Latest / The Chief Customer Officer Podcast / EP021: Agent Sprawl

Transcript

EP021: Agent Sprawl · The Chief Customer Officer Podcast

• Jeff Breunsbach: Alright, what's up, Jay? Welcome back to another episode of Chiefcustomer Officer.io. It is gonna be Thursday, June eleventh. I'll actually be in the hospital when this comes out. ⁓ if if ⁓ yeah, we'll be ⁓ we will be ⁓ deliver or I will not be delivering a baby, but I will be supporting my wife. ⁓ currently scheduled, yes. ⁓ could be, you know, we could plans could change, but ⁓ as of right now. Jay Nathan: Hey. ⁓ wow. So are you i is it gonna be induced? Very cool. Well fun fun fact, June eleventh is my brother's birthday, so you know. Jeff Breunsbach: ⁓ a Thursday will be. Okay, there you go. We we are just trying to avoid Monday because that's Caleb's birthday and we just in our minds I think we're just we're we're probably making too big of a deal of this, but I'm just like I I feel like my four year old's gonna have a real tough time if he's gotta share a birthday ⁓ with ⁓ with a sister. Jay Nathan: Yeah. Yeah. Yeah. Yeah. Exactly. All right. Cool. All right. We got a lot going on. Jeff Breunsbach: ⁓ well what right off yeah, right off the bat. ⁓ so I would say a couple things that we're playing around with right now, just right off the bat, and that you encouraged me to go download. So ⁓ nanoclaw is I would say a k it seems like and this is ⁓ I'll I'll give my version, then you can correct me because ⁓ I feel like you're more you know about this more than I do, but ⁓ open claw became a very popular ⁓ way to engage with agents and like have basically a personal agent. Jay Nathan: Yeah. Jeff Breunsbach: But I think it pretty clearly or pretty quickly almost like got out of hand where there's a lot of ⁓ kind of security issues, there was a lot of permissioning issues where all of a sudden like people felt unsafe giving this autonomous agent access to like all of their files and all of their things and then and then could kind of like act erroneously. So my interpretation is nanoclaw essentially is looking to ⁓ build a more secure version ⁓ where you can essentially have agents working on your behalf. But ⁓ in secure ways where maybe credentials are ⁓ hidden in specific ways, maybe access is hi is like only for specific things that you can ⁓ still engage with this type of technology, but ⁓ in a way that you feel like your information and data isn't going to be compromised. How do I do that? Jay Nathan: Yeah, yeah, it's good. to to me, OpenClaw is default like open and accessible. Like it's nanoclaw is the open source version of ⁓ OpenClaw. Time for a haircut, Jeff. nanoclaw is the open source version, I I believe that Nvidia sponsored the open source project. And they they said open claw is great, but let's make it default secure, which means it's a little bit more and there's a little bit more involved with setting it up. I can hear that. and but it's it's more secure by default, right? Like all your credentials go into a a data store ⁓ that is encrypted and they're they're pulled out of there at runtime. So it's more like what you would expect in in maybe an enterprise kind of environment. And so I think everybody as as I've played with this, by by the way, ⁓ I feel like as much talk as we've done about AI, as much as we've been Playing with and doing, you know, in almost like enterprise grade AI. The fact that I hadn't yet set up nanoclaw was a huge gap for me in my understanding of the possibilities here. so when I when I got it set up, the you you you basically start chatting with it and and you build agents by chatting with it. And then you give them tasks, they return, you know. Jeff Breunsbach: Good day. Jay Nathan: return, you know, whatever you ask them to return. And then you coach them to make it better. And they log all of that behind the scenes and and and store it as memory, essentially, that they call upon later to improve the output of whatever you're asking them to do. So Jeff, how did did you meet our new podcast producer, Mav Maverick? Jeff Breunsbach: Yes. Yeah. ⁓ I'd I was gonna ask if that was a ⁓ so just to be clear, or I guess just for so that the record is straight, ⁓ NVIDIA did something different, which is called Nemo Claw. NE MO Claw. Nanoclaw though, ⁓ you you are correct ⁓ that like NanoClaw is ⁓ it's built by two brothers, but they're ⁓ they started a a company called Nano Company and the whole idea like you're Jay Nathan: ⁓ Nemo call. ⁓ my bad. Jeff Breunsbach: Your point is correct that they're trying to create an open source alternative to open claw that is more secure enterprise grade. ⁓ so Nemo Claw is is the NVIDIA version though. Jay Nathan: Okay. Okay, that's right. And and Jack had told me that he tried to set up Nemo Claw and it was just a giant pain in the butt. ⁓ so nanoclaw is what he recommended and that's what we're doing now. Jeff Breunsbach: Yeah. ⁓ but yes, so ⁓ I imagine now I guess Mav is our ⁓ is your ⁓ is an agent that's part of your nanoclaw setup that is ⁓ essentially doing this. And so when you or I guess there's two questions that I have when I when I keep thinking about this. I guess like ⁓ we talked a little bit about this in our office hours actually. We had another uncommon you know, AI hackathon office hours and I actually thought it was is ⁓ it Jay Nathan: Correct. Yeah. Jeff Breunsbach: It was slightly different than the first one where I actually think we got into more discussion and people were asking more questions around these things, ⁓ which I actually thought was extremely helpful. So I think one of the topics that we were talking about was ⁓ and I'm curious your thought on this. I I'm right now, I guess, struggling in the in as a builder who's growing a CS team, trying to think about like how do I enable my team on these things. I'm starting to, I guess, get stressed out about. having agents running in multiple places or different places and not being able to contain that. Like I can already see it happening. And I've like I've got it, I've got things running in Claude Cowork, I have things running in Claude Code, and I have things running in ⁓ Plan hat, and I've got I've got some stuff in NAN. And so now I'm I'm already starting to like be like, ⁓ no, like this is like I need to probably ⁓ I'm actually out for the next couple of weeks. So like one of the things I'm gonna do in my spare time while I'm out is Jay Nathan: Yeah. Jeff Breunsbach: you know, not work, but I'm gonna go research on like, okay, how