Latest / The AI Ecosystem Podcast / The Rise of the AI Service Industry and A2A Adoption Trends

Transcript

The Rise of the AI Service Industry and A2A Adoption Trends · The AI Ecosystem Podcast

• Bob Moore: Here we are, folks, back at it again with another edition of the AI Ecosystem podcast. I'm Bob Moore, ⁓ CEO of Crossbeam. Lindsey DeFalco: Hello, I am Lindsay DeFalco, CPO of Crossbeam. Bob Moore: All right, and we're here to talk all things AI and partner ecosystems and the intersection of those two. know, we let a couple of weeks go by. We had an engineering off-site and board meeting and all kinds of other stuff. ⁓ so we're coming back after a whopping, what, two and a half weeks since the podcast episode, and of course there's been billions of dollars of acquisitions that have happened and all kinds of other ⁓ stuff released. what's ⁓ what's going on out there in the AI world ⁓ today, Lindsay? Lindsey DeFalco: Yeah, you know, it's never slow. every week there's a ⁓ a wave of new updates. ⁓ but I think the one thing that I thought was really interesting that I wanted to spend time talking about ⁓ this week ⁓ ⁓ ⁓ services trend we're seeing with them the model companies. So you know, back at the beginning of May, so this is like three weeks old at this point, open AI and Anthropic, that's a tongue twister, ⁓ both announced these consulting or and services arms, which at face I think is quite interesting. ⁓ and it makes you go like, what the heck is going on over there? And then when you dig into it it it it makes a bit more sense. But anyway, I think that's like a a interesting topic us to kind of dig into as we're watching what these model companies were kinda like leading the way with all this stuff or or doing over there. Bob Moore: Yeah, so I this like right out of the gate on this, I think the logical question is, ⁓ why now? ⁓ I I feel like we kind of have gone through the era of this stuff is gonna ⁓ get deployed ⁓ kind of solve for its own ⁓ that it may create from an implementation standpoint or complexity standpoint, and like the the previously existing human rules associated with why doing stuff in big companies is hard don't apply because because AI. ⁓ and and there's just like a simplification of all things that are complex that come with it. And I would argue that probably to a certain extent, like the the model the foundational model companies were like maybe a little drunk on that Kool Aid, as we all were, ⁓ I think in the early days. And, you know, the there's been this massive surge in in Consumption and projects getting kicked off and other things internally. And then you start to see these scary numbers about the percentage of AI product projects that fail, or kind of these like kind of money pit style, siloed skunk work style things that get invested in inside of big companies, but then don't reach the general population of people doing the work. And, you know, meanwhile. You parallel track that with everybody doing a ton of layoffs kind of in anticipation of these in efficiencies and like the the two like the rubber doesn't necessarily meet the road in a lot of these cases. And the maybe the why now is that that realization is is there. I I think maybe the real why now, if I'm an open AI or anthropic, is that that's because of all the things I just said, that is currently where all the money is. Lindsey DeFalco: Mm-hmm. Bob Moore: whether or not ⁓ those two companies ⁓ really fundamentally believe that part of fulfilling their mission to the world ⁓ and you know, bringing about AGI or like making humanity better AI requires this step ⁓ of ⁓ kind of a glue layer of these forward-deployed engineers and consultants to actually make this stuff work inside of companies. Whether or not they believe in that, what I think they are seeing is that they're hitting the limits of how many like tokens they can sell and tokens can be consumed inside of these enterprise environments because like this stuff just isn't working well enough to automatically create rollouts and scale and training and consumption. And companies instead of wanting to like pay money to stand up the AI thing, They want to pay money to ⁓ you know, train whatever employees are left and connect all those systems together to build like more robust, full featured, useful applications out of these things, which very often requires integrating them with systems that have nothing to do with AI. and there's like a shitload of work to be done there, and they'll probably pay a big premium to be able to do it. So, like this is what the big model companies can sell. Therefore, that's where the revenue is, therefore that's like what they want to invest in. ⁓ And by the way, whoever can deploy it better digs the deeper moat too. Right. It's it's a little bit of an arms race of just like footprints inside of these ⁓ these environments that neither one of these companies can afford to lose. So the ground game suddenly becomes important and like having the best model, nobody gives a shit anymore. Like they're all better than all humans are anyway. So now it's like, okay, who who can actually run the play to get it deployed internally in a way that's so sticky it can never be ripped out? nobody wants to lose that. Yeah. Like I so from that point of view it it makes sense. Lindsey DeFalco: Yeah, and it's it's interesting 'cause like the two different approaches the two, like Anthropic went this route where, you know, they have this whole fund that's this like PE backed ⁓ or this like PE firm group of companies and they're pushing it through all of their you know, PE portfolio companies. Yeah. Blackstone, Goldman Sachs. Yeah. And that's like ⁓ that's interesting 'cause it's of screams like, ⁓ yeah, this is a you know Bob Moore: ⁓ yeah, right. They're partnering up with Apollo and ⁓ a bunch of other like they're trying to work the portfolios. Yeah, yeah. Lindsey DeFalco: a a revenue play for trying to ⁓ from the PE companies trying to push this into all of their their portfolio companies. the thing I think that's like really interesting about all of this is just I and we've talked about this before, but I think we just overestimate humans and technology. and the ⁓ there it felt like