Why Your AI GTM Stack Will Fail (And How to Fix It)

Julia Nimchinski:
Five. Welcome back to AI Practice Sessions. Today’s theme is AI-native GTM Architecture, and we’re going to showcase practical methods for embedding Agentic AI into your GTM strategy. We’ll explore where these initiatives typically break down or succeed, and assess the risks, constraints, and trade-offs to various GTM architectures. Follow along in our HSC Slack, and please feel free to just ask any questions, comments, or share your concerns. And let’s get into it! First up… We have the black swan himself! Amos, been a long time, how have you been?

Amos Bar Joseph:
I love that nickname. I’m gonna… I’m gonna adopt it, the Black Swan, Julia. Yeah, it’s great having… it’s great being here.

Julia Nimchinski:
Cool, what’s new in this one world?

Amos Bar Joseph:
Wow, so, I feel like so many things have changed in the last 60 minutes or so, so it’s really hard to just, you know, keep track. We’re moving so fast, and every hour, I feel like there’s, like, new insights that we’re learning, more customers to treat, more challenges that are new. It’s just so hard to keep up with the pace.

Julia Nimchinski:
For sure. Well, just for the purpose of today’s show, though, I’m really curious to hear more about, you know, your thinking, philosophy, and practice on AI-native GTM architecture. We’ve all seen this stat, 95% of AI deployments fail. 20, 25 LinkedIn feeds. I know that you’re really particular about the symptoms. So, can you share your, just, thinking here?

Amos Bar Joseph:
Yeah, so there are two reasons why, AI implementations fail. One is because of the architecture of your entire approach of how you’re enabling AI implementations. And the second one is actually human factor. And I hope that we get to cover both, and it’s always… it’s always a problem of both the machine and the AI, of the human and the machine, basically. So, when we look at the architecture point of view. There are 3 fundamental layers that, if you’re not thinking in this… point of view, basically, then you’re… you’re wondering why your AI isn’t working, you’re thinking it’s because of the vendor that tried to sell you that, you think maybe it’s because of the employees didn’t implement it correctly, maybe the technology’s not there, but after seeing over 200 successful AI implementations, in the last 12 months, and a lot more failing implementations, I can tell you that the technology is there, you just need to master these three layers, and if you do that, then you can enable very powerful AI implementations within your company. So, 3 layers are context layer. the orchestration layer in the interface layer. And I hope that we’re going to cover all of them in this session, Julia, and I’m happy to start with context. Are we ready for that?

Julia Nimchinski:
Let’s do it, let’s dive in.

Amos Bar Joseph:
Amazing. So… AI without good context. It’s worth nothing. So, imagine, and maybe you had that feeling before, when you brought, like, a really smart person. to consult with about something in your company, and they don’t know anything about your company, they don’t know anything about your business, they don’t know anything about, you know, your… if it’s in GTM, so it’s anything about who you sell to, what you sell, all the things that you learned, the battles that you lost and won, they don’t know anything, but they’re super smart. And then you try to ask them. how should we generate pipeline? How should we reach out to this person? And they will just give you bad advice, and you were so disappointed. You brought the smartest person to the room. And they just give you very disappointed advice, right? So, it’s the same thing with AI, right? AI without the right context is just like having the smartest person in the room, but just that doesn’t know anything about your business. They can’t really contribute. And… What people often get wrong about context, they think context equals data.

Julia Nimchinski:
That’s not… that’s not true.