if I go back in thirty days, like what am I going to how am I going to consolidate this and what's it going to be consolidated too? Because I feel like you basically need this orchestration platform that can see everything that's running and be able to control it. But how do you think about that problem? Jay Nathan: ⁓ okay, yeah, this is a very good question. Agent sprawl. And it actually plays into like our lead story that that Maverick put together for us this week, which is the whole idea of like maybe, maybe even the underlying problem here is just the the increasing cost of token spend, right? And the fact that, you know, are all these agents even doing anything valuable for you? Or are they just experiments that are just running? I I realized the other day that I had a Jeff Breunsbach: Yeah. Jay Nathan: Cowork thing running every day that is basically going to pull everything out of my calendar, all my emails, and basically gives me a daily briefing every morning. I haven't looked at that daily briefing in like three weeks. So I'm just burning tokens, like you know, killing seagulls, whatever. ⁓ so ⁓ but yeah, I I think that's a real challenge and something, something that we probably all need to be reining in. I don't know the answer right off the bat, but Jeff Breunsbach: Yeah. Jay Nathan: I will tell you that The whole nanoclaw experiment has it it inspired me to go build something else for my company, which is something I I'm calling an agent command center, but it's basically an agent studio for us, right? And by the way, I think you could just use to be really clear, I think you could just use if you're running Microsoft, you could run Azure. they have an agent design studio. Jeff Breunsbach: Yeah. Jay Nathan: If you are using GCP or Google, you could use GCP. They have an agent design studio. Everybody's got these things, right? ⁓ so, but what it what it is is basically a nano claw for enterprise. Okay. And it lets me chat with it and create an agent, define skills, and it like saves them away. ⁓ it lets me actually, this is the cool part. It lets me create. Jeff Breunsbach: Yeah. Jay Nathan: What are what I'm calling verified data sets, right? So I'm connecting to Snowflake or I'm connecting to ⁓ databricks or some cloud-based data set that I can pull into the context window for an agent. So I can I can associate a a structured data set with an agent. ⁓ and I can define all that within this interface that I created. And then I can just chat with it to go create all this, which is the cool part. So it sort of behaves like nanoclaw. But it's creating shared agents ⁓ that work across the organization. So I think the key to this, part of the problem, and I made a post about this a few weeks ago, is that Claude and Chat GPT, they they're just contributing to the sprawl because you can create agents everywhere now. You can create them in like you saying, Planhat, you can create them in HubSpot, you can create them in Salesforce, you can create them everywhere. Gain site released at Agent Studio, you can create them there. ⁓ Jeff Breunsbach: Yes. Jay Nathan: Everybody's got their little agent studio, but I think every organization needs to pick one, right? That's outside of any application because the real power of this stuff is being able to connect it to all the different data sources, right? Via MCP or like structured data sets like I just described, and have an agent platform, an agent command center for your team, right? And then maybe everybody is running a personal agent as well, like NanoClaw. Enterprise grade that can connect to some of the same data sources. But at some point, I'm not just doing things for myself. Like even the nanoclaw, I should be able to share what I'm doing so that other people can pick that up and and and those become company level tasks, right? That that just happen on behalf of everybody. Like I don't want, if I have 50 sales reps and they're all on the phone and they're having, you know, conversations, I don't want them all creating. Jeff Breunsbach: Yeah. Jay Nathan: separate proposal decks. I want I want an agent that classifies every call that we have and determines whether that call can still hear me? Okay. And determines whether that call was a proposal call or not, or a discovery call, and then automatically goes and uses our pitch deck template and format and then takes that call transcript, takes our pricing model, our packaging, and creates the presentation for the next call, which is the proposal call. Right. And it sends it ahead of time. So like I don't want everybody doing that. I just want them on the phones having that conversation. So it really does, you know, take so anyway, ha that that's the way I think about it. You've got to have a a team level agent platform that handles everything for you, no matter what system you need to talk to. Jeff Breunsbach: Yeah. Yeah. ⁓ but I think the or I I guess like what's what's becoming interesting to me as I think about this is, you know, we're we're a fast growing company, but we're still I would say relatively small, right? Like we're less than a hundred people. And so like to me, I can start to see how this problem gets I I can already see how it gets exacerbated, right? Because like what platform am I gonna build agents on versus my sales leader, versus my product leader, versus my engineering leader? Like you said, now I now we've got so now I think this is Jay Nathan: Yeah. Jeff Breunsbach: To me, this is actually becoming the same problem that every software company faces, which is it's the same problem that AI is going to have, right? Which is this whole idea of change management in the organization, and that you've got to understand like, okay, how is the leadership level? And and honestly, right, like the the best thing that I think these AI companies can do, and and or like if you're a company that's operating in the space is to effectively go help me have this conversation with my leaders, right? Like if I'm if I am any of these CSPs, if I'm any of these CRMs, if I'm ever whoever right now, like I it actually would behoove you to come to me and say, Hey, you're gonna start to build agents in all these places. And you know what? It doesn't matter if you build it in mine or somewhere else. But we w what we want is make sure that you get our data from our system into those agents and that it's operating whatever. So let us go help you have that conversation. Like, you know, you I I guess like I start to think about