Integrative AI Where Human Intelligence Becomes the Killer App

(00:29) So uh you know seeing how they don't just build ad tech but they build community and uh and uh I've known known his team for quite some time. So it's been wonderful to have Coachava supporting and involved with March in a lot of ways. you uh you might have seen Charles colleague Trevor at Marcher live recently and uh and so so uh Charles and and his team CEO CTO Ethan is brilliant and just like amazing folks where like you see some of the companies I'm getting a lot of my knowledge from uh and uh and a company that's doing so

(01:09) much to not just lead but educate the industry. Um, enough about me talking. Everyone's here to learn from you, Charles. Well, first of all, thanks so much for for the invitation. Um, I'm a I'm a big fan of the community and Dave and I had the opportunity to meet many many years ago um at uh something I don't really recall.

(01:32) I don't remember what it was but it was from one of our team members and uh we were just really uh enjoyed the community and just loved um what it was all about just bringing marketing together in terms of actionable execution of AI and not not just powerpoints and concepts and so we're thrilled to be involved. So thanks for the invite. Yeah. Well, uh, excited to dive in because it it's just so funny because we were talking about what's happening with agents early in the year and and there weren't many examples, were there, right? Like of like deployed like for the ad industry uh and and now it seems like a few things have changed since then. For sure. Yeah. So, um, but I mean,

(02:12) obviously a number of things have changed and, um, I I I think there's some some really interesting themes about what we're working on, and I'll I'll back up and speak to what some of our drivers have been as a company. I think many of them will resonate with the people in the audience. Um but you know if you think about AI and how it has transformed not only the advertising industry but just computing in general and any kind of any kind of SAS technology out there.

(02:43) What you're seeing is a um a chat prompt that's getting added to a SAS dashboard UI and that's their kind of tick box of that's how they're incorporating AI in their SAS technology. And certainly there's a lot of value to that. It's like how do you create data to be conversational? It can be accessible.

(03:02) Um, it can be usable by analysts in ways that maybe traditional dashboards and reporting has not been. Um, but fundamentally that interaction, that conversational interaction is still siloed specific to that vendor or that specific tool or that specific SAS tool. So, you know, since November 20th, 2022, that's been the prevailing thesis of layering AI on top of an existing SAS offering.

(03:29) We've observed all of that and said and in fact David was at our summit last year. We had actually given a an early insight into what I'm going to share here that we've just opened up as a as um kind of private beta uh for which anyone here that's interested.

(03:48) I'm happy to engage with David and get invites out to the folks that are here if you're interested in having access to it. I'll walk through some of the attributes of what this thing is. But in our in our early execution of AI as a as a you know large SAS company that deals with lots of data, it was a prompt to SQL interface.

(04:08) Very common like how do you have a conversational prompt? It turns into something very unique as a SQL statement and it comes back out. The problem is that if you're an individual operator, if you're an individual that's part of a team, you're dealing with 30 of said SAS tools. You're not dealing on a one by one basis. And so we started to back up from that and said, what where is computing going to go for team members who want to leverage AI in a way that 10xes their productivity? M and the conclusion we reached was that there's kind of this new category opportunity called integrative AI.

(04:44) So you've got kind of ML and AI and generative AI. And you know I'm I'm certainly not suggesting that this category that I'm describing is well understood or known yet or that there's fine boundaries around it. But this is kind of our mission and our evangelistic exercise. We think integrative AI is really the value prop and the the attributes that we're establishing on what integrative AI is is variable model.

(05:13) So you should be able to work with any model you want and work with any model key that you want because many people work for companies that have very specific uh expectations around governance of what model you use and what key you use and what account you use. The second attribute is uh a templated pre-prompt library so that you can have consistent execution with those models across your teams.

(05:40) you don't run into the problem where you know Jane is a particularly capable individual that knows how to pre-prompt things and so always has better responses than John who just phones it in and it and it shows in this case where you have common pre-prompts you have a framework where you can start to share these things around workflows for team members and then the third element um third or fourth is connectivity that whatever you're doing with AI Instead of copying, pasting, importing, uh, and you know, doing all this kind of manual dragging of information, the the system needs to be able to fundamentally connect through authenticated tooling to

(06:19) the tools you already use. So you log into Tradeesk, awesome. You should have whatever this tool is authenticatable against Tradeesk. You use Kachava for measurement, awesome, that's another endpoint. you use Samba for, you know, linear uh measurement or targeting. Awesome. You use your credentiing for for Samba.