it f has felt like there's been this huge gap from between the model companies and then like What we talk about on this podcast and like what you know we see with our customers, it's just like eons behind what you hear ⁓ you know, people at OpenAI and people at Anthropic talking about and like what the people that are actually building this stuff are doing. ⁓ and so it's no surprise to me that they're like, come on, hurry up, everyone needs to get ⁓ get on board and and catch up. but I do find it fascinating that they're doing it with humans when like the entire thesis is sure, you could probably talk to Anthropic or Open AI and they could come up with a plan to do it. But their answer is like, let's put humans in these companies to do it. and there's been a lot of patterns like that that I think are interesting where, you know, you mentioned layoffs. There have been a a lot of layoffs. and at the same time, there's a lot of roles opening for specific types of jobs. And so, you know, with these like forward deployed engineers says that ⁓ postings jumped from April of last year to this year, 729%. Bob Moore: From one job posting to seven hundred and twenty nine job postings. ⁓ Lindsey DeFalco: And there's a similar trend in to 729. Yeah. Crazy. Yeah, I'm not reading the raw numbers. Actually, no, it says it's a six hundred and forty-three to over five thousand. So close, close on the one to one. yeah, and it's interesting too. Bob Moore: ⁓ okay. Yeah, ⁓ it's like big enough numbers. ⁓ It does it does make me like this is the the interesting thing is like I feel like Seinfeld, it's like who are these people? ⁓ those I I ⁓ like maybe it's ⁓ there's like some hubris in this statement, but I feel like we've been able to at Crossbeam from our existing like engineering and ⁓ ⁓ Lindsey DeFalco: Yeah. Bob Moore: Superstar folks cultivate very, very specialized and talented implementers of AI systems and reconcilers of what AI can do against what our business needs are, without any to go like pay anthropic millions of dollars to do it. And I guess like, look, at our scale, maybe We have not yet experienced this like exponential complexity curve, ⁓ and we get all these these like startup y kind of benefits. But it is it's interesting that there is so much demand out there in the market for some external entity to just come in and like figure this out for big companies, ⁓ as opposed to what very likely really needs to happen. And frankly, is the spirit of all these layoffs that are happening. Which is like an internal reinvention of how these companies do business from the ground up. And th this is like two there's like a cognitive dissonance that I think is going on in a company that like lays off ⁓ you know, a double digit percentage of its workforce in order to drive AI efficiencies, and then in that same breath goes and spends millions of dollars on a third-party company to come in and implement AI for them. It's like if your own people internally aren't the ones standing this stuff up, then Maybe you need to lay some more people off. ⁓ or like it's just not gonna it's gonna work in the short term, but there's gonna be some fundamental underpinnings about your understanding of how to run a business in this era that's gonna come back to bite you again in another another one year, two year, five year cycle in like a pretty big way. and by by outsourcing those skills you are eliminating a giant opportunity to train and educate and level up the people that you do have on how to do this stuff, as opposed to just making it some like external vendor dependency thing, ⁓ which which is very much is not. Like it it's it's gotta work its way into the DNA of like how work gets done in your company. So I don't know, there's something so that feels very broken about this. ⁓ I'm not surprised by it at all, and it's very human nature and it's in line with Lindsey DeFalco: Mm-hmm. Yeah. Bob Moore: the way companies use these consulting firms full scale pre AI for all kinds of shit that they probably should know how to do internally. ⁓ Lindsey DeFalco: Yeah, the thing I'm super curious about too is like to what extent they're going to embed this into the companies. So like I'm, you know, and I'm obviously it'll vary per mileage will vary, but it's like, is their goal to go in and replace complete functions? Or is the goal to set up, you know, replace all current operations that exist between functions with ⁓ you know, more AI native ⁓ or AI first tooling? And that's where the combination of the layoff patterns plus this ⁓ is interesting. I do think it's different if like a company's trying to replace an entire department ⁓ with, you know, a purely AI ⁓ tooling and maybe they don't want to do that with the people that are in the department. but it is interesting then that they're hiring people to come in and do that. it's all cluster, but I'm really curious about like the extent to which like what you know, what's like the internal goal of OpenAI and Anthropic in terms of like what depth do they get embedded into these organizations? Bob Moore: Yeah. Yeah, because if you look if you're trying to replicate like the business model of like like a Deloitte or like ⁓ you know some of these consulting firms they're kind of like employees by proxy that will like I have friends that have worked in in these shops that work for the same client for over a decade, right? You just you end up being like embedded there and you're ⁓ you're like on the payroll of your parent firm, but like you end up with this like deeply embedded model with all this institutional knowledge. And ⁓ it is kind of a forever business in that way. ⁓ and it it'd be an interesting like vision question to for you know the the leadership in those companies of like, does this feel like a necessary step in the grand transition of AI coming to market and coming to be the predominant like thinking engine for these businesses? That replaces the f the vast majority of humans. or is this a forever business that is gonna is gonna persist in ⁓ providing the human glue between what the AI systems can do and what these companies need of them? And I think if you ask all the big consulting firms, they're they're in the forever