like what is a CSM's role? Like, you know, everyone's talking about like what should CSMs be doing and like account managers, whoever else, right? Like I think this is a fundamental thing that you could be doing at any business right now, which is like, what is your AI strategy? How are we connected to it? And if if you guys don't have one yet, how do we help you broker the conversation with the rest of the executive team? Right. Like it just to me, that is the the existential question of like, cause right now I'm I'll tell you what, like to your point, like I'm gonna spend the next 30 days while I'm I'm not in a a full time job to go figure out this. And then when I get back, I'm gonna go sprint as hard as I can to figure out, okay. What's our central platform? And how can I go start spild building agents on that, like you said, that then unlock my team? Because right now, I I mean, I I posted about this the other day, but right now, the number one challenge that I s I have on my team ⁓ is single player versus multiplayer AI. Like we're all using AI, and like you said, we we've developed some skills that we've shared, and that's cool, but that's still to me is like single player AI where we're all in our own context windows, we're all doing our own little thing. And like we've got to ⁓ i if you really want to, I guess like unlock the Jay Nathan: Yeah. Yeah. Jeff Breunsbach: level of AI that everyone is talking about is available at like the enterprise level. It is like you need the infrastructure to go do that. And we currently like have not solved that. Jay Nathan: Yeah. Yep. That y you nailed it. ⁓ yeah. Jeff Breunsbach: The the other well the other the other question that ⁓ I guess has come up, I don't know if you've seen Gav have you ever heard of Gavin Baker? Have you seen his name? ⁓ I'll send you ⁓ he's a ⁓ well known investor in the AI space. And and I would say it's from the I've watched now probably like five interviews, six interviews with him over the last like week. I've become enamored with him. But I saw him ⁓ on the Invest Like the ⁓ Invest Like the Best podcast, ⁓ which is Patrick O'Shaughnessy. ⁓ and he Jay Nathan: No. ⁓ Jeff Breunsbach: So he's Patrick O'Shaughnessy recently did the one with the Claude CFO that became ⁓ quite popular. And ⁓ so, anyways, Gavin Baker just came on again and was talking about a number of ⁓ talking about a number of things ⁓ around AI. And so one of the things that I'm also interested in your perspective on, and it kind of goes to this first story, is ⁓ I guess like is our token prices actually going to fall? Because I think that has been like the That has been like the what every AI company has said. It's like, ⁓ over time these will get better and then it'll be cheaper and cheaper and cheaper. But he was kind of alluding to the fact that like that can only go to some point ⁓ like at some point there is actually going to be something where it's like this can't like we actually can't get these things cheaper and this is just the cost. And he so he was just saying that like, I don't know, that game will eventually run out. We're like, ⁓ yes, this is gonna be cheaper. But he so but he did allude to like, yes, there already are versions where like you could run I could run this same I guess agent on a cheaper model. And so now there's actually like this now there's actually software and stuff that or I guess orchestration that needs to happen that actually helps you optimize your token usage that says, Hey, you know what? This this is a simple task. You're like you said, you're asking for a daily briefing that doesn't actually need this model. It needs this model and it's 20% cheaper. So I don't know I I guess like I don't know if you've got any opinions on like it it you know, does it seem like this stuff's getting cheaper or you think there's the diminishing returns and or like it seems like this software is the next horizon of like, okay, we actually need to have our teams making sure that they're optimizing token usage and like there probably should be software that's helping us like basically meter that stuff. Jay Nathan: The the model, yes. So absolutely. And by the way, like when you choose a company level or a team level agent studio, that's part of it, which is part of the challenge with using Claude Code or Claude for your agent platform, because it's just using Claude models, right? Now or anthropics models. Now Jeff Breunsbach: Yes. Jay Nathan: There are cheaper models that you can utilize inside of Anth Anthropex tools. Like in Cloud Code, you can use something called haiku, which is like a very small model, right? It's very lightweight. I use that for for call classification in my ⁓ call transcript tool that we've talked ad nauseum about. ⁓ I don't use like Opus 4.6 or 4.8 because that's a very expensive model, relatively speaking, in terms of token usage. ⁓ so the reality is token prices are actually coming down. But I think you're seeing Javon's principle play out in real time, which is absolutely the token prices are coming down. But as people learn what's possible here and continue to experiment, like we're just using so much more of them, right? so ⁓ yeah, they they're gonna come down. Jeff Breunsbach: Yeah. Jay Nathan: And we've talked about this before, but the bottleneck is not the bottleneck is actually physical, right? It's energy and data centers being built. And there's this whole there's this whole ⁓ paradox happening, at least in the US right now. There's some some very negative sentiment around AI in the United States in particular. And part of the reason for that is because you've got these clowns, Sam Altnan being one, and what's the CEO of ⁓ Jeff Breunsbach: Are ya? Jay Nathan: Dario of Anthropic, they've been AI doomers, right? This is gonna take your job. Like that's the they're they're they're they're basically pitching this ⁓ what's the word? Like story about what what they're building. They have completely screwed up the narrative. And you see, even just recently that as they've both, you know. Jeff Breunsbach: Doom and gloom story, like a ⁓ Jay Nathan: move toward going public, they've started to shift their narrative a little bit toward the opportunity. But those are not the guys to listen to. Of course they run the model companies, super smart, whatever. They're blowhards. The guys to listen to are Jensen Wong, the CEO of Nvidia, very grounded in a very practical and pragmatic view of the future with AI. ⁓ listen to ⁓ Google CEO, Sundar