(06:43) So that the notion is is this composite or mosaic of connectivity that brings us together. And then finally, and the reason why this is the last one, and I'll and I'll pause for a moment, David, so that I'm not just talking, but um is this agentic piece. Now our thesis is not that we're going to live in a world tomorrow, you know, with this immediate transitory step where agents are going to just be working authenticating on their own doing tasks and activities um autonomously.

(07:07) We don't believe sociologically, we don't believe psych psychologically that is going to work organizationally. Instead, we think this notion of integrative AI is an extension of the human as a team member who can 10x their productivity so that the agents are performed or or created that they can get created by a noode analyst, a noode operator and that they fundamentally composite connectivity templating and models so that it starts to be an extension of what they're already doing in the workflows that they want um reduce the repetition on. So, I'll

(07:47) pause there, David. I know that was a mouthful. I hope it's okay. These terrific and and I love these four parts to it. One thing I'm also curious on the connectivity side that also might apply to some people in this room or who watch after is is do these other players need to do something themselves to enable that or is this something that the system should just be able to do with a standard login? Yeah, so um a awesome question.

(08:20) I'm in a I'm a much more visual person than I am anything else. Do you mind if I share and I Please do. Yeah, I'm fair. Well, I'll love that. Yeah, we took a different approach. So, you know, I'm I'm um I've been through and worked in the client server computing age and then the web age, the web 1.0 age, web 2, um you know, full SAS.

(08:46) what we're doing with this product called station one which you know our focus is to focus on this category of integrative AI this is a downloadable app so this is like old school this is not a SAS tool and we did that intentionally because of the privacy and kind of data leakage concerns that we're finding and we're hearing about from our customers about um not wanting anything related to AI to proxy through another SAS system because there's a concern of what that means from a carriage perspective of the access to that data, access to those APIs, access

(09:28) to the models, etc. And so we flipped the model entirely. We said, "Okay, we're going to build a Slack-like downloadable app that has workspaces just like Slack." And those workspaces can segment buckets of connectivity, templated pre-prompts, and model access and allow for the productivity of individuals to be extended and multiplied because they've got this kind of universal AI hub.

(10:00) And so as as an example, I've got a um I've got a workspace that I can use um you know, OpenAI as my designated model, but I can also use a local model. I can use an internal model uh against my um backend data center if if I've got a team that has built a small language model that's specific to my business or my industry that's not even loaded in one of the foundation model providers. I can use a a Google model and anthropic model.

(10:34) Effectively, I can put in keys for any one of the models that exist out there. And now all of a sudden within this downloadable environment I now have access like this universal mosaic system of of uh connecting models and then again based on which um which workspace I'm in I may have different pre-prompt templates.

(11:00) So for example, I'm in a product management workspace that we've been using internally is like a almost eat your own dog food approach to if our product management team can consistently spec out requirements documents and sprint planning documents, then we have something super interesting because that's a great example where you've got a team of PE people working against a common objective with common artifacts. And now all of a sudden you can use consistent tools around it.

(11:35) So this is a totally unrelated to ads but it's our kind of product management workspace that has things that help you build requirements documents. Separate from that we have this sample workspace that for any of you that participate in the data want to get an invite you'll get access to this.

(11:54) We have a effectively like a gallery or a marketplace of workspaces that are published publicly. There's only one right now, just so you guys know. But in the future, there's going to be more. There's going to be a whole bunch that vendors want to publish these workspaces. And what we have in this this workspace is like everything you need for omni channel advertising. And the reason is that's the business that we're in.

(12:12) You know, we're we're in the business of buying media across a whole bunch of channels. And so the experts that we framed were industrypecific um experts around monetization around how do you translate your your campaign execution for a CFO? It's about um things like uh you know we're really good at mobile but we don't know anything about digital out of home and CTV.

(12:43) How can I have an assistant that's helpful to our team as we think about extending our media plans across these other channels? How should we think about measurement? So, there's continuity and consistency as I do activation. Um, and then kind of your standard media planning and activation. Now, these these experts that we built are pages and pages of experts. Like, it's it's a lot of content that pre-primes all this uh historical knowledge that we have around the advertising space. Nothing stopping anyone from saying that's cute, Charles.