business version of this thing because it's kind of like in their DNA. I I doubt that the AI companies think about it in that way. And it does lead to this thing of like, It's like a dating site. It's like them if you're successful, they don't need you anymore. So like they they they all whereas like Deloitte, the most profitable version of that engagement might be when they have that like proxy employee that stays for 15 years at the client. The most strategic thing that OpenAI or Anthropic could do in these situations is deeply embed people, so much so that they're so successful that those people don't need to be embedded anymore. And there are actually like fully stood up autonomous systems that can kind of like perpetuate their their value creation without a bunch of human glue. in which case these may not be forever businesses. ⁓ like comes a point where the dinosaurs have died out and the newer folks just don't need this. ⁓ but it's like it's like this glue moment. Lindsey DeFalco: Yeah. And it's ⁓ it's interesting too because I think like, you know, we talked about that whole like Citrini twenty twenty eight and there's all those like doomsday articles. it's been, I don't know, six months, I think at least since then. ⁓ and I think a lot like a a lot of the AI stuff that's been happening over the year has been scary 'cause it feels like just overnight you're gonna snap your fingers and like all these companies are gonna be gone and it's gonna be like agents talking to agents and there's no humans. ⁓ and so it's this ⁓ it's this is like very grounding to me that this is the direction that they're going right now because it's like that's not coming that quickly. Like if you are now talking about embedding you're spinning up full services and consulting arms to get people to get these enterprises to actually adopt AI. Like if you've ever worked with, you know, an enterprise, those are like multi-year projects and engagements, like transformation is ⁓ you know slow. And I think we've seen that happen a significantly faster, and that's probably part of the entire goal here. But it is ⁓ not only that like they're bringing in humans to do it, but they're like investing in these existing companies and these existing people by doing it. which for me, I don't know, it makes the ⁓ it makes the future seem more exciting and less scary in terms of all the doomsday stuff. And maybe that's because it feels more more comfortable because it's like close to home, but it it does make it feel like the ⁓ the agent to agent and there are no humans, reality is ⁓ is is further off than. You know, a lot of the doomsday. Bob Moore: Yeah, there's a there is a messy middle here. It feels like we're in the beginning of the messy middle. Like the end of the beginning and the beginning of the messy middle. And the way the time you know, when when did the first chat GPT come out? Was that 2023? Twenty twenty two? I should know this. We're hosting an AI podcast. there's like the Chat GPT moment, right? ⁓ Yeah, it was late. It was late 2022, right? So we're now we're going on three and a half years. Lindsey DeFalco: Yeah. Yeah. Yeah, no idea. Bob Moore: ⁓ since the ChatGBT moment, the world to some of us feels like it's been turned upside down, and probably to others it hasn't changed all that much. and so that's like ⁓ you know, call it a four-year cycle just to get to the point where we're having this conversation about human glue layers that are needed in order to bring this to the enterprise in a disruptive way. So, like, as much as everything feels like it changes month to month, like that. That cycle in the enterprise could be another four year cycle, ⁓ until you see like fundamental disruption there in the way that we've kind of seen it in or starting to see some maturity in it in things like consumer search and startups and like all the other like tip of the spear kind of areas. then the world can end in the in the next four years. So we got a good, you know. We got a good ⁓ a good five to eight ⁓ left in the tank here as humanity. It's gonna be it's gonna be good. Lindsey DeFalco: It's extending before it creeps back in right now, I guess. ⁓ Yeah, it is like I mean, if you look back at other trends we've talked about too on the podcast, it's like you look at like, you know, A2A the A to A protocols, like agent to agent protocol. ⁓ and that that was announced back in January, I think. So it's been like almost months or so then since that was like official. Maybe it was November. ⁓ ⁓ know, I have not talked to a single Bob Moore: Yeah, that's exactly right. Yeah. Lindsey DeFalco: Company in our sphere who is using A to A has built an agent that can talk to another company's agent. ⁓ most of them are still shipping MCP. Like, I mean, our MCP is ⁓ in a very long beta, ⁓ you know, hopefully GA soon, but that's like pretty much the the norm in our sphere. So, you know, talking about this like gap of the people who are building the stuff, ⁓ like the AI stuff at the model level and the protocol level, and then the people that are actually implementing it. There's still this long tail of like, I bet you it's gonna be another six months to a year before it's standard that people have an agent they can actually talk to another agent. Bob Moore: ⁓ god, so here, you know what this reminds me of? Because I I followed this closely and tinkered around with it in the like earliest days. So ChatGPT comes out fall of 2022. There was this moment in like the spring through summer of 2023 where there was this thing called auto GPT. And it was this idea that yeah, like it like you could run it kind of in your console, and it was this idea that like ChatGPT, which at the time was just like a question and answer exchange that could happen, people were building this tooling to allow ChatGPT to like operate somewhat autonomously against a task that you give it. And the context windows were still really small and it required like this local kind of quasi memory management stuff, and like you couldn't do a lot of dispatching, it was kind of just like one exchange at a time. But there would be like this idea of The agent could spawn requests to an agent and there could be this like continuously