Pachai. And listen to Microsoft's. Listen to listen to Satya Nadella, the way he talks about the what's happening here. Those are the people to listen to. They're adults, they're mature, they're not freaking weirdos. Those are the guys to listen to when it comes to this stuff. But anyway, I'm way off target of of what you were talking about. But the the the the point is the the model prices will continue to come down, but the use is just gonna continue to go up. And I think it's gonna be more expensive to use these models until more data centers come online. But the but you see the bottleneck is is almost social and political, then it's physical, and then it's you know, what what what are what are the you know, what do these services actually cost? So Jeff Breunsbach: Yeah. ⁓ my last little pi ⁓ pitch I'll give you about ⁓ Gavin Baker is he ⁓ he he's pretty adamant that at some point we're gonna see ⁓ data centers in space because it's like actually an optimal it's like actu actually optimal conditions for it and that like he thinks that that's like Elon like Elon is like going to sol like figure that out or do it. And basically we're gonna have these data centers that are floating in space that like actually require less energy and you like all these things that don't have to take up physical space from us. ⁓ Jay Nathan: ⁓ yeah. Yes. Jeff Breunsbach: So it's like interesting h listening to him talk about that too. Jay Nathan: Yeah, ⁓ you get five times the exposure to solar energy when you're orbiting the earth in a geostationary orbit. So yes, I'd believe that too. I just don't believe SpaceX is the company to invest in at $1.8 trillion. That that's the problem. Okay, that's a whole nother diatribe. We will save that. But that my advice is don't invest in SpaceX. I thought I wanted to. Jeff Breunsbach: Yeah. Jay Nathan: But then I saw the valuation and it's ⁓ it's it's sort of interesting. Well the the Yeah, okay, we'll we'll skip past that topic so I don't get off on a tangent. Jeff Breunsbach: Crazy. ⁓ All right. Well we ⁓ well we've been like chatting through I think that's some of what we're building. ⁓ we talked a little bit last week about uncommon as well. I think we're we're still moving down that path. ⁓ we're still building a lot of things for uncommon right now and figuring out like how do we enable this community and I think bring more of these like discussions that we're having basically into a community setting so that we can like more and more people can be having ⁓ be having this. But ⁓ where where else do you want to go? I think we've got about probably ⁓ Ten ten minutes left. So I what's what's the next story you wanna try and hit? Jay Nathan: ⁓ let's see. So let's see. I mean, we've got all this great, great stuff to talk about now. It's sort of amazing, isn't it? By the way, ⁓ let let's and I'm saying this for Maverick, our podcast producer, because he listens to every episode and then helps us tee up the next one. But let's put some some content in there about Gavin Baker next time. Maybe maybe we can sort of riff on some of this stuff. Yeah. Jeff Breunsbach: Yeah. Actually well let's talk about that real quick. So how did you set up ⁓ how did you set up Maverick? Like and what's like you said, like what's he doing? So it's ⁓ it's so it's on a nanoclaw, which allows us allows you to basically ⁓ I don't know if you're using Telegram or WhatsApp, but like allows you to text with him ⁓ Slack. and ⁓ and so then you essentially just set up ⁓ requirements or rules that you wanted to run, ⁓ go listen to all of our transcripts, ⁓ figure out like basically the topics that we're interested in, where we're going, and then it goes and does research online. Jay Nathan: Slack. Using Slack. Yeah. ⁓ yeah, goes and does research online, pulls out stories. And and it actually started by it sort of seeded the content by going back and reviewing all the transcripts from our prior podcast episodes. And so it automatically does that now every week. So the first thing it's gonna do is go pull down the latest episode, add it to the archive of transcripts so it knows what we've been talking about. Like in our document here, which is so cool, like there are references to to prior episodes. this is a little long-winded still. I think we gotta we gotta train Mav to be a little little less wordy. But ⁓ but yeah it references back to to old old podcast episodes that that we've done sort of tie it back together. So that's why, you know, all we have to do is say, hey, like we want to talk about Gavin Baker on the next podcast and he's gonna go pull, you know, research on that ⁓ and do it. So yeah, I just fed him the link to our RSS feed. Jeff Breunsbach: Yeah. Jay Nathan: in the podcast and he goes and and pulls down the transcripts before he does anything else. Jeff Breunsbach: It's called Nice. Jay Nathan: So let's talk about ⁓ okay, so this is sort of interesting. Canva, sixteen billion AI interactions, post sale thesis that should embarrass most companies. So ⁓ I think the story here is that Canva gave their entire team Well, gave five thousand employees an entire week to do nothing but learn AI. The tools were ready, but the team was not. Evidently they they froze. They didn't really know what to do with it. ⁓ So what are your thoughts what are your thoughts on that? Like how do you get your team ready? And you're we're not dealing with 5,000 person teams here, luckily, but like you're running a CS team at Junction. Like what are how are you getting your team ready for this stuff? Jeff Breunsbach: Well, I think yeah. ⁓ I think this is something I've been thinking a lot about ⁓ about the single player versus multiplayer, to be honest. And so I guess there's like two two ways that I'm trying to approach this right now. One is just through sheer modeling behavior, like as a leader. So ⁓ you know, on our like on calls with my team, during one on ones we might be pulling, but like I'm I'm at some point trying to like revert back into like, okay, ⁓ How are we, you know, how are we going? Is this something that you can use AI for? Is this something that's an automation that we need to clean up? Like trying to revert it back to something and showing them how to do it? Like for instance, you know, ⁓ we recently set up a ⁓ similar to you, we we recently set up ⁓ customer quotes channel ⁓ in our Slack. And so I built a automation with Claude Cowork that scrapes all of our Fathom calls and looks for Quotes from our customers that say nice things