(13:13) I'm going to create my own workspace for us as an agency or us as a brand and I like what you've done, but I'm going to morph this and I'm going to change it and I'm going to use my set of pre-prompts because they're specific to my team's execution. Does that make sense on those pre-prompts? Mhm.

(13:33) So, um that's an example of our our our integrative um advertising uh pre-prompt. And with each one of these workspaces, you have this notion of connectors. And these connectors are uh for those of you that have been following, these are effectively curated authenticated MCP flows. So MCP is the model context protocol originally invented by anthropic but being used very very broadly and and extensively.

(14:05) One of the problems with MCP is that it's a great protocol and everyone's um adopting it, but how you authenticate with MCP is still a bit of a a nightmare. And so what we did was we built a gallery of MCP servers that are be behind common authentication OOTH frameworks so that there's pointclick two-factor authentication. Done.

(14:31) It's integrated. And if you want to live on the wild side and you want to connect your own manual MCP uh directly, you can, you know, have an STDIO uh MCP, a streamable HTTP MCP, and you can set up authentication. You can still do that if you want to be a little bit more technical, but we we're building this kind of gallery of of connectors that are specific to what customers want.

(14:58) Now, this is just a small subset. We're going to have Trade Desk, AppLin, uh Amazon, uh you know, AM, uh AMS, the Amazon marketing, uh cloud, AMC. Uh we're going to have all the buying tools here. Uh but we're also going to have a whole bunch of measurement tools. So like we as an example um you know we're we're an omni channel measurement company for sure but it doesn't mean that we won't have other um vendors who are in the um category of measurement exist here because otherwise there's kind of no point. So one of our competitors on the

(15:37) advertiser side within the mobile measurement space is a company called Appsflyer. We're integrating the Appsfire MCP server directly. You can see all the tooling that's available. Um, you authenticate and now all of a sudden Appsfire is part of the picture. And so you can have conversational engagement across multiple tools where they're being synthesized together with common experts and pre-prompts for consistency.

(16:08) connectivity is an extension to the person and then um you know variable model support. Does that make sense? Uh it does. Uh and if there are questions from the community uh feel free to chime in and chat or share your questions but uh yeah I mean yeah please uh so quick question for you. This sounds really great. Uh Charles, thank you for breaking it down for us.

(16:32) Quick question. So on the client side for people using the tool, who's going to be the person at the company responsible for the governance and making sure that the data model is aligned so that the platform can leverage the insights from all the different integrations that you have. Yeah, awesome question.

(16:52) Um that is I think the uh long pole in the tent. Um so I'll I'll tell you what's happening as we engage with our customers. I don't know that that's prescriptive and that what's going to happen everywhere. Obviously, this is all fresh ground that we're all tilling together organizationally. Absolutely.

(17:16) But we we think that the value of station one and being a downloadable client is that it's not proxying um you know all this through our SAS system. So um uh it's a direct connection. Having said that, when corporate entities, when organizations deem that a client tool or even a serverside tool is acceptable and usable, uh there's usually a checklist and we're certainly seeing that across the AI category.

(17:42) Our our customers are typically larger um you know, larger brands. We have 20 of the top 25 streaming media companies that use Coachava. We've got a lot of large QSRs. So they don't, you know, willy-nilly let organizations within their company just start installing software and connecting AI models.

(18:02) And so what we're observing is like a a check list, an AI SWAT team checklist because there's such an influx of tools that are coming in and we're we're getting better at um how do we prepare ourselves for answering those just like we would another DPA or any other kind of support elements. I think you expect that. what we're what I'm obser what I'm learning.

(18:21) So the former is what I'm expecting what I'm learning is there's interest in things like traceability and governance. So organizations saying I love that you can set up variable models for anyone that is on our domain. I don't want them to connect any model. I want to designate server side what model they're allowed to talk to and limit it at that.

(18:50) makes tons of sense because they want to make sure if it's a company asset and it's company exercises, it's limited to just the models they're deeming appropriate. Um the second big thing is uh traceability for um you know what devices are connecting agentically versus from a human perspective. So as an example um Salesforce when I'm connecting doesn't know that I'm connecting agentically. It just thinks of this as another session within Salesforce.

(19:20) The same with Slack. And so what we're observing is this kind of interest level of how are you a enabling individuals to proxy their um you know their arbbacks their um role-based access control um exposure to agentic interfaces because that's effectively what we're doing is we're bringing that to bear for your tooling. Does that answer your question bro? Yes and no. I'll be honest.