running thing until theoretically you got to some completion. And there was so much hype about it and so much talk about it. And it like totally flamed out and didn't work. And like the models just weren't ready. and the use cases just weren't there. And like the the rollouts, like the like the interfaces and ability to like communicate with other systems, like MCP didn't exist, none of this. So ⁓ so yeah, like auto GPT kind of flamed out. I feel like we're in the auto GPT moment for A to A. And the this was summer 2023, and here we are in like summer 2026, right? So it took three years. Now auto now every single thing that we do with LLMs feels like it's auto GPT. Like these things code while I sleep. Lindsey DeFalco: Mm. Yeah. Bob Moore: Everybody at our company uses cloud cowork to like go, you know, build some document that's too long that we all have to read later while they go take a walk. and you ask one really simple question and it dispatches five or six little sub agents to go hit a bunch of MCPs and recombine the data. And it's like all the things that Auto GPT kind of like promised it could be, we were able to see it. Humanity saw it, and we tried to build these like very, very ⁓ rough early prototypes, but it just couldn't deliver. I think we're in that moment with A to A, like we can see it, we can feel it, we know what it should look like. But there's there's an adoption curve, there's a protocol curve, and maybe there's like a model readiness or systems readiness thing that this needs to coalesce with a little bit better, ⁓ or dovetail with a little better. in order for it to get real. But it is I I agree. Like the agent to agent stuff, I'm surprised We're not at least seeing more pressure on it in a lot of these cases. And I think we're getting by with MCP, which gives people the high of like an agent to agent type interaction, but it's really just an agent to API interaction. ⁓ and what's what's coming out of the other end of the MCP is is ultimately like a dumb endpoint wrapped in a tool description. Lindsey DeFalco: Mm-hmm. Yeah, and I think i a lot of it has to do with like an application's ability to have actions taken on it by something that's more autonomous as well. So it's like and this gets back to all the like foundational work that needs to happen at every company for all this to work, but ⁓ you know, a lot of apps are built for humans to do things to them and sure, you know There are APIs, and some companies are more advanced in terms of like their APIs being able to actually take actions. but like MCP, ⁓ mean, regardless of whether or not you're just getting data or taking action, in a lot of ways, solves a ton of the use cases, the basic use cases that people have now that they would solve with ⁓ with ADA. And I think it's just that like no one's built an agent yet, because they're all still building their MCPs. ⁓ ⁓ Yeah, you see these, you know, you see some companies like ⁓ spinning up tools, like enterprises spinning up tools that have to do with agent orchestration and you know, laying the foundation for like the structural foundation for when these things exist, but then you get into the reality and it's like, Okay, but nobody has an an agent hookup yet. ⁓ not to mention agent means a million different things in a million different ⁓ contexts. ⁓ and a lot of them I think are getting implemented just like many apps or or integrations in these tools. yeah, but I'm yeah, I'm also super fascinated by like Ada because there's been a ton invested in it from like a, ⁓ know, community and reputation standpoint, with ⁓ you know, them like rolling it up to be owned by the Linux Foundation. And ⁓ so it feels like conceptually it's here to stay, but the ⁓ is still so incredibly slow that I'm I'm waiting to see if like ⁓ after the MCP wave, that's ⁓ that's the next trend that we see at least in our our kind of like you know, B2B SAS bubble. Bob Moore: Yeah, it's a good can you think of a not to put you on the spot, ⁓ but I'm not coming up with an answer for this as quickly as I would hope I would, and maybe I will at some point here, but like I'm trying to think about this modality of like when agent to agent is actually more useful or interesting than agent to MCP. Like, if you use the humans doing work as a like a an analogy, I can think of an enormous amount of human work where In order to do the job, you need to go access information and like access truths or protocols or facts or you know rules of engagement or what or whatever as a human ⁓ you ⁓ and retrieve that information and then you can go and do that work. So maybe that's a customer service job where you need to deeply understand a product, maybe it is ⁓ that you are working at a fast food counter and you need to know what prices are and ⁓ speak with a voice and interact and do math or ⁓ you know maybe you're the CEO of a company who kind of needs to go and maybe there's my answer. but then there's like so there's like these classes of jobs, right? Where it's really just like there's a human that needs to reason based on a a bunch of truths and facts and data sources. That really is more like an agent to MCP model. Like The only reason that the customer success human today needs to interact with other humans is that those other humans are their customers who are asking the question, or internally, those other humans are the keepers of the knowledge. And like there's some like institutional knowledge that gets stuck in inside of humans' heads, or they know how to like figure out more complex problems, or they need help. ⁓ and there's like companionship and social and all that stuff. So in a world where agents are everything, that Particular like customer success agent, like knows all of its like internal stuff, and its job predominantly is an agent to MCP. Most of the work that agent's doing is still going out to MCPs. But the exchange between the customer success person and their customer is an agent-to-agent interaction because maybe the customer is not allowed to access those same MCPs that exist internally. Because they involve like sensitive customer data or they have the ability to go in and rifle through internal product documentation to like figure out why a