I didn't write a single line of code. You don't need to be an engineer. You just need to be able to capably communicate your vision.
SPEAKER_01What's one thing that you have been able to hand off to generative AI that has been a huge win for you?
SPEAKER_00My favorite app that I built, I call it the Pocket Sales Engineer. I built an application that listens to calls in real time. When it detects a question, it sends that question over to a knowledge base and hunts for the answer to that. It's really cool because it turns our salespeople into weaponized sales engineers. You've made a lot of mistakes. Pick one that happened with this business getting started. You know, I think we took a little bit longer than I felt comfortable with to adopt modern AI coding tools.
SPEAKER_01But you mentioned something. You said that it's almost contraindicative to their terms of service for them to keep your stuff private. What do you mean by that?
SPEAKER_00I'll give you three. One covert way that they're using your data, one over way that they're using your data, and even one nefarious way in which these organizations are using your data. We'll start with the covert.
SPEAKER_01Hey everybody, welcome to another episode of Using AI at Work. My name is Chris Daigle, and I'm the host of the show. And before we get started, I want to mention that some of you have uh sent me an email that the opportunity to uh work with our team on identifying some $50,000 opportunity we call that exercise in our ads, the link was dead. So I've now fixed that link. You should be good to go. And today, you know, uh we talk to executives all the time. I think at this point we've trained over 18,000 non-technical business professionals over the past three and a half years that we've been doing this at Chief AI officer and through the podcast and everything like that. And there's always a few top concerns from executives. And one of them that I hear regularly is we'd love to use AI more, but we don't quite understand the risk. And, you know, they don't necessarily just want to turn their teams loose with ChatGPT or a Claude license, although they know they need to do something. The time to sit on the fence or to watch and see if this stuff is working, that's over. And I talk to them, and most companies are thinking that the AI decision is simple. We'll pick some tools, we'll block some tools, but today's guest actually sees a third path. And that's giving people the room to use AI, but putting real controls around the data. And that's what most executives are looking for. So today's episode is going to help your company decide how to introduce the tools to your teams with the safest governance protocol in place. Our guest today is Steven Walchak, the founder of liminal.ai, and this is what they do. Um, Stephen started as a software engineer, then spent roughly a decade without coding. He had some successful exits. But modern AI has pulled him back into building because he knew that he could create the applications without personally writing the code, including the infrastructure heavy work that we've talked about in our pre-interview. And he's now leading liminal uh with some known brands that you would uh recognize for sure on how they're using AI adoption through the lens of security and governance. So, Stephen, before we get started, anything that you think the audience needs to know about what we're going to talk about today?
SPEAKER_00No, I think you covered off the notion that you can actually adopt these tools in a world where the providers themselves are not aligned with how you think your data should be treated, um, where the incentive is actually stacked against data privacy in this case, uh, where they are categorically ready to to abuse in many circumstances their terms of service in order to uh to achieve the outcomes that they're looking for. Yeah, I think there's a lot of questions that need to be answered. And I think you've you kind of hit it on the head, how do we do this in a way that's safe, secure? Because these tools aren't going away. And in fact, they're so valuable. Uh we we certainly believe in an optimistic promise of AI, and we've seen the the value can deliver for teams that how do you deliver the how do you how do you achieve that value without actually uh um having to bury your head in the sand and say that we'll do we're never gonna deploy these things internally. So uh that's that's the goal for today, and I'm looking forward to having the conversation.
SPEAKER_01But you mentioned something, you said that it's it's almost uh contraindicative to their terms of service for them to keep your stuff private. What do you mean by that?