(19:47) Uh I think the question I have for you is just understanding the the data models as somebody who's been in the industry for a while now. One of the biggest challenges I've seen whether working in different platforms is that they all have different metrics and dimensions that need to be stable within an organizations especially with clients that are retail for example.

(20:05) So my question was really trying to understand from a client side perspective who would be in charge of that integration because yes logically that makes sense but I'm talking about terms of the data model making sure that the data that all the platforms are sharing makes sense given the needs of the request or the prompter.

(20:23) Yeah I now understand better. I'm sorry. Um it's an awesome question. I'm going to rephrase that question into something we're seeing as well. Um, if an agency is using Kochava as a measurement technology for some some clients and apps flyer as a measurement technology for other clients, how do I create consistent workflows irrespective of which tool they use so we can normalize the data models between them as as one example or if I've got uh data models from multiple tools. All of that is going to show up in how the MCP

(21:00) connectors are built and merchandised. And for those of you that don't know, and there's nothing wrong with not knowing this, this is like one of the most interesting things about I think this notion of integrative AI is that these MCPs are um connections that expose these tools with these text descriptions that kind of merchandise the tool.

(21:30) Okay. And so what happens from a flow perspective is if I'm prompting and I'm I'm interacting with a with a model and I say um and I can I'll go through an example. I say you know I've been busy in a meeting for the last 27 minutes. Please summarize my emails and prioritize what to focus on first. The fact that I said summarize my emails prompts the inference engine to call the Gmail connector and query Google's mail.

(22:05) Your example is let's say I've got a ERP system or you know two different systems that have unnormalized data models. Not only you you don't have to change the source system but the MCP connectors and h and the tool calling and the normalization of data that's output and the way in which those tools are merchandised with this text that will be where your boulder will be to turn into rocks as you bring those pieces together. Got it. Okay.

(22:33) Organizationally what we're doing um we have a we have a um let me see if I can find it. So we we have our own you know measurement um MCP as you can imagine and um when I look at them um let me just see if I can find one where I've got it connected. So what's interesting about um this is that this is our customerf facing MCP set of tools and it's things like getting attribution results on apps you know run search queries you know all the things you'd expect you know lifetime value details settings for apps

(23:17) and um separate from that we have customer success individuals on our team who are managing portfol portfolios of customers, right? So this MCP is a onesie twzy MCP. Um it's so that a singular account can talk to their singular account in Coachaba. We have a separate MCP for our CSMs that are looking at a bulk of a whole bunch of customers across a portfolio of their customer base and looking at settings across a broader set.

(23:52) And we ourselves here we are you know uh plowing this new fresh ground really ran into some organizational questions of do we designate a part of the engineering team for MCPs separate from our product team or are they an extension of it because the needs of how these things are going to show up agentically are going to um they're not really affecting the engineering team of station one they're affect affecting the product team on how they expose that data model to use um the the summary you you provided. Yeah, that makes sense. I may have missed it in the demo just because I

(24:31) know we're using a demo account right now, but I'm curious have there been is there plans on your product road map to integrate tools like u uh CMS or CDPs cuz those are trying to you know that's what most clients I found have been using for their customer whether it's B2B or TOC and so I'm curious any integrations with those kind of platforms 100%.

(25:01) So not only are I mean one of the great things about standardizing around MCP is that I think every vendor is already thinking about how they build an MCP that's like their product supported MCP for their product. companies like Tradeesk, they don't have an MCP server, but we're building one around their APIs and um so it it it the you know CDPs are in that category as well of folks that um we're going to be building um uh MCP integrations around as well as CMS.

(25:27) That's that's uh two very important categories. Awesome. Thank you. Yeah. Um and uh we have a question that came in the chat and then uh and then I'll go to Adam who's also raising his hand. But uh Ory is asking uh could you elaborate on how this is different from other agentic infras do you mitigate the risks when connecting to an external LM via an API key? Yeah. U good good question.

(25:55) So how it's different you know this this approach is really an extension of the individual as opposed to a server side amorphous um agentic engine where you give it instructions and you hope for the best in terms of governance rules because of pre-prompt and prompt u elements. So I I think of this and I think of integrative AI generally as like a tooling mechanism where it's my common interface to increase my productivity.