about us, junction, the team, and ⁓ so now it automatically pushes in there. And so like I like walked my team through that example and like in real time showed them, right? Because I could press the run button and it actually, ⁓ look, it it populates these calls in there, right? Like you can actually see it working. So like trying to find examples like that of just showing them, okay, here's the power of like these little things that we can do. And so ⁓ that actually was helpful though, because what spurred that is they said. My team immediately started jumping to other use cases of listening to c call transcripts. They're like, ⁓ you know what'd be cool is if instead of just listening for customer quotes, like what if it was listening for product feedback? And could we actually push that product feedback to our product team versus the CSM having to jot down the note or like go file the linear ticket themselves and it in and you kind of miss, like you said, like every time you do that, there's a game of telephone, right? So I feel like by the time it reaches the actual product manager, it's gone through like three or four hands and and or like three or four iterations and all of a sudden like I you know, I think our head of product, Boris, actually does a great job of this. We're like in any instance we bring up a piece of product feedback, he knows that we record all of our calls and he's like, Can you just go clip that card part of the call? He's like, Not that I don't believe you, but like I just love to hear it from the horse's mouth. And so, ⁓ so anyways, that so that actually was like a my team saw that I did that, customer quotes, and immediately they were like, Hey, can we listen for these signals? And so then I've asked them to basically go put together like, Okay, what how would you train the model on like product feedback? Like what What are what are the names of our products? What are the key words that we listen for? How do we train it on what we do? We actually have our API docs. So can we actually train the model? Just hey, can you can you basically go ingest all of our API docs and these are the signals and things to be looking out for? So that's like one I guess like showing the behavior is like one thing that I think about. Jay Nathan: Yep. Yep. So I think this also comes back to the initial conversation we were having too is if you're gonna do something like this, you have to go something like what Canva did, and maybe they did this, right? So I'm not saying they didn't, but you have to go give people a structure to start with. So one of our companies actually did this ⁓ within a division or or a department at Walmart over the past couple of months. We actually did an AI hackathon. And what we did there was we basically set up all the tools. We set up some data sets that were going to be available to those tools. What I think you could do is like nanoclaw is a great example. Let's figure out how to go install that for everybody in a safe, secure way across the team, right? Ahead of ahead of the week of hackathon, whatever. Maybe it's not a week, but ahead of that. And then also build some data sets like. Maybe it's your customer data set, maybe it's your ARR, maybe it's your invoicing, maybe it's, you know, whatever. And then when you go to do the hackathon, the first thing you do is training, right? Okay, here's how we installed NanoClaw on your machine. Then here's the data sets that we made available to you and the things that you can do. Now we just want you to start chatting with nanoclaw over the next 24 hours. What ideas do you have of what to do with that data set that we verified for you already? And then see what comes out of that. And then I think you can sort of aggregate that stuff and say, okay, like what of what we built needs to be a company or a team level agent versus you know sitting on your nanoclaw running in your little personal agent space. Jeff Breunsbach: Yeah. And will you ⁓ 'cause you mentioned it a couple of times and I think I understand the importance of like why you're you're referencing it so many times, but like why do you I guess why do you believe in like this verify verified data sets idea? Like what what to you is like critical about like making sure that that is like a part of the equation? Jay Nathan: ⁓ good question. So I think you know a lot of our systems of record are designed to track data transactionally. Right? Like for example, we have I don't know, I'm gonna talk through this in real time. We have time tracking system again, like we're we're a services business. ⁓ we have a CRM, all that kind of stuff. But we also take all that data out of those systems and put it in one centralized Data warehouse basically that we can do aggregates on. So we can do weekly summaries, monthly summaries of that data, especially the the numerical data. ⁓ financial data is a good candidate for this as well, right? Being able to sort of aggregate ⁓ data. So there's two reasons for it, in my opinion. One is to make put the data in a sh and I've got a my background is data warehousing, ETL, like analytics. Like early in my career, that's what I did, right? A lot of that. And so to me, there's still a lot of power in having structured data, right? That we know what it means. We know why we're looking at the data in the certain way that we are the analytical views versus the the transactional views of the data, because most of these systems are made to to process transactions very quickly. And I think that's only going to get more important because agents are going to use them much more than we've than humans ever could. So That's one thing, just n being able to know with certainty that the data is accurate for whatever use case you want to build on top of it. The other thing, and I think it comes back to the first conversation, is token efficiency. Right. If I've got Claude calculating everything from scratch every single time, it doesn't make a lot of sense to do that, right? So why would you burn all those tokens when you could actually Jeff Breunsbach: Yeah. Jay Nathan: And by the way, we're AI is doing all this, right? It's doing the aggregation of the data, it's building the ETL, whatever. Like those are just structured deterministic programs at some point that are loading data on a you know minute by minute, hour by hour, day by day kind of basis. But then ⁓ but then you you're you're actually burning fewer tokens by not connecting to an MCP, connecting to a verified data set, going