bug is showing up or file a ticket internally with the engineering agent that needs to go fix a thing. there's an interaction between parties where you need to have reasoning on both sides in order to reach a resolution. ⁓ And like I think that's kind of the class of Lindsey DeFalco: Mm. Bob Moore: problems, but in order for those folks reasoning on both sides to do their work, first we need the MCP layer stood up so that they can go and actually retrieve the facts that will inform that reasoning. So like I I feel like part of the reason ⁓ sorry I said it was going to be a question, but then my example turned into my answer. ⁓ it's like part of the reason why A to A is not here yet is because we're j we're like in the beginning of the messy middle of building out the MCPs and the other like fact retrieval stuff. Lindsey DeFalco: Rhetorical. Bob Moore: Which is part of what all these services arms are actually building out, right? It's like these fundamental how do we translate all this institutional knowledge and how do we translate all these internal workflows and processes into things that AI can interface with and talk to and interact with, even if it is deterministic, even if the agent is the only source of LLM firepower and reasoning, and the rest of it is just a bunch of connectivity that never got built or that got built with a bunch of tech debt, so the agents can't talk to it. So we can't get into the A to A era until we really have institutions vary soundly on agent to MCP footing or agent to data layer footing within their own operations. But once the first place, the second will come into focus. Yeah. Lindsey DeFalco: Yeah. Yeah, I mean there's like Yeah, yeah. And I think that's like I think that's all true. And there's also like you know, we're talking about the different level of agents as well. And there was a great Lenny's article about this. I think I I shout them out every podcast. Oops. ⁓ but it was talking about like the three different ⁓ Bob Moore: Just a photocopy of a photocopy of ⁓ Lenny's newsletter from four weeks ago. ⁓ our listeners love it. Lindsey DeFalco: yeah, yeah. but it was talking about the different, you we're talking about like everyone defines agents differently. And it was talking the different classes of agents and like ⁓ most people still only have need for the first level of agents, which really are mostly ⁓ but like AI and the output and like the ability to be more creative and less in the output. And I think that's like a really ⁓ good metaphor for just where everyone is, which is like we're still solving mostly deterministic problems with AI. And we're like graduating into these more complex problems. And it's like the more complex you get, the more complex architecture you need, which is when you get into a lot of this like the need for A2A And also like, you know, you're asking for an example there and I was like crossbeam. Like Crossbeam is a great example. ⁓ if you have Yeah, if you have like an AI, SDR, BD, you know, rep ⁓ that's an agent at a company ⁓ and they're, know, trying to move this account forward, and then they realize, ⁓ hey, I know because have access to the crossbeam MCP in this case, that, you know, this other company has information that I need about this account. They go to that other company and they go talk to their agent that's their like rep, that's an agent. And now that's like across company lines, which is you know, you don't need to be across company lines to be using A2A, but that in itself is like each one of those needs to be reasoning and needs to be accessing multiple different data sources to come to the conclusion that they even need to talk to each other. and you know, what about? And so I actually think like it's interesting because like we're sitting in the the middle of think one of the most interesting use cases for where A2A will be fascinating once these individual Companies that are, you know, our customers and like a portion of the whole B2B SaaS world have these agents that are going out and autonomously trying to actually move deals forward ⁓ you know get information about about accounts. But again, all those companies are still just building their MCPs right now. and so ⁓ here we Bob Moore: It is a good it's a really it's of interesting where pre-Crossbeam the typical pattern for doing cross beamy stuff was very like human to human on individual accounts. So you have like a partner manager at company A calls up his partner manager buddy at company B and says, Hey, do you know anything about IBM we're trying to sell into them? Do do y'all work with them? ⁓ have they bought any of your products? Is there anything we could collaborate on? Because our products work really well together. and it's a yes or it's a no or it's an I can't tell you or something in between. And then that same partner manager maybe calls up five other partners and just kind of sniffs around because the the rep on the deal asks, like, hey, is there any partner support? And that was a very, very slow and annoying and manual and human hours consuming process, but it's a little bit of this like dance of like, can we exchange secrets across company lines through a human intermediary? And then Companies realized this worked really well and and like turned into like very valuable intelligence. So they started doing this practice called account mapping, which was emailing spreadsheets around to do this in mass and you know with all their partners on some kind of regular basis, like massively compare customer list and CRM ⁓ like opportunities and like open pipeline and things like that. ⁓ ⁓ ⁓ was very, very painful and resulted in a lot of leaked data, but also Like when it worked, it really worked, and it gave you data of that nature, this like second-party intelligence data at scale. and that was like even more compelling. And then Crossbeam comes along and basically takes that account mapping process and makes it radically more secure with way more governance wrapped around it, where only the data that is worth sharing ends up getting shared as governed by some some rules of the road that your company can put in place and manage around who can see what. And then you unearth this like ma