(26:26) It's certainly how I'm using it. What I what I foresee is turning some of the workflows and the tasks that I perform into repeatable efforts that I no longer have to manually do each and every time. And so I I haven't shown you this uh part but I'll I'll give you an example. I have a um you know agent forge which is where you build agents um is a place where you can create an agent. You're not coding anything.

(27:01) It's not like n where you're like dragging and dropping and then check you know picking modules and doing all these things. You're literally describing what you want it to do. Arguably, you'll have already done this probably a number of times in the conversational UI and now you've decided you want this to be repetitive. And I'll give you an example.

(27:25) I've got a personal dossier agent and um I'm sure like many of you um you find yourselves in a meeting with a person that you've met before, but you don't quite remember the last time you talked to them. And you also don't quite remember who on your team has talked to them and you don't quite remember all the context of what you want to make sure you nail like if there was like any leftover things that were not addressed.

(27:58) Um and so I put this together and I described what I wanted. um you know perform thorough research on individuals by name, email, providing actionable intelligence and context for meetings with a special focus on coach related business opportunities and it generated this monster which is the pre-prompt and then that monster generated this workflow of the tooling.

(28:24) So again, I didn't I didn't create all these steps. The system created these steps. And so the first step is that I'm going to put in a name and an email address. And given that name and email address, it's going to carefully extract that person's name and um email provided and determine the company domain to do as a hypothetical, you know, employer based on that domain name.

(28:52) It's going to then search all my email archives for any interactions I've had with that person. and it's going to search Slack messages in case they've ever been talked about within our company. It's going to search Salesforce records to see if there's opportunities that are open and related parties so I can know who I can reference and namerop.

(29:12) And it's going to conduct comprehensive web searches that are just kind of public in nature. When that runs and I can run it on demand or I can have it on a schedule. I can actually tie it to my calendar so that it generates it in advance and that runs. Um, it then has an output and I've got an example. This is a partner of ours, U Samba. I used it as a demo. This takes I don't know a minute and a half.

(29:39) I don't want to waste everyone's time and watch a a pulsing uh progress bar. But background on the uh chief commercial officer um that we have um some a lot of internal and external interactions. There's a whole bunch of different opportunities that we have from partnering with them. Um we have uh some press activities that they've been doing. Um and this was just before you know just after the ADC CP announcement which is really MCP for advertising.

(30:03) Um potential talking points, value proposition, other queries etc. So totally unrelated to advertising, but an example of how agents are approachable and they're not just this amorphous thing that's happening out in the space. It's like an extension of your productivity. Terrific. Thanks, Charles. Adam, thanks for your patience. Yeah. Hey, Charles.

(30:27) Uh, thanks for doing this. Um, a former very early customer of Kachava when I was running partnerships at Jumptap way. Oh my goodness. Juptap was our third network integration in our history as a company. It's great to see you. Thank you. Yeah. And and at the time uh we had already been working with so many others.

(30:49) We're like too crowded, don't need it, whatever. And then you guys came in and just blew everyone away and we leaned into it big time. So thank you for that. Thank you for sticking with the industry for so long. Um and a you know good buddy of Garrett McDonald and Grant Cohen. So good stuff. Um thank you. So, this is getting into the cool new territory.

(31:08) I may have missed it, but you know, if people are using a Looker Studio, Looker Studio, a funnel.io, more sophisticated like integrated reporting, are those just API hooks, another integration like anything else? Um, is there a more sophisticated Q&A or NCP kind of like how do you think about because a lot of times what what I'm doing as I'm working with senior leadership or board and trying to explain the eb and flow of where you know the media is going.

(31:42) So, anyhow, that's a Yeah. So, I think what you're hitting on is like the the the progressive um maturity curve of some of these tools and I think you're spot on. No one's ever accused me of being mature. Charles, keep going. And but I think that's happening. Um so, there's going to be basic integrations that are just getters and setters. Yeah.

(32:08) I would say 90% of the activities that we talk to partners, candidate partners that we want to build uh connectors with, they only want getters. They're scared of setters. They don't want the thing to be able to do anything. They just want to be able to view things. Now, that's progressing and they want to do both soon. And um in particular those companies that have APIs where you can actually affect change they seem to be comfortable because this is just an extension of affecting that change.