to get the whole data set at once. You already know what the shape of it is, you're feeding tons of context about that data set. into the LLM before you're asking it to do anything with it. So it doesn't have to go figure all that out every single time an agent runs. Does that make sense, those two things? Jeff Breunsbach: Yeah. It does. Yeah. and I mean I think at the end of the day too, like I guess like the ⁓ the word that comes to mind for me is like that you have trust in the data that you're look that you're that you're using, right? And like 'cause at the end of the day, if I'm gonna go use an agent to do something, like we've actually already run into this scenario at our company where ⁓ we've got an agent actually that surrounds our our data. ⁓ we're using a tool. I'm not gonna call the tool out right now, ⁓ because we're still on like a beta with them. But ⁓ there have been questions where somebody like goes and you can chat, right? I saw I'm chatting with it and there have been times where our team chats with it, it pulls a data set, it gives them like a nice visualization. And then they're kind of like, ooh, I don't know if I believe that. And it's like, well like why? You know, like what's what's and so we have to we we've we've actually had to dig into it and say like, ⁓ that doesn't look right or that doesn't and and like our team's saying that just from like a feeling of like, ⁓ I kind of know generally the numbers and like that just doesn't look that doesn't like kind of play. And so we've had to go do a lot of digging, but like that that mistrust now is like Jay Nathan: That's exactly right. Jeff Breunsbach: Almost like lingering where it's like all of a sudden like, okay, we put this tool in front of our team, but our teams don't trust it and therefore they're using it less. And therefore they're actually going back and reverting to pulling stuff manually. And it's like, okay, this isn't the the whole reason to do this actually is like ⁓ we we've lost it. We've lost the plot. We actually need to go back and like verify the data and make sure that's like like the accuracy is there and like almost come out with like a white paper for our internal teams to be like, hey, we've gone our head of data and AI have gone and verified this. It's like true true and accurate, you can use it. Jay Nathan: Yeah. Yeah. That's such a great way to say it. So two things which make my long statement ⁓ much simpler trust of the data and feeding context to the L L ⁓ to to make it more efficient and faster, frankly. So Jeff Breunsbach: Yeah. And that that to me like makes a ton of sense too. Like you don't want ⁓ 'cause I've already done a couple of things where I've had it like start calculating certain things or I've had it just like, Hey, pull this data in, can you do this analysis? Like you and all of a sudden like my token usage just like burned through and I'm like, ⁓ that was I I had it like doing addition, you it's like I had it doing like some simple equations that I just didn't want to do in Excel before I fed the data shit seed in, and it's like I should have just done that because it would have actually saved me in the long run. Jay Nathan: Yeah. Yeah. Have you ever had the experience of, you know, doing something before eight AM in the morning in Claude and you're like, Well, shoot, I just burned through all the tokens that I had for the next five hours. That happened to me multiple times last week and I just freaking upgraded. I was like, Forget this, I'm not dealing. Jeff Breunsbach: Yeah. Well I think the worst I think the worst situation is when you're trying to get something done and then it's like you know, like you've got a couple of things lingering. It's like at the end of the day or you're trying to get, you know, it's like, ⁓ I gotta get this one more thing done and it's like you're out of tokens. You can't until like nine PM at night, you're like, shit, you know? ⁓ Jay Nathan: No, that's when I that's when I could I switch over to I have two I have two Claude premium accounts. I'm like, okay, I gotta use the other account now. And then I can always sometimes I'll go when I ⁓ am just asking questions, I'll go search Google so I get the Gemini I use Gemini tokens for tokens of the new water. They're the new currency. It's crazy. It's like Star Wars. Jeff Breunsbach: Yeah, yeah, yeah, yeah. Just get the yeah, that's a good idea. ⁓ yeah, they are. ⁓ all right, well, ⁓ I think what it's a good good wrap up, but I think ⁓ it was cool to talk through. So I think we talked through a bunch of nanoclaw ⁓ elements. I think trying to talk through like orchestration, how do you think about like really kind of building a platform for these agents to run and try and get your teams kind of on this multiplayer AI idea. ⁓ and then I think we talked through a little bit of ⁓ and then I think we're gonna go and ⁓ or we talked a little bit about Canva and some of the ⁓ challenges of how do you roll this stuff out to teams, how do you start to to get adoption and start to to build around that. So ⁓ some good topics. But Maverick, I think you've got some work to do for the next time. But this should be a good ⁓ well in real time, I think we're we're gonna see how Maverick gets better and better at ⁓ I think some of the some of the ways that we pull these stories together. Jay Nathan: Yeah, I mean th there's some interesting like quick hits in here that he and maybe we talk through a few of these just to figure out like for ourselves like what is interesting for for our crowd, right? Who's listening to this? Because what I what I've also found, and I actually had Maverick add just added this morning, you know, I said, Hey, every week I want you to pull, you know, four or five new tools that are out there that we should be talking about. Because What I'm finding, even with my own team, it people have a hard time keeping up with this stuff. Like you and I, like, I think we're eating, sleeping, and breathing it right now. ⁓ but if we weren't, like we, you know, listening to podcasts basically six hours a day and you know, reading everything on Hacker News and whatever, like we we might not know about some of the tools that we've heard about. No, it's like the trend and the zeitgeist of this stuff. So ⁓ I had them add that. So that might be an interesting thing for our crowd. Jeff Breunsbach: Yeah. Jay Nathan: ⁓ as well. But here's some here's some quick hits. And and ⁓ I'm curious like what what out of this peaks piques your