc., I'm like, I feel like I should need to ask money for this instead of just putting it out there for free. But I might do like a Cloud Code one and it's on like YouTube, LinkedIn.

I would subscribe.

Awesome.

Maybe I'll finally download Cloud Code.

Hopefully, else my course isn't good enough.

All right, Lars, thanks for coming on the podcast and yeah. Yeah, I'll subscribe when you launch that.

Yeah. Awesome. Yeah. I'm thinking like, I'm not sure you have YouTube too, right?

I do. This is where this is gonna be posted.

Yeah. I think I'm subscribed, but I'm not sure. I'll check. I'll subscribe if I'm not.

I'm gonna, I'm gonna release season 2 and 3 soon. We have like a lot of people. Lawrence, you can start to cut and paste what my, my video editor, he'll see this, but like, uh, could like cut it off a little bit forward. But yeah, I will put, this will probably be on like season 2, maybe season 3. I'm not sure. I have a a lot of other sessions tomorrow. I talked to Zach, Zach Vitebor, and then I also talked to a few other people. It's going to be like crazy insane.

Wait, Zach from the We Have a Meeting guy or?

No, no, someone else. From Oxef.

Oh, from Oxef. Yeah, yeah. Never met him, but, uh, super cool guy. Yeah. I texted him once when he was, um, they had the GTM engineering, uh, role open.

I was like, agency I, dude, I asked him the same thing and I was like, damn, what if I was in the Bay? No, all of these guys are hiring full-time. Like, I understand, like, you don't want to outsource your, uh, you don't want to outsource your strategy. I, that's why I like, I'm going to be fractional. I, I don't, and this is, by the time this is published, it'll, it'll be public, but like, I'm actually shifting to like fractional strategy instead of just like just executing. That's what I've been doing a lot of. It's kind of like, I I, you know, I just build everything custom, but it's, you know, do they have, do they have a head of growth? No. Do they have a head of rev ops? No. Do they have a head of marketing? Maybe. Do they have a head of sales? Yes. But that head of sales can't balance all of building everything and doing rev ops and doing like all of this like coding stuff. So that's where I'm repositioning for right now. Like positioning-wise.

Yeah. Yeah. I mean, where's the line between strategy and execution? I feel like it's so blurry sometimes.

Between strategy and execution?

Yeah, I mean, the extremes are obvious, but I feel like there's a big middle part where it's like, because a lot of execution obviously is like based on strategy, but it might be because it's like startups and startups, they don't really have a defined, here's our goals, marks strategy.

Well, I think between strategy and execution, like for me, what it comes down to is what is the situation? Because I'm not going to put together a strategy or I'm not going to put together a solution or like a coalition of solutions for you if you don't have the end goal in sight. Like, okay, what is the main goal? Well, the main goal is to increase, decrease CAC, increase demand, or, you know, look at, go or whatever, have, have perfect data quality, which is like CRM hygiene, da da da da da, you know? Okay. What is the, what is the overarching objective there? It's like, we need to have We want to have data hygiene at 95% plus across all of our record entities inside of our CRM. Okay, awesome. We want to have all of the buyer committee mapped for every single target account. Okay, top, top 50, like top 500 accounts. We want to have like 6 buyers, 6 potential buying committee people mapped inside of our CRM for every account. Great. That's data quality. And then the second one is the go-to-market, the demand generation. How, what are the other channels that we want to do ads, organic? Built like out-of-home, email, LinkedIn marketing, right? Like video, stuff like this. Putting together that type of strategy usually comes from like a, like a marketing strategist. And then there's the whole issue of executing it all together. Like, how do we, how do we connect all the different marketing mix channels to sales enablement? Sales, how do we take the sales enablement initiatives of like doing the cold calls? How do we give them the cold calling scripts? And then from there, making it so that they can execute better, but without changing their workflow. For me, in my head, the bridge between strategy and execution is just enablement and like how smoothly you can do it without disrupting everybody. So that's what I would say. And I think, I think you need sometimes more than just a go-to-market engineer to be able to like sit down and like systematically like map, hey, what are the pain points of account executives? What are the pain points of marketing? What are the pain points of SDRs? Gather all of this intelligence by themselves. And then be like, hey, here's the plan of attack. Yeah, dude, I guarantee you, like, the VP of sales did not do the one-on-ones with the AEs, and the AEs have very specific complaints about the SDRs, and the SDRs don't like marketing because they don't give them demand. It's like sales hates marketing, marketing hates sales, XYZ. So yeah.

Yeah.

I don't know. I feel like, I feel like the keyword there is enablement, and that's kind of like what I'm going to start to pitch, my opinion.