Amos Bar Joseph:
Context is data organized in an opinionated way to try to solve a specific problem. Okay? That’s interesting. If you look at, you know, the saying of, let me try to contextualize it for you for a second. When I say that, I’m trying to tell you, look, there’s a lot of data that we’re looking at, I’ll try to package it differently for you so you can understand what I’m saying, right? So, the biggest fallacy in AI is thinking that context equals data, and then if you just throw a lot of data on the AI, we’ll just… Becomes very proficient in how you run your business. And so you think, yeah, just connect it to my CRM, and amazing, but your CRM data is so messy, no one can learn anything from that, and actually, they can learn a lot of, you know, the wrong things out of that. And so… Context is actually how you run your business. In the context of go-to-market, okay? That’s interesting. And how you run your business, it’s not that easy to find in your CRM. It’s not that easy to figure that out. And so… you need to actually make sure that you provide the AI with the right context about how you run your business. So, the simplest example for context is ICP definitions, okay? So, you know, if you don’t have your ICP definitions. working, then the AI won’t really know who to target. And then you have, like, your funnel. So, like, how does your funnel look like? You know, what does it mean, MQL, for you? If you don’t specify what does it mean, MQL, then, okay, the AI could just look at an MQL and think, wait. what does it mean here? It can assume so many things, and MQL means different things for different companies. So does SQL, right? And so you need to define what does that mean? The more specific you are about these definitions of how your business runs, the more… the better the AI could run on top of it. But then. If you get to that point, if you had. you know, written down your context and provided the AI with the right context, you’re halfway there, that’s good. And we can talk about other, you know, primitives within the context model that are important outside of your ICPs and funnels. There’s so much more. It’s how do you run your outreach strategy, how do you do research and qualification, how do you assign leads, etc. There’s so many things to fill in this context model, it’s endless. But even if you did that job, you’re still halfway there. Why? If you don’t have a unified, single context system that all of your AI agents are using. Then that’s a different… mess, okay? And maybe some of you work with, you know, amazing solutions like NA10 or Clay, and, you know. you… you went through the following realization, bear with me. So, you had an ICP, let’s assume that, you’re selling to mid-market companies, in the U.S. that are SaaS businesses, but then… all of a sudden, like, an enterprise company in the services business closed a massive deal in the company. Then another enterprise company, you know, just closed a massive deal, and you say, wow, we can just double down on that ICP. It’s interesting, right? We can just go upmarket, let’s just do that. Let’s start going upmarket. Then you need to define and update your ICP definition. Okay. And what will happen is you start going into every workflow in NA10, in every table in Clay, and every place that you’ve built an AI agent and you define that ICP, you need to make sure that you updated it. Not only you updated that specific point of the ICP definition, maybe it proliferates downstream. Maybe you have a research step that you’re researching that lead based on mid-market, but researching an enterprise companies are totally different. different researching. process. And so if you don’t have a unified context layer where all agents rely on, then you start living in silos, you start creating this context chaos, which you can’t really tame. So that’s… that’s the context layer in a nutshell. We can jump into the orchestration layer. If you have any follow-up questions, Julie, I’m happy to answer.

Julia Nimchinski:
Yeah, this is super helpful, and you know, once in a while, especially the LinkedIn community picks up a word, and then it just… dominate. So context is the word, obviously. context graphs, the concept is obviously not new, but I love your definition, Amos, and I’m curious, what else context, like, based on your experience and all of the deployments that you’re doing internally, externally, because obviously, like, you, I mean. I don’t know how many times we can mention, but it’s 3 founders running, like, I don’t know, like, how would you compare the operation? 10 million ARR? What is it? What are the numbers, actually? And In terms of.

Amos Bar Joseph:
Yes. Yeah, so we’re gonna share a lot, like, very soon, and we have something big coming, Julia, but we’re working with more than 200 customers across all over the world. We’re just 3 people, and… We’re scaling with intelligence in the company, using Swan ourselves to grow our go-to-market, but other agents as well at the company, and we don’t see them as employees, by the way. We’re trying to scale ourselves in this process, but that’s for a different discussion.

Julia Nimchinski:
Really cool. Well, so, in terms of context, you mentioned ICP criteria, CRM data. How else do you see it? What else is essential in order for AgendaKai to work?