interest. So Sierra AI, and if you don't know what Sierra AI is, ⁓ that is a like a CX ⁓ agent platform. So predominant use case being contact centers customer ⁓ experience. a lot of a lot of B2C use cases ⁓ Jeff Breunsbach: Yeah. I'd say a lot of B to C use cases. Jay Nathan: Brett Taylor is a CEO. He is the former co-CEO of Salesforce, right? So he's basically leveraging his his personal brand to go build this thing. They're at 150 million in ARR. ⁓ they raised at 900 and ⁓ no, they raised 950 million at a 15.8 billion dollar valuation. What is that? A hundred times? Is that right? Is my math right? Jeff Breunsbach: Salesforce. Yep. No. Fifteen ⁓ yeah it is. Yeah. Jay Nathan: Yes, it is. Hundred times revenue. so that seems crazy. but okay. Side note, I tried to have some conversations with Sierra about, you know, what would it be like to for us to spin up a a partner firm, you know, a solution partner firm. It's like, no, no, we can't talk to you. We only are talking to the big five right now. I'm like, really? So I just had a friend who went there. It'd be interesting to see what what that looks like from the inside out. but what do you think? Have you run across Sierra? One one of our clients uses Sierra today. And they like it. Jeff Breunsbach: ⁓ yeah, I've I've I've run ⁓ yeah, I've I've seen it a bunch. ⁓ I don't know, I'm a skeptic around all these like CX tools because like my interactions with them n not ⁓ only more so because I feel like every single interaction that I have with a ⁓ an automated CX tool, AI CX tool, I'm I'm always just like, Can I please just talk to the person? Because I've I've now had to like chat with you longer than it would take to solve the problem and it always just annoys me. Jay Nathan: Yep. Yep. Jeff Breunsbach: So I don't I I I mean I've heard of C C R I have not like I I guess like I don't know if I've interacted or engaged with like the actual product itself, but like I guess to me the promise of those tools continues to feel very like you said, it feels like you need deterministic ⁓ a lot ⁓ maybe in like a lot of ways, it feels like it needs deterministic ⁓ patterns in order for you to actually release it to customers because you don't want almost like these probabilistic ones to like go haywire and I don't know. So it just feels all ⁓ in all these cases I I end up just trying to talk to a person 'cause I'm like this and now I'm getting annoyed that I have to write paragraphs back to you about like what I'm trying to solve for. Jay Nathan: Yeah. Yeah. Yeah. Yeah. I think ⁓ a a much better use case, especially if you have a like a digital product, which that may not always be the case, right? But is if it's like monitoring what you're actually doing in the product and responding proactively instead of waiting for you to have to tell it there's an issue. But I I do think like in some B2C use cases where you're like emailing support, those are those are pretty good. And ⁓ I was talking with a private equity operating Jeff Breunsbach: Yeah. Jay Nathan: advisor kind of person in the go to market space, ⁓ a couple of weeks ago. And she basically said they are almost giving their all their portfolio companies a mandate to have like ninety-five plus percent of all like technical support issues handled handled via an agent by the end of the year. I th that that's the kind of ⁓ Jeff Breunsbach: Support. Yeah. Yeah. Jay Nathan: And it's it's totally possible. Right. It's totally possible. Because ⁓ i in really what you were saying, it comes down to what context are you feeding it and what are you allowing those agents to do on behalf of the customer behind the scenes. Right? I mean, that's it. I mean there's there are use cases already. I think we talked about did we talk about this already? Where a bug gets submit submitted, right? An issue is found in the product, that Jeff Breunsbach: Yeah. Jay Nathan: Issue is then taken and handed off to a coding agent, which actually goes in and figures out where the problem is, solves the problem, creates a pull request for a human to review, and then the human reviews it, checks it in, goes to production, email automatically goes back out to the customer that says that their issue is solved. Like that is 100% possible today. Now, how many companies are doing that? Probably not many, relatively speaking. Jeff Breunsbach: Yeah. Yeah. Yeah. But also but also right, like the humans in the loop in that. And that's also mainly in the back end, right? Like that's just my point about I think like some of these things like is that ⁓ it feels like we're trying to go ⁓ like in some of these cases, right, go put this stuff in front of the customer so so quickly that like in a in actuality, like would it actually be better if you just like optimize the internal process to make those people just much more effective and then go push it? Like for example, like we've got some automations running on the back end for our technical support team and Jay Nathan: Right. Yep. Jeff Breunsbach: So it'll it'll basically take a technical problem, it'll go auto triage it, it'll basically try like you said, it'll try and solve the problem first or try and go pull documentation to solve the problem and then put it in front of the person and then like they are the ones who are basically, you know, trying to go figure out the right next step. Jay Nathan: Yep. I just think that I saw a really interesting article this week on ⁓ actually, you know, it was from the it was from the click up CEO. If you haven't read that one yet, like why they just laid off twenty two percent of their company, it's a good read. And his whole point there is that incrementality is not good enough for a lot of for most companies, especially tech tech companies, right? You've gotta really like what you just described is the incremental way to do it, right? It's like, okay, let's make this Jeff Breunsbach: ⁓ yeah. Jay Nathan: Process 10%, 20% better. But what he's pushing for is 100 times better. Right. Like, how do I how do how do I go do like go all the way to the end of the story, which we know is probably possible, but let's go shoot for that and figure out where it falls short. And then maybe we land at 50% better, right? Versus, you know. So anyway, interesting. All right. You want to do the next one here? Jeff Breunsbach: ⁓ Wix cuts a thousand jobs. Jay Nathan: Or maybe let's just do one more. Let's just do one more. How about that? Jeff Breunsbach: ⁓ yeah, I mean I think well ⁓ tech layoffs hit hundred and forty