absolutely massive and validated and relevant like second-party data set. And that's kind of where we are today. It's like the crossbeam data layer is like this you know super badass, like what does my entire ecosystem know about every account? But here's the thing about Crossbeam that is like a little bit of sometimes like hitting a fly with a sledgehammer. You draw these Venn diagrams. Of every single company in my sales pipeline versus every single company in your customer set across hundreds and hundreds of different partners. And you end up with this like mega, mega, super enormous data set, just in case you need to pull one data point on this deal and one data point on that deal. And this rep is about to get on a call with that company. And Crossbeam has basically like built at the ready this giant data layer that you can pull from. That is very, very valuable in aggregate, but in reality, ⁓ there are certain data points that are always more valuable than others, and certain data points that are more timely than others, and that are needed for different reasons from different companies. And what you get from CrossFream is just this giant fire hose blob that most companies don't really know what to do with. So, like, then here comes agent to agent. And to your example, Lindsay, it's like, what if the right model was actually where this whole thing started? Where it's The human to human, that partner manager who calls up partner manager A and then partner manager B and says, Can you tell me anything? Yes, I can tell you this. No, there's nothing to tell. No, I can't tell you. ⁓ what if that is actually the thing that needed to get scaled infinitely? And that you could have those questions and answers and exchanges and yes, no, maybe happen with governance wrapped around it and happen at scale and happen instantaneously across company lines. Well, that Is basically agent-to-agent crossbeam. ⁓ And in that world, then the need for mass account mapping and these giant intermediary like data pipelines and data storage kind of goes away. But what you do have is extremely high frequency, high-volume communications across company lines about individual accounts consistently happening. And then all you need is a system of record for who can see what, when, and under what circumstances. And you need to let your agents access that at all times. And be able to say, hey, am I allowed to tell this partner this thing about this company we're working with? And like that layer is basically like what we will probably end up providing in the AI era when you take it far enough. Better damn well believe ⁓ that's gonna be our our like our domain to go after. And ironically, you could argue that we already have it for 30,000 companies. ⁓ and it's just like the A to A isn't fully baked yet, but we're kind of there ready for it. because we have the rules. Lindsey DeFalco: Which Bob Moore: ⁓ and the agent we have the MCP. So like the agents just need to engage with like, can I share this? And maybe the data layer, crossing actually holding onto the data in the middle doesn't actually matter that much anymore. Lindsey DeFalco: It's I mean it's like what we've always called our platform, right? Like it's like it's the complete abstraction of anything that's application layer plus shaving off part of the platform, but like the core platform is ⁓ remains the value, which I think is like the ⁓ the thing that's getting tested now with ⁓ all companies is like if you shave away the application layer. Bob Moore: Yeah. Lindsey DeFalco: Is there enough of a platform there that remains that actually works when it's just agents and not humans? ⁓ Bob Moore: Yeah, the ⁓ both sides of the sandwich actually. It's like the application layer and the data layer. Cause like I I don't even like we the fact that we house and replicate and run this crazy matching algorithm against like so many thousands and thousands and thousands of companies, CRM instances in order to make the product work may not be necessary in the form that we have it right now if you can do massively scaled point to point on demand. Lindsey DeFalco: Yeah. Yeah. Go fish. Bob Moore: individual account mapping. Yeah, playing Battleship with your partners. ⁓ like you can finally play Battleship without ⁓ over ⁓ over revealing stuff. ⁓ like that that that's kinda well. And it does take me back like get out your bingo cards, folks, like to RJ metrics where at at RJ we were three products at once. We were an application layer, which was the dashboards and the charts and like the business intelligence layer. And we were the ETL Lindsey DeFalco: It's a good fish. Yeah. Yeah. Yeah. Yeah. Bob Moore: the extract transform load, which was all the data pipelines that suck the data out of all your SaaS tools and e-commerce cards and stuff. And we were the data warehouse where you physically stored the data ⁓ after you pulled it out of all those places. And Crosswind is a very similar like layer cake. Like we're basically, you know, starting at the bottom there, where that we're the data storage layer. We actually physically house the data and we organize it in a way that that, you know, unifies it across this global data model so we can do these exchanges and comparisons really well. And then we are the actual governance layer, the rules, ⁓ which is our network graph and our network effect and like who's allowed to see what under what circumstance we are the system of record for how data gets exchanged. And it's just the rule layer is like a sandal and thing. And then we have an application layer where people go in and like the dashboard equivalent where people go in and do stuff and see the data and pull it down. And in an AI world, very similar to how at RJ Metrics we we kind of we dropped the the front end layer and we dropped the data warehouse layer and we We spun out Stitch, which was just ETL, just the data pipelines. In that same exact way, like there's a similar kind of like unbundling of this account mapping stack that I think can happen, where there's consumption that can happen totally independently, and there's storage ⁓ and like systems of record for the raw first-party data that can happen independently. And what you really