(32:37) So adding postbacks or enriching this data through my CDP or a looker merging a bunch of different uh views together to get um you know unified output. Um so that's happening. I think as you as you run the tape fast forward I think two things are going to happen over the next 24 months. uh two specific things around your question.

(33:00) One is these MCPs are going to get really interestingly um sophisticated and they're going to self-describe their sophistication in that merchandising of the tools better and better and things are going to work better just because they're describing their tooling better because the inference engine can now make more sense of how the tools start to merge and progress. That's the first thing that's happening.

(33:22) The second thing that I think is going to happen is companies are going to get into this category. They're going to think about how they build their own GUF models. So these are compiled downloadable models that look like an open-source model, but they're internal to their proprietary data. Those that's like the other end of the spectrum. Super sophisticated.

(33:49) Imagine a world where um you're not interacting with OpenAI's chat GPT4.1, you're interacting with company XYZ, my internal company CRM SLM, you know, small language model. Yeah, I think and and this supports that as well. So, I think both of those things are going to end up happening. Um, I guess there's one other one other movement I think is going to happen and I'm hearing a lot about this just in the last week is this notion of workflow steps as an extension of human workflows.

(34:26) So like when I do a a campaign um plan and and I'll just give you an example. I have a digital I have a streaming media app that I uh that is supported across Roku, Samsung, LG, and Vizio. Um sorry, I'm launching in 2026. I need a holistic um uh media plan. I use my expert of media planning and activation which is like reams and reams of best practices and avails around talking about strategy versus tactics around the things that are important for measurement around how re-engagement is as important as new acquisition as you know every related

(35:22) just just quick real time reaction and I think you may have touched on this but if you're talking about avails and things like that. It's great to have this like sort of abstract project plan or task list or check or something. The ability obviously to sort of go and pull. Yes. And those that you're hitting right, you're the straight man.

(35:42) You're you're hitting right on my point that there's going to be a new class of MCPs that are the avails. And then there will be another class of MCPs. It's like, okay, if these are the avails, I've got a I've got a media plan. I want to lock those avails because of my contracts that I have in existence so that I can now progress that to the next workflow step.

(36:09) I think the next level of maturity is where it's actually stepping through the workflow. It's not just a pretty word doc that I can present to someone. So when we talk about configuring inapp events for example, it's not just talking about configuring the inapp events. It's confirming that my inapp events in fact are configured, that my trackability is configured. Um I'll give you I'll give you another example.

(36:32) Um so I've got a u an example with um universal ads. So we've got an integration universal ads. I need to make sure that all of my tracking links. How often have you guys trafficked a campaign that's sizable and spendy and you don't have your impression verification tag set up correctly and you realize it 30 days later when you're reporting is awful. That's an example.

(37:00) Um or same thing with outcomes measurement that for folks like Coachava and our customers I need to make sure that all my tracking links are configured across my campaigns and bveral ads and you confirm. So the multi-step looking at active campaigns um qualifying tracking links do they exist or not and everything at least in this phase of the maturity curve is confirmationally based.

(37:35) So without calling it human in the loop and confusing everyone and scaring everyone that there's like this human it's just conversational. So um you know I'm going to for the sake of discussion say make your best assessment each review prior to making change. So it gives me my summary where they're configured where it's not applicable um drill in and follow through. Right.

(38:04) So you can't do that without consistency of the prompts and then also the tooling as an extension of the person. We're not saying now go make this a magical faceless agent that runs in the ether. This is an extension of you. Um, one last thing cuz it's kind of fun and we just added this feature in the end and um, we call it operators, but this is kind of the closest thing to what people are thinking about as agents that they don't create, but in fact they engage with almost like personal coaches or personal assistants.

(38:36) So, we have like these these five out of the box ones in this integrative advertising workspace. The cool thing is you can create your own operators and I'll show you that how possible that is. But Selkerk and Monarch, those are the names of the mountain ranges around where our headquarters is located and we just call them these CSMs.

(39:00) One for our advertiser, our brand products and one for our publisher products. But then we've got this like inspired by Mark Pritchard. How do I translate my performance results so that my CFO loves me? And so it's like a p these operators are like these persona differences that are coaches and helping you make sense. Um Frank is a you know virtual consulting services member of our team that helps you make sense of the data that's in Kachava in a way that our consulting services team would look at that and um Allison is an expert in MM and starts to think about how incremental lift is different than attribution gives you

(39:40) pointers on things you can do. you can just as easily with station one create these other operators and I'll I'll give you an example. Um you can have an operator that's like we'll call him David. Um, you're a podcast and community leader that speaks to the issues of AI and marketing helps audiences understand how to be more productive.