two thousand in first five months, ⁓ as hyperscalers commit, you know, some more. I think Wix cuts jobs. There's a bunch, you know, I think there's this is like a constant theme. ⁓ and so I'm curious from your standpoint, you know, ⁓ it feels like there is a lot of ⁓ or I guess like it feels like there's two camps. Is it really AI that is like forcing us to cut these jobs, or is it that the companies have overhired in the first place and that they're using this as a a way to, you know, trim back their teams and get back to maybe where they feel like they're actually capitally efficient. Jay Nathan: Ooh, do we have another 10 minutes for me to go off on this? I think ⁓ and I think this is probably coming to an end now, but every single one of these companies that have basically said AI is the reason, like AI efficiency is the reason that they're laying people off. It's all bullshit. A hundred percent. I the reason I like the clickup guy Zeb's posts. on why they're laying off is good because it's honest, which is like we're doing really well as a company. We we're, you know, we have great profit margins. We're actually going to we're gonna double down on AI in the future. Are but they're making a bet, a future bet, right? They're not saying AI has already provided these benefits to the business so that we can lay all these people off. That is not happening today, right? Everything we've just been talking about. Jeff Breunsbach: Yeah. Jay Nathan: I think we've all created more work for ourselves. There are there are there is there's more work to be done. Engineering hiring is up, by the way. The the biggest use case for AI platforms and tools has been coding. So if it were the if it were true that, and we've talked about this before, so I won't belabor it, but if it were true that AI is replacing jobs, those would be the first ones to go. They're not going right now. People who are not buying into this and who who are not getting up to speed on how to leverage these tools to make themselves at least 10X better, they're the ones that are gonna go. Right. And by the way, that's gonna apply to every, every, every role inside of a tech company. ⁓ but there are gonna be plenty of jobs. There are gonna be plenty of jobs, but you're gonna have to learn AI to hold. Jeff Breunsbach: Yeah. Yeah. ⁓ and I think that's like the I I guess like to me that's I I I think I largely fall in that camp too of like I think they're you know, they're using it as a nice wrapper of and ⁓ they feel like that's gonna give them almost a PR bump to talk about AI as you know, how driving their business efficiency, but ⁓ continues to feel like Jay Nathan: If you're using it for efficiency, though, it's the wrong answer, right? The efficiency is a is a race to the bottom. Right? It's a it's it's a race to an exit strategy, basically. It's not it's not the it's not the most valuable use case for for AI, in my opinion. It's innovation. Jeff Breunsbach: Yeah. Like you should be Yeah. But I think like the I guess maybe the the other point that I continue to see or that I believe in is that like every job is changing. Like don't just think that like it's you know, don't think that your job is safe or that don't even think that like your job is going away. It's just that every job is is functionally changing what's happening, right? Like even as a chief customer officer, right? Like the the landscape of like what you do has changed because now you've got access to these tools and therefore like It should change. Right, right. Like the when you see like a step function changes in technology, like it should change the type of work that's being done. And so like you should lean into that. You should figure out, okay, if like you said, if I kind of like take this to the nth degree, like what does my job look like and how do I start moving myself there? Right. Even if your company is not moving you there, can you move yourself there? Because that means you're gonna be better better positioned. And most likely that means like you're actually on the almost like cutting edge of what your team is doing if you're the one who's like moving down that path faster than anybody else. ⁓ and like Jay Nathan: Mm-hmm. Jeff Breunsbach: Ultimately that's going to be looked at as a good thing, right? Like if you were I mean, I I think one of the things that I'm I'm hiring for CSMs right now, and ⁓ I think, you know, I don't spend a large time of a a large amount of time in the interview on it, but I you know, you are sh you are correct in that I ask many questions around AI. Like, how do you use it? What do you do in your personal life with it? Like, what are the ways that you've thought about this? Have you you know, how do you stay up to news ⁓ on this? And because like to me, it it's become a part of all of us in technology and therefore like if you're telling me that you don't do anything with it, that you don't play with it in your personal life, that you don't read about it, like, I'm sorry, but like that's not the type of person that I I think I can that can be successful in in this business here because like I need people who are like continually thinking about the next step ahead and like are curious enough about it even in their current role to do it. Jay Nathan: Yep. Yeah, a hundred percent. I think the mo the highest paid role on any team is going to be the ops role that is basically the AI builder for that team. That enables everybody else to just do what they're really good at, you know, if they're not AI pilled like they are. Jeff Breunsbach: Yeah. So Yeah. Yeah. I think that's a ⁓ I think Maverick it like one of the topics for next week should this be this idea of almost like ⁓ pulling apart ops ops because I feel like it we've basically just shoved a number of things into ⁓ like what a C S ops team, what a Rev ops team does, like even a a COO at a business, right? But like if you start to pull the onion back on those things, I think it would be interesting to like look at okay, like, you know, how did how have these jobs changed, right? Because like you used to have a BI and analytics team. Like imagine what that team looks like now or how that team functions versus what they used to do, you know? ⁓ so ⁓ all right, let's ⁓ we'll we'll end it there. Good episode and ⁓ we'll see you next week and I'll have most likely another another child. Jay Nathan: Hi. Ha ha! Yeah, congratulations, man. Big news. All right. See ya. Jeff Breunsbach: ⁓ all right. We'll see you next week.