need in the AI era is this governance layer. And like, what is the trusted system of record that everybody is So that we can all use it to make these determinations. that may be where we actually end up getting getting pulled into in this world. ⁓ because it's gonna be hard. It's gonna be hard to persist like a monolith under this dynamic. Lindsey DeFalco: It's going to be happening. Yeah. Yeah. And it is like it is interesting and I I find it fascinating that anytime, you know, there's all the exercises that every company are doing right now, which is like, you know, for the past eighteen months, which is like, how is AI going to disrupt me? And then it's like, what is what is our product and our company look like in a world where it's just agents and no humans, you know, is a thought exercise. ⁓ but we continue to talk about like I think it's very natural to talk about agents as ⁓ replacing humans in terms of them doing tasks in the same way. then I thinking there's there's reality we're not even thinking about where like, ⁓ you know, they think nothing like humans and it's a there's a totally different paradigm for how ⁓ for how works. But ⁓ anyway, in the meantime, you're mentioning the system of record and I do think ⁓ you know the big thing with ⁓ A to A right now as well is all of the like verification and trust. So, you know, akin to like a Twitter check mark. ⁓ is this actually the Bob Moore, the Robert Moore of RJ Metrics? and that, yeah, and that problem I think luckily exists independently of us and is being solved by, you know, some very, very large companies. But then you have this question of ⁓ kind of like that same stamp, but like an official Bob Moore: Yeah. Yeah. That's a whole other layer, yeah, yeah. Lindsey DeFalco: partnership or relationship where ⁓ you know you can do business as as agents with each other. and I think that that's where our our platform gets really interesting because that's where the network of the actual humans who have made these agreements, both, you know, legal, ⁓ like these legal agreements to be, to be partners, ⁓ all of that, like we're the source of of truth for that. And so ⁓ there's all this kind of stuff I think that like has been handled within an organization by people in a different way that like once you have these agents that are basically like these little like floating people out there that can go do stuff with other companies, it's like how do we regulate that? ⁓ and so and and Crossbeam is an interesting ⁓ input to that because we're the regulation around like what you can do from a commercial standpoint with each other. Bob Moore: It's kind of exciting. like that that I ⁓ it ⁓ even Bob Crossbeam is really Bob from Crossbeam and ⁓ AJ from Quotopath is really AJ from Quotopath, Bob can still ask AJ to tell him a secret that AJ's not supposed to tell him. ⁓ So like ⁓ these independent of trust, which is like, am I who I say I am? And even though I am that person Lindsey DeFalco: Mm. Bob Moore: ⁓ am i asking you to do something or disclose something or say something that I really shouldn't. And in a world where it's it is truly agent to agents interacting, it's gonna be a really big deal because ⁓ these things are they're probabilistic and they can ask or say anything and you can put in system queries and guardrails and like it's very well known ⁓ that there's there are ways to kind of engineer engineer around that. In the same way that humans are flawed and ⁓ you know, you can trick your average random employee to send Google gift cards, ⁓ or like go to the Bitcoin ATM. ⁓ Lindsey DeFalco: I say ⁓ ⁓ I was gonna say I feel like CEOs are like the leakiest buckets at companies, so maybe this will actually help ⁓ tighten some of that up. Bob Moore: ⁓ yeah. I I just accidentally gave up our whole ⁓ three year product roadmap on this podcast. Lindsey DeFalco: Like I said. yeah, another another thing in this in this similar vein that I I think gets very interesting too is like the, you know Crossbeam's been centered around this idea that you can share certain data across company lines. And it's always been CRM data, right? It's like data about you know where the account is, who the who the buyers are, ⁓ you know, how much they bought another ⁓ product for and you know what your like what your potential path is with this this account. ⁓ and I think it gets really interesting in a way where nobody has to build an entirely new product that all of a sudden ⁓ you can share more data. ⁓ so if you know that you know if an agent talks to your if an agent at one company is talking to an agent at another company and it's like, hey tell me this account, then you know this other company can decide, hey, I actually want to share all of my call recordings. I'm actually going to share like product usage data. I want to share this Slack message that happened in our deal channel about this. And a company like Crossbeam of having to build that into their UI and their interface this like massive multi-integration project and ⁓ you know it's always been it's like the the data pipeline part of that is is daunting and having to store, normalize, standardize and figure out how to share all that data has always been like a ⁓ a really large task you know, no one's tackled. ⁓ but it kind of goes away in this world. Like if if an agent can just send another agent data, then you just need the parameters. ⁓ And they can pretty much decide exactly what they want the relationship to look like with their partner, what they want to share across their entire internal data set. ⁓ and like the world is your oyster as long as you ⁓ you have codified somewhere what your boundaries are and what your parameters are around around what data crosses company lines. And that's where I think this gets like very, very interesting. Bob Moore: Yeah, I've been da for years I've been dying to expand the footprint of what data you can like pre-AI you can exchange on Crossbeam. Like I've always wanted to have a crossbeam like pixel you could drop on your website and like, hey, sure would be great to know if like my customer landed on your website and and ⁓ you know things like things like call recording data and things that kind of live maybe a little