Amos Bar Joseph:
Yeah, so… I’ll take one step backwards before that, because I’ve heard you say something that, I have, like, strong emotions against it, and it’s the context graph. Okay, it’s a context graph. It’s all the… all the rave. So there was, like, if you… if you missed that, and that’s totally fine that you missed that, there was an article about, you know, why context graphs are trillion dollar opportunities, right? Yeah. And… First of all, if you see something circling around that fast, and with so many people, it means that there’s something really shallow there, that it’s not really hitting the point, and people really love and sharing, and that’s how social media works, right? So just remember that, and I’m sharing something that… if you missed that, that’s really good that you missed that, you didn’t miss anything. But… people are trying to create this picture that context is so complex to create, and is this something super technical that we need to use AI to build out, and we’re not there yet, but eventually we’ll get there, right? And this… Feeds that… this narrative that we’re not there, it’s a technology problem, it’s not a human problem, it’s not… we’re not the fault, it’s, you know, the machine is the problem. But context models can be very simple to build and create, and, you can just… everyone should and can do it, okay? And… primitives of context models, people think that it’s like AI should just mine the data and create that, but these… but they can be human-generated, right? So I talked about ICP definitions, right? So you probably should have, like, a Notion page or a Google Doc page, you know, describing your ICP definition. Wow, that’s the beginning of a context model, guys. You already… you’re… that’s one step towards building your first context graph. for AI. And… Basically, beyond ICP, so we have, kind of like multiple types of context pieces, you can call it, basically. So, one. is, you know, the core fundamental GTM primitives, like ICPs, personas, like the foundations, right? Your GTM foundations, you can call it, like your value proposition, like your USP, This is, like, out of the book. Nothing new here. Everything that you’ve been taught in, like, GTM 101 is, like, the most fundamental thing in your context model, okay? And if you’re doing things properly, these things constantly evolve, by the way. That’s why you need a unified context layer, because you’re always learning about your ICP, about your personas, about the value propositions, about your positioning. All of this goes into the context model. Guys, this is context, basically, okay? Another layer that you have in context is how you… how you run shit internally? And excuse me for my French. It’s just, how do we do things, okay? It’s not the fun… it’s not the foundations, it’s like, how do we do things? So, how do we assign leads? How do we assign leads? That’s… that goes into the context. If you have a good, SOP, Okay, forgive me for the corporate jargon, it’s just worse than my French before that. If you have an SOP, it’s called the Standard Operating Procedure, for those of you who don’t know. It’s basically, corporate America, jargon saying that this is how we run things internally. So, you need to start having SOPs for everything, basically. The more SOPs you have, then the better AI can understand the most nuanced operations within your company. So, at Swan. And just FYI, Swan has skills basically built into it as a product, so you can create all these SOPs within Swan, and so, what you do, basically what we did, we codified how we run our go-to-market. What does it mean that we codified it? We created these documents. It’s a knowledge base, basically. It’s not a context graph with a lot of, you know, machines operating on it or something like that. It’s a knowledge base, and… It says, like, how do we do outreach? How do we qualify a lead? How do we research? How do we, assign leads that… an inbound? How do we run our ABM? How do we, the list goes on and on. How do we prep for calls, etc. Okay, so… These are, like, the processes that you have, and how you run your go-to-market. Then there… Two more layers in the context, okay? that are more familiar, actually. So one is in account context. So, you know, the old way of doing account context was, like, storing data in a CRM. That’s okay. That could still work for an AI. It’s better if you have a dedicated context model for an account, like we do, but we can talk about it in another subject, but a CRM is still good for an account context. Basically, so an AI could go into the CRM and hopefully just gather context about that account. And if you’re operating, you know, in a good way, then you have, by the way. folder for that specific account, by the way. And in that folder, there’s all the things that that account did. through the buyer journey, basically. And all the people who’ve interacted, they downloaded a form. they, you know, it’s a closed loss, why it’s a closed loss, etc. That’s the account context. And for AI, that could just be a file, with all that context in it, right? But it’s okay if it’s in your CRM as well, and AI could just constantly check that. So account context is the easiest part, because your CRM is designed for that, right? So that’s account context. And then finally. You have personal context, personal preferences, so it’s, you know, we… there’s some things that go top-down, so how do we do outreach at the company, but then there’s, like, how Julia likes doing outreach, right? And, That’s part of running a company. There’s some contradictions between how we run things top-down and how we do things bottom-up, and that’s healthy for a lot of companies, and there’s a balance there that every company wants to take. There’s some companies that don’t give almost any autonomy to employees, that’s how we do things, and so they won’t really have personal context. They won’t be, like, personal preferences, like how Julia runs things. No, we don’t want to hear that, because, Julia, this is how we do things in this company. And there’s some companies that, you know, they enable more agency to the reps, and they allow them to kind of lock in into their zone of genius and understand, yeah, I love, you know, reaching out to these types of personas, because I get these personas. And sometimes, it’s just better to work with that persona, but I get CISOs, and I know the psychology behind them, and I can just work seesaws. And so, if I… have that preferences in the context model, same like in account model. I have a file. that is, like, my file, basically, that I could write things into that file, then the AI could always know what are my preferences. That’s it. That’s the context question.

Julia Nimchinski:
Super helpful. What happens with exceptions, or anything that wasn’t documented? Do you observe it? Does the AI take on that function? Or, like, is it the part of the onboarding process? How does it work?