SPEAKER_00Yeah, so uh we can give a couple of examples over the last few years, and I'll give you I'll give you three. One covert way that they're using your data, one overt way that they're using your data, and even one nefarious way in which these organizations are using your data. Um we'll go with the we'll we'll start with the covert. If you guys recall last year, um maybe some of you listeners will recall, but many of you probably won't, because it wasn't well broadcast that OpenAI was required by the U.S. government, uh, federal judge required them to indefinitely retain and log all data both through their APIs and through their end user tools. Um, this was, of course, due to the fact that they were being sued by the New York Times, and it was on behalf of that case that the judge ordered this, but here you are now having attorneys getting access to privileged information, um, your information, um, all data. It didn't matter if you were an enterprise and sign an agreement or not. Again, a third-party judge coming in and saying that this needs to be retained. Um, and when you consider the fact that there were 1,400 plus breaches last year of the top five model providers alone. Ouch. Uh that data is now subject to breach. So that's that's a starting point, right? That's the the covert um where they they didn't exactly announce that publicly, that all your data is now subject to uh somebody looking for looking through it in some um some sort of uh legal legal case. Then there's this the the the the overt ways. We had Anthropic last year updated their terms of service that if you were under the vast majority of their licensing, that you were now uh subject to a zero data retention. But in fact, uh not sorry, you were you were not subject or not allowed to have zero data retention, but in fact, that they were for up to three to five years storing your data, retaining it on site, um, and in many cases they would be using your data, and they explicitly said this to train their next generation models. Um so while I applaud Anthropic for being completely overt about their usage of your data, that that wasn't the initial terms of service when people signed up. So that they have the ability to say, listen, you don't have another tool. You're you you do or you don't. But the reality is all of these major model providers are incentivized to get to the next generation of models to continue to put out the best. And the the best way to do that is through data, data, power, and compute. And with the power and compute components being solved for, well, they need to get more data. And so where do they get that? Well, why do you think ChatGPT and and is is free for their level licenses? Why do you think Anthropic's free? And you guys know when when the product is free, you are the product. Um so I I think that that's the the incentive is stacked in their favor. It's a trillion dollar pot of gold at the end of that rain, but of course they're gonna do everything they can to get there. And then let's talk about the nefarious. You have companies like Google, right? They came out and um for 20 years swore up and down that they weren't using their Android users' data for any purpose whatsoever. And in fact, we're sued and lost a class action lawsuit last year. $440 million that they had to pay out to their Android users is nothing, by the way, in the in the grant scheme for Google. So that that's the that's the dilemma these these these folks are facing. It's real. Um the incentive is not uh not certainly not stacked in our favor. But but the tools they're providing are amazing. So I'm I like I don't want to beat them up. Like they actually are providing a very useful tool. So how do you how do you work in coordination with that where you know you don't want your data shared with them, um, where there are in many cases significant regulatory penalties for your data being shared with them. How are you equipping your organizations to do this? And that's kind of the problem that we we set out to solve for at Liminal.
SPEAKER_01So uh explain what Liminal is and what it does.
SPEAKER_00Aaron Powell I'll be really simple. Um because I we don't need to belabor this point. I think we can just jump into the why and the there's certainly great stories around this stuff. But uh I think the the best way to describe what we do, and very simply, um we provide an end user interface so that users can select from multiple model providers. And to the end users, we provide unlimited access to the latest and greatest models from the top five providers. Um actually top six. We added Meta this week too. Um Spark models are actually pretty cool. Um so you have unlimited use of the top six model providers, um, OpenAI, Anthropic, Google, Perplexity, Grok, and now Meta as well, all for a very low cost. So that's for the end users. Um for the administrators of these accounts, they have complete and total governance and control over how data is treated before it ever leaves their network boundary. And because what the way that we're deployed, which is called single tenancy, which just means that we are acting as an extension of their network boundary, all of their data is contained. And they get to choose from you know, uh as granular as how a name is treated, up to how documentation, how MCP calls are made, um, how that data is treated before it ever reaches an LLM. So their end users are able to use these tools in an interface that is extremely familiar to them that feels really neat. Um I we continually hear delightful stories from our end users, but the administrators crucially get the governance that they need, the tools to control how data is treated, um, they get the observability that they need so they can see what people are putting into the uh into these LLMs. Um, and crucially they have an auditability, uh they have auditability within that. So end users get a delightful way to engage with models and the familiar voice and chat experiences that they're looking for. They get unlimited usage of the top six model providers. Administrators get governance and control over how data is treated and they get tracking and observability into what's going into these LLMs. That all at a cost that is typically 30 to 60% cheaper than single model provider subscriptions. So we're able to do everything there in a really cost-effective way.
SPEAKER_01Okay, so for the listener, I want to give you an example. We were on site with a client doing a one-day training that we do, and we make it very clear with them hey, make sure that you're using the company account. So look in the bottom left, they were on Chat GPT, look in the bottom left, make sure that it's showing the company account, not your personal account. We made that very clear. Why, if somebody is using their personal account, we don't know their security config. But also, when they leave, all of that information that they've done in their personal account leaves with them. That's a no-no. So I was walking around the room, we were doing an exercise, and one of the senior execs, even after we had talked about this, I looked and I was like, hey, buddy, you're in your personal account doing all this stuff. So I know that those things are happening. We we we see it happening. But for the listener, what Steven's saying is that before you you got somebody, you just did the training, they've got their license, they're excited to use it. They kind of remember maybe what we talked about in the AI use policy, but they enter some stuff into the chat interface and they hit enter. Now, in that case, if you're a liminal, if you know, participating in the liminal environment, liminal is actually catching that before it or evaluating it before it gets sent to the model for processing. Is that correct?