higher up in the funnel, ⁓ like Lindsey DeFalco: Yeah, yeah. Bob Moore: Who attended my webinars and who filled out what forms, ⁓ and like turning that into signals that could actually help inform across company lines. And I think the the biggest roadblock to that, outside of just like inertia needing to do market making for all those things because nothing like it exists, even on the backs of the CRM data exchange, like there's so much to like the personas start to change and like anyway, that's all complicated. ⁓ but the other thing has just been that ⁓ like humans adopt and understand these things a lot more slowly than agents very likely will. and I guess it's related to the market making point, but it's just like there may be in a world where you can just have every single exchange happen as this individually authenticated and validated thing that has been approved to change hands across company lines, then why limit that to CRM data? And like maybe that's been the challenge is like with the pixel thing, it's all or nothing. It's like everybody in the world needs to install this is network effects where like everybody needs to put this pixel down and everybody needs to adopt this protocol and everybody needs to roll it out and everybody else needs to operationalize it. W with the agent to agent singular data point, like can you tell me more? What more can you tell me? You can have these like one off, very atomic transfers of information that don't have all that much lift associated with them, but could prove extremely valuable because it's just that surgically identified data point that matters the most to that agent in that moment. ⁓ Lindsey DeFalco: You know what's interesting? I wonder I wonder if this will ⁓ in some way compress certain partner integrations. So like if we think about so just in this specific use case, right? Like if and maybe just for us actually. ⁓ but if you think about Bob Moore: Nobody's listening. Just this is just our one-on-one. Lindsey DeFalco: Yeah, yeah. ⁓ actually to that point, fun anecdote. I got ⁓ an outreach email from a product analytics company today and it had a quote that I'm pretty sure only came it had a stat that only came from our podcast. ⁓ but there was Huh? Bob Moore: I love it. Because we made it up and it's fake. Because it's fake and totally made up. That's the only way to validate. Lindsey DeFalco: Yeah, yeah, yeah. No, it was it was hyper specific. It was like you just had your first hackathon, which is wrong, but you just had your hackathon. ⁓ you had twenty three participants and over half of those were sales. And I was like, How do you know that? It's yeah, it's v it's Bob Moore: Yeah. ⁓ that's liter that is extremely specific and not at all made up. ⁓ yeah. Lindsey DeFalco: And I went and checked, neither of us I don't think put well, I checked your LinkedIn. I actually didn't check my LinkedIn. ⁓ I don't think it came from LinkedIn and I was like, huh, I wonder if we have a anyway, thank you, listener. and Bob Moore: Yeah. It was definitely some agent that was like ⁓ scanning every podcast transcript for some ⁓ some fight but hey, I'll take it. A to A, right? Lindsey DeFalco: Yeah. Yeah. but yeah, what I was what I was starting to go down the path of is like if the data that lives ⁓ to your own systems ⁓ be shared externally by your own internal systems, then there's no need for any intermediary to build integrations with those ⁓ So like but I think in our case that actually just matters for Crossbeam ⁓ who's it? Bob Moore: Yeah. Yeah, it's also this like it there's this weird like dimensionality to it where like I might want to get data that is my data out of like post hog because it is like housing data that belongs to me because it's the collection vehicle for it and it has the infrastructure under it. And I need a rule of engagement to like get my own data out of some third party system that is like a walled garden, like multi tenant thing. So I just need to get my data out of my thing. Lindsey DeFalco: Mm-hmm. Bob Moore: And I can go ask for that individual data point. but again, there, yeah, it's like why copy all the post hoc data over to some like Snowflake data warehouse just so I can query it to learn about the one thing I need. and I think it just becomes like these point-to-point individual, like I need to do a task related to a particular account. That's true. But when you want to do like analytics and roll up the question of ⁓ you know, across all visits in the last 30 days, like, you know, if I want to retroactively A/B test this particular ⁓ path or behavior, like how did it perform? You do need it all. And in the same way of like, hey, I want to validate across all my partners across all accounts, like who's the most valuable, who has the most open opportunities, which partner should I invest more energy in? Again, you do need that massive data coming over to do those bigger analytical things. ⁓ But it it's it's a different set of tasks. ⁓ and ⁓ it may just be that you can get significantly more specialized and have like a least required permissions model where any given agent can only get at what they really, really need, you know, through whatever channel. But ⁓ yeah, why this stuff is hard. We should hire some ⁓ forward deployed engineers to figure it all out for us. Lindsey DeFalco: man. ⁓ right. Well did we tap ourselves out on that topic? Bob Moore: I think we did. I think we did a great job. I think we're at time here. is always a pleasure. I'm glad to be back ⁓ back in the swing of things. Lindsey DeFalco: Yeah, back at it after a a nice little mix of vacation and work. Bob Moore: Love it. ⁓ all right, back at it again with the white vans. I'll ⁓ see you on the other side. That's an old you remember that reference? This is getting cut out. It's like a thing from ⁓ it was a viral video. It might have been from like the ⁓ Damn Daniel, it was Daniel. damn Daniel, back at it again with the white vans. Lindsey DeFalco: Yeah, what what was it? It was some guy? What was his name though? Like Brandon or something? Damn dick. ⁓ incredible. I do remember Damn Daniel. That's great. All right. Bob Moore: Bringing it back. All right. Till next time. Lindsey DeFalco: Bye bye.