Amos Bar Joseph:
Yeah, so… eventually, now we’re kind of in the realm of, like, the relationship between the user and the solution provider. It could be an internal solution that is being built, right? You can build things, and there’s a lot of cool things you can do right now internally, but… Right now, we’re talking about the relationship between those two, and from my perspective, for example, us as a vendor, we’re… for those of you who don’t know, we’re creating, like, an AI go-to-market engineer. It really helps you create that context model. It helps you run these motions and build, basically, whatever you want in go-to-market, so you can basically scale with intelligence, not with headcount. And the way that we approach it. Is that there’s a relationship between the user and the machine, and the way that some of the things the user can proactively share and can think about in advance, and some of them are emergent from the machine executing them, okay? And just bumping into stuff, right? And we designed Swan, and I think It’s a good segue to the other reason why a lot of people fail with AI implementations, is that they try to design an AI solution for perfection, not for adaptation. What does that mean? It means that at Swan, we’re designed for the realization that you can’t really think of everything at day one. And then, when you work with SWAN, it’s designed for these feedback loops. And so it could raise something, like, I’ve encountered an error, or this is a use case that hasn’t been covered, and it could surface that, and the user could give its feedback to Swan, and it will adapt the context model accordingly. Okay, maybe this is a new ICP, or maybe this… we haven’t had an SOP for that specific process, etc. And so, basically, it’s not like you need to do everything at day one. If you have a system that is designed for adaptation, not for perfection, then the way that you build stuff is not by thinking actually about everything in advance. It’s just, you know, observing things break, and observing things, you know, evolve, etc, and being able to refine it.

Julia Nimchinski:
Perfect. Checkbox. Context layer. Orchestration.

Amos Bar Joseph:
Amazing. So… Orchestration is… You can think of it as the connection of the agent to the outside world. Okay, so an agent is just tokens, just… or words, or just, you know, something that, you have an input and output, but then what makes it really an agent is that it can, like, do stuff, right? And so, orchestration is, like, all of the connections to the outside world, basically, and… What we’ve realized is that having a siloed orchestration layer also creates chaos. And what we’ve seen ourselves, and a lot of companies try to do, is that every type… every time they try to implement an agent, they need to kind of reinvent the wheel, right? is try to connect everything and understand how it should do stuff. And when you have a unified orchestration layer. Then, any addition, a new agent. can just pick and choose whatever they want, basically, and you can just say, yeah, you know what, from a permission perspective, maybe we don’t want that agent to access that, or we don’t want to overload them with too many choices, let’s just make it confined to these set of tools. But this… idea of having to connect everything, every single time we want to run an agentic operation, we need to think about all the tools it needs to use, and to connect it, that creates a bottleneck of engineering, and it creates a chaos, because what’s happened is that you switched an outreach tool, for example. Now you’re using a different sequencer. Okay, what does that mean? You need to go to each one of these agents and now update them to work with that sequencer. And actually, you know, email outreach, for example, is a very common tool for agents in go-to-market. And so, you just switch provider. What do you do? You just fix all these agents right now? That creates a chaos. But if you have a unified orchestration layer, then it just takes you… one time to update everything, right? So that’s the orchestration layer, and that’s simpler to grasp and easier to understand.

Julia Nimchinski:
Love it. Amos, just to look at this in one second. I’m sure you’ve seen it, Frontier. their framework. I’m curious, in terms of orchestration layer, how would you compare it to their model? Is it the agent execution, evaluation, and optimization? just comparing, you know, Swan AI Opening iFrontier.

Amos Bar Joseph:
Yeah, so basically it’s, it’s in the agent execution layer, so there’s, it says there, model intelligence and tools for agents to plan, act, and recover real-world tasks. So, and why is it called agent execution? It’s, it’s a different terminology, it encompasses two things. There, basically, that I think are separate, but… It’s the context of how to use these tools. Okay? So you can… and that’s just a philosophical argument, it’s for the nerds and the geeks out there, it doesn’t matter and doesn’t have any implication on anything, but you can think of the context of how to use a tool. Is it in the context layer, or is it in the orchestration layer? So they have business context, and then agent execution is, like, the tools, the orchestration, plus the planning and acting and how to work with real-world tasks, like, all the context of how to do stuff. So, for example, we have, in the context layer, we do things differently, we have a skill that says, how to update HubSpot. And a judge. this is the process of how we run our HubSpot updating, but you can think, if I was OpenAI, maybe they will just… they will have it somehow differently, maybe it was just, you know, very attached to the orchestration layer in some way, so they’re not standing on that difference like we do, but for us. We’re very fanatic about it. Everything that is context is in the context layer. And orchestration is about connectivity. It’s different.

Julia Nimchinski:
In terms of orchestration layer, what happens so that the Frontier and Promise, that they encompass all of your existing agents, third-party agents, everything you do, you can just, you know, access it through the platform. What happens with Swan? What’s your philosophy here?