(40:20) Please refer to David Bowitz architecture and the AI marketers guild as a reference. love being a reference. That's not an awesome pre-prompt for those of you that have futed around with pre-prompts a lot. And so, we have a tool that helps you build the prompt for you if you're not all that good um and is a little bit more thoughtful.

(40:44) I can then apply my expertise that I've set up in Station One for that workspace. So, if you've done lots and lots of work to like set up expertise, you can do that. And you can also limit the tools. So like let's say for example in this operator I want to link it to um you know the example before of this looker instance or you know this trade desk instance maybe this is my trade desk professional and it's only allowed to talk to trade desk I can hit save um and I can close this and um um it it's going to introduce himself um what is the best um set of topics I can consider for

(41:28) December given my load of topics over the last six months. I don't know. I mean, I wasn't even planning to give this as a demo, which is probably the world's most dangerous thing to do. But um you know, if you think about your own job, I'm thinking about David's job right now, like how do you have a built-in coach that helps you do the things you want to do with these templated objectives so that as guard rails about your success and that's really the idea behind integrative AI.

(42:00) This is amazing. Are these topics any good, David, or is this all nonsense? Th this is me. This is what what I'm I'm done. That's that's it. I won't even be here next week. Sweet. This this really good. I got I've got I've got a whole next book written. Well, we'll we'll work with David.

(42:25) I won't have any of you guys' emails, but we'll work with David and let him invite you guys as audiences into the beta and um um for those of you that are interested and and um we'll go from there. Yeah. Yeah. And all the the follow-ups. Amazing. Thank you for sharing this. Awesome. Awesome. Yeah. Yeah. In the in the home stretch here, what are the questions uh do others have here? What can our resident expert Charles answer for you? How did you address the the people aspect of this um in in Kachava and your organization? I feel like that is such a fun question.

(43:04) I wish we had another whole hour. I don't have finality and conclusion on that topic. I feel like we're still mid-process. I think it's such an awesome question. We I I I visualized it kind of like um a gardener hype cycle, like a miniature hype cycle in the context of adoption within the organization.

(43:29) What I didn't want to do is to add any more work on the people who already had lots to do. Um we we segmented out a different team. they specifically focused on um building this thing and we had a vision and a conviction and we did it. But what was really fascinating was as we started to roll this out, you had enthusiasts who really understood the value and they immediately like gravitational pole started to play around with it and they started doing really interesting things.

(44:03) And then you had people who didn't even register um and you know when I say register it's like you know log into the system like even when it was just an internal tool and it was like I I don't I don't see where that's going to change my life. Um, and over time what we've started to do is think about how do we build features that facilitate social adoption within teams? Because what we discovered, even though the tooling we have today is not very good for this topic, um what we discovered was that when we could socially share the artifacts,

(44:45) people were like, "Holy crap, wait, you just did that. Oh, I see. Let me connect my stuff." And I'll I'll give you a really I'll give you a really good example. Um, so like so your early adopters brought the rest of the moths to the light bulb. Yes. But it wasn't very effective uh without sharing tools.

(45:17) In fact, one of the biggest questions we were asked internally was, isn't there a feature where I can share an expert without sharing a whole workspace? So people loved this expert and then they wanted to tweak it and extend it almost like they wanted to version control the expert and like extend it have it do more things. Um we um and and so social sharing of experts of workspaces and of artifacts were the three things that I think have u been an asked for element.

(45:53) But the the the human the human component we think is the real long pole in the tent here. We need to make this tool approachable like beyond recognition because this is not about will it make you more productive. It's about will you believe it will help you as opposed to believe it will hinder your your work or change the way you work. Girls, I want to jump in with a question.

(46:17) Um, imagine I'm on a college campus and a fold out table with a sign that says, "Change my mind." I love it. The analogy is the Excel spreadsheet. I have a bunch of people who are really smart. They've created their own spreadsheets to do their job. One of them leaves and a new person comes in and looks at the spreadsheets and goes, "I I don't know how this was built. I don't know what these formulas mean. I don't know how to edit it.