Yeah. Yeah, no, that makes sense. I mean, I see GTM and engineering also mainly as like leverage across go-to-market, and I tend to like I tend to stay away from process as much as possible unless I like, unless I go in and it's like, dude, what the hell are you doing? It's just stupid. But process management is like hard. Um, cause it's people and AI is very against people as like way more.

I don't think so. You think AI is against people?

No, I mean like AI has variance in like you give it a prompt and then it's output and the output will like be different over time.

Okay.

Like in theory, right? Because it's a model. Like with people, it's feel like it's like AI you can manage. People is just harder, but I don't know. That's like—

I think that's a whole other discussion. I think, I mean, I've done some operations work in the backend. I would say that people are just, it's more or less their attitude in conjunction with their abilities. It's, uh, that's what it comes down to. The good thing is about the problem that you have is like you can train a monkey to pick up a pencil and start writing something, but like a human and an agent can also do the same thing. The agent's just going to do it with no attitude, but it's going to have inherent— Yeah. —calibration as well as like the consistency issues. A human, if they're really, really well trained and they value their job and the money that they're getting paid, then they probably won't make that mistake, but the SOP has to exist. So in my opinion, if you, if you tell a system exactly what to do with no variance, do not go outside of that, that, that becomes a game of optimization. But I always thought, I always thought that as like humans as, as well as like people on operations teams, like they understand processes, but everybody outside of operations, they're like, oh, I'm just gonna do it my way. Luckily agents don't really discriminate. They're just like, we're gonna mess up here, or we don't know what's going on, so we're just gonna host an agent.

Yeah. Yeah. But yeah, I don't know. I've always kind of stayed away from the, the people stuff in terms of changing processes, cuz The idea mainly just was like changing processes is harder than changing, like, basically the alpha I get on taking the current process and making it just way more efficient is probably higher for, like, not higher maybe, because you can make the argument that like changing the process, the alpha might be higher, but it's way faster to just take whatever they're doing and make it better versus changing whatever they're doing. Yeah, interesting. But that's a skill issue maybe, like. It's a skill issue? Yeah, it's like a skill, it's managing people, right? That's a skill. Yeah.

I think, uh, I think if you can dilute anything into like an A to Z linear process and you can systemize it like an SOP, then, and like a, like an agent will have no problem following it. But then again, it's not, it's not gonna give you attitude. It's just gonna give you variance. And if you give the same SOP to a person, they're gonna be like, oh, I don't like doing this. It's like, it's like when you go to work in a restaurant, when you do your first first, you know, 5 to 6 days of training, like you've got the head sales barman or whatever, uh, bar lady, waitress, goddammit. Waitress, yeah. The waitress like showing you like, oh, this is how you're supposed to do it. And then after your training's done, they're like, this is how we do it. And it's like, okay, this is how you do it. Go like this. And of course this happens in like taverns and stuff like that. But yeah, the good thing about LLMs is that they have they can do a million and one times the volume of practice that a human could.

Yeah. I mean, it's the good and the bad, right? Because it's also why there's so much noise in emails right now. Yeah.

I mean, I use a matrix for like the outbound emails. Usually I'm actually seeing like the one, it's either one-to-one emails that are written through something like Octave, which I'm super bullish on. It's like contextualization of data from multiple sources or or you have Spintax, which is just like very hyper-segmented Spintax campaigns that are also going to be super relevant no matter what. No personalization, screw personalization, it doesn't really matter right now. But like, the— if you make it relevant, that it's one-to-one relevant anyways. Yeah, you know, yeah, just a matter of what's the similarity in the, in the micro segment.

Yeah, I think that's a big thing as well. It's like, I think it's, it's It's like offer relevance, personalization, maybe something else, and then trust as multiplier across it. But I think if you have a business and your offer is like, I don't know, free money basically, and you have trust, then like, why the hell would you personalize? Just send it to everybody. I don't know.

But it's like, you know, email marketing is dead, whatever. Outbound is dead. Yes, but your offer's probably shit. Yeah, exactly. And this is where like a lot of Like, I don't know, like this is why I don't work with startups. I'm very selective of the startups that I work with. Actually, it's, you know, like Series A, Series B startups, but yeah, I just don't think, I don't know, your offer just has to be killer and it has to, I have to be convinced that email marketing is going to be your channel. If it's not, it's not. I mean, I'm playing around with a lot of direct mail with like ABM, super fun.

Yeah. I know there's an API for it, but I've never, like, it's been on my list of things I kind of want to do. To, but I haven't done it yet.

I might do it for our own outbound, just like not send any emails, just send letters. Yeah. Yeah.

It's like, it's so more fun than getting an email and the open rate has to be, you can't track it, but it has to be good. Yeah.