Amos Bar Joseph:
Yeah, so we are competing with them. We are building, basically, Claude for GTM. Our most successful customers don’t go to ChatGPT and Cloud anymore. Why? Because they have a dedicated context model that is designed for GTM. They have these ICPs and funnel steps and everything that I’ve talked about already built in, right? These platforms are horizontal, right? They try to solve these use cases for everyone. We just do it for them. So they have a dedicated context model. Their orchestration layer is really dedicated for GTM, so deep integrations into HubSpot and Salesforce, understanding all the small details of, you know, when you try to update a contact in HubSpot, doesn’t have contact details, you get an error. We have a fallback solution for that. So, it has all the orchestration layer tailored for that. The interface layer is not like just chat. You don’t just jump on a chatting interface and that’s all you do. There’s tasks and so on. You can look at a company’s list, you know, things that are more suitable for GTM. And finally, the most important thing, by the way, I think, our GTM engineer has taste. What does it mean that it has taste to it? It’s our taste of how we think Agenta go-to-market should be. And that’s something that you can’t copy no matter what, so when you want… when you tell Swan, I want to run an ABM motion, it goes to a skill that we gave it. This is how SWAN think. you know, us, this is how we think, you know, ABM should be run, and it tries to customize it based on your context and how you see things, and there’s that ping-pong between our best practices and what you want to accomplish that creates something that is kind of like you can call it expertise of the AI in some way. And so, we’re competing with them.

Julia Nimchinski:
Love it. third layer, is this the interface, or what is it? Interface.

Amos Bar Joseph:
Thirdly is the interface, and so it’s just basic. If everyone’s just going to a chat interface all day. Of course, that’s inefficient for go-to-market. Of course, that go-to-market is more than just a chat interface. And more than that, if you don’t have a unified interface layer, so there’s two problems here in it. So, if you don’t have a unified interface layer, and if you don’t have a go-to-market interface layer, right? Two different problems. You don’t have a unified interface layer means that, well, everyone, like, where do I… I have, like, hundreds of agents, I need to go to different platforms to engage with them, etc. It’s very confusing. We know about this problem of, you know, this… dashboard saturation, so you just get that with agents, right? The second problem is that it’s not go-to-market specific, so you just can’t work in a chat experience all day, right? If you look at… if you go to Swan, you want to look at an account, you have the buying committee, you have, like, the latest, you know, context about that account, what’s their inner ICP, what’s the funnel stage, etc. Instantly, you get all the context you need. We have a desk. So that these are the tasks that Swan tells you, these are the tasks you need to run through. You can’t really go to a chatting interface every time and try to chat your way into everything, so… and dedicated interface for go-to-market is super important. And so, to just wrap that. If you have… An amazing unified context layer. you have a unified orchestration layer, and you have a unified interface layer, then you’re almost there for an AI successful implementation.

Julia Nimchinski:
Love it. Aligned, we have one minute, and I was dying to ask this. What are your thoughts about MoldBot, OpenClaw, whatever it’s called now, and I don’t know if you’ve seen it, but this ambient AI type of innovation as evangelized by, you know, ClickUp. Have you seen it? What do you think about it?

Amos Bar Joseph:
Yeah, I haven’t seen the Ambient AI, but Sam Malt book, big fan of it. Basically… What we’re seeing is the first instance of agentic entertainment. Okay, what does it mean? It means that the content crea… what I… why I think people love about that, and what was fascinating about it, and I’ll be honest, I’m trying to… I’m cooking something on my own. But what I think people… fell for across the world, doesn’t matter if you’re on the side of, this is, like, AI taking over the world, this is humans, or, like, doesn’t matter that fight, what everyone realized there is that when agents has full autonomy about creating content, it’s interesting. That’s interesting. If there isn’t a human that is pulling the trigger behind the scenes and fine-tuning the content and making it PC, if there’s real autonomy in the content creation process. Then that is an uncharted territory, and that is interesting for humans to look at machines creating content. And that’s the first agentic entertainment we’ve ever seen, and you might see more coming out of Swan soon.

Julia Nimchinski:
Love the metaphor, and looking forward. The time flew by! 5 minutes. Thank you so much again. Where should our folks go to… to learn more?

Amos Bar Joseph:
Yeah, so if you want to learn more, we have LinkedIn, I’m very active there, Amos Barr Joseph, you can follow me there, and getswan.com, if you want to have an AI go-to-market engineer that could turn any go-to-market process into an agentic workflow in seconds, stop by.

Julia Nimchinski:
Awesome. Thank you so much again, Amos.

Amos Bar Joseph:
Thank you for having me, Julia.