(46:35) I need to create my own." Yeah. And And that person who left thinks they can take the spreadsheet with them and apply it to a different job. Yeah. How how are these things universal? Yeah, I I think that very much relates to the social sharing bits. So, I've um there's a whole community, if you guys aren't familiar, I'll tell you there's a whole community of people who um if you put in your email address, they'll share with you a Google doc of pre-prompts for different industry vectors, you know, segments. And then there's another whole community that takes that to the next level. If you pay them $25, they'll

(47:15) share with you an even bigger library of pre-prompts. And the way in which you pay the $25 is like an online class system. I think there's a content creator marketplace opportunity where if we do our job right, you can have people who really understand their business well, just like you described the individual who knows those macros in Excel and they can share either experts in a gallery of experts or share their workspace to a gallery of people who want to download those workspaces. And if and our job should be how do we facilitate

(47:54) helping them make money on sharing that content almost like a marketplace today. We're not worried about that because we just want it to work really well. But the thing I continue to ask myself is the same the kind of the question you're you're you're posing. Do you get faster adoption if content creators have a way in which they can make money on that awesome Excel macro they produced? And my question is, how do we know that that macro will apply across an industry? Because it feels like it's very personalized to a job, a task in a job

(48:31) in a company and an industry. I I think it is too. I mean, the idea here is that these are either tool specific or they're company specific. And if you've you've heard of this phrase like forward deployed engineers, FDEES, these are the people who like want to understand your workflows and then orchestrate your AI to work with your your workflows.

(48:56) You know, nothing is uh more challenging for corporate adoption than having a bunch of consultants tell you what your workflow should change into. We want to flip that where it's like we're extending that so that this is just helping you with your micro teams and then maybe over time it starts to get syndicated with larger teams. Very helpful.

(49:13) Thank you Rory. Go ahead. Hi there. Uh thank you. This is really wonderful seeing uh station one in action. And uh my question centers around local LLMs because the company I work for we have everything on prem and uh we self-host um you know eating our own dog food everything um using open-source technology and I saw that you had uh yes as one of yours and I'm wondering what sort of success you had in with your alpha testers for local LLMs.

(49:47) Yeah. So we support a llama lm studio. Um there's a few other kind of run times for local models. And for those of you that are on the audience that don't know what this means, it's like these are um physical um you know local open source models that have been compiled. Hugging face is a great destination if you've never played around with it with just oodles and oodles of models.

(50:12) Some of which are awful, some of which are kind of interesting. and sounds like Rory's doing a bunch of stuff which is cool around that area. So, we've like used the the Llama 3X uh models for some testing. We have uh some internal um Nvidia hardware just for some testing. Um but we're trying to optimize towards online models because our observation is that despite all the kind of governance reservation and concern around model access and trainability, that concern is being prosecuted by legal teams telling Open AI and Enthropic, you can't train off of our data, but we're going to still use your online models. I think

(50:57) over time they're going to want to do stuff that's on prem and um there's just such a demand for hardware that it's super expensive and I think that'll go down over time and I think I think we're going to see in our in all of all of industries more decentralized distributed models not these centralized online models and that's why we're taking the approach that we're taking.

(51:24) We think over time there's a myriad of models that start to come together across workflows and testing is awesome. Like um depending on how you've juiced your hardware, it it's faster than I mean like on a on a on a Mac M3 you can run Olama locally with GPTOSS and it works just as fast as chat GPT. It just has different responses because it's a different parameter count. It's juiced differently in terms of pre-prompts.

(51:52) Yeah. Wonderful. Thank you. Um Charles, we could easily go another session or two here as you mentioned, but uh sadly we've got the hour for now. We'd love to have you back sometime. We know that you're all uh you know, we know that you're all active with the community and uh know how to find us.

(52:18) I'll make sure to help get the invites out to everyone and uh and get anyone access who wants to dive way deeper into this. So, always appreciate that. Awesome. Well, Sam, thank you, David. Thanks for your community. I know you do recordings and I'm sure, you know, lots of other audiences that will hear this and for those that that are interested, we're happy to be supportive and helpful.

(52:35) We we just think it's a really cool um it's a really cool time where so much is changing and it's really fun to be part of that authorship with with you guys. Well well yeah great to have these conversations with folks who are actively building all this and uh showing us the various stages of progress too so we can track it along the way. So it's so helpful I I know to so many of us here. So uh thank you and to your team for setting this up.

(52:59) Thank you.