SPEAKER_00Not only are we catching it, we're doing it within your network boundaries. So everything is deployed as a virtual private cloud. So it's not going to us. We don't have access to it to that tool for evaluating how that data is uh is meant to be respected. So we we catch it, we work kind of in three-phase. We are able to detect noncompliant data. We're able to uh de-identify that data. So we place a, we we treat it according to a policy you put in place. Do I want to treat an age um by calling it you know nothing, just age, or do I want to treat an age as a range? Uh range gives the model more context, more to toy with without actually violating any HIPAA agreements that you may have in place. Um so you get to determine how different data types are treated before it ever reaches the LLM, all happening within your fixed tenant. So that means it doesn't leave your network boundary. Um then once that's de-identified and sent to the LLM, the LLM responds. We actually what we call rehydrate that non-compliant data back into the output. So back in the network boundary, we rehydrate that age, that name, whatever it looks like back into that output so that the user, the end user, is not getting a bunch of dear redacted, thank you for redacted, therefore redacted, which is a useless outcome from the LLM. But in fact, they're getting the exact outcome that they would look for from the LLM because we've been able to again push that that data back into the output the output from the LLM. So the user, again, they're getting what they're looking for, um, and and they have a totally unbroken experience. It's super neat because again, what you can track this through the the the logs, what was actually taken out of these, whether it's documents or prompts themselves, voice input, um, all of that is able to be appropriately de-identified and then re-identified back and rehydrated back into the output. So yeah, it's it's cool. It all happens inside of their tenant too. I just I can't clarify how important that is because it's not going to us. Um we're not looking at it. There's no human in the loop here. Um we've just been uh built some models that are extremely accurate and able to be deployed again and housed within your own virtual private cloud.
SPEAKER_01So for the listener, if you're in, especially in an environment where you're dealing with a lot of, say HIPAA or legal or anything like that, this sounds like maybe uh uh an issue that you weren't even aware about, but now that you are, you can't unsee it.
SPEAKER_00100%. Gosh, that couldn't be said better, Chris. Like once you know that, oh my God, uh a HIPAA violation could be somebody uploading a spreadsheet with a thousand names in it, and well, all of a sudden, that's a thousand HIPAA violations at a hundred bucks a pop. Oof. So so you you can, but but this is who we this is the people that we sell to. We sell to people that have regulatory uh penalties for data noncompliance. Um, and we see it in banks and financial services, insurance companies, life sciences, healthcare, state, local government, education, all of these are subject to um regulatory compliance, and all of them have significant penalties for not meeting that regulatory compliance.
SPEAKER_01Such a tricky environment for everybody because like you can't not use the tools, and you don't want your people to not use the tools. And you can put a policy in place and you can do training and you can, you know, spot check or monitor, but even then somebody might take a shortcut just because it's faster and easier, just oh, you know, it's no big deal, or nobody will catch me, or anything like that. So I I want to actually kind of dig into that a little bit. Uh in our pre-interview, you had mentioned that you had uh, you know, you had a coding background, a degree in computer technology, computer sciences, but that you had kind of dropped out of that space because you'd had some exits and you weren't really, you know, engaged in that anymore. Um what was the what was the event or what was the the news story or whatever that kind of pulled you back into building after that long away?
SPEAKER_00Yeah, so I I started my first company when I was 19. And I uh it was an online platform for banks to trade whole loan debt. But we ended up selling that company to uh Intercontinental Exchange. And after that point forward, I I took a more kind of keen eye on technical business development for um a few different startups and joined their founding teams and um and we had a you know a bunch of exits there. I did it and then I went to AWS and I did much of the same thing there. I led technology partnerships, I led go-to-market for their uh their emerging technology stack. And again, once you get into AWS and you get to see these things operate at scale, it really changes the game. Um I think I've always been a builder at heart. I I've always loved building companies. I think at this point, getting back to it was a it was a bit of a lifestyle choice. I mean, Chrissy mentioned, you know, you have a bunch of kids and um and and you have a wife, and you know, family tends to trump all else. And yeah, I think the decision for my wife and I around this time around was do we are we ready to do this again? Are we ready to get in knowing the sacrifice, the pain, the difficulty, the the I mean, the it's way lower pay. This is a terrible financial decision than working for a large company. Um it's the one of the worst financial decisions and stepbacks that have taken a long time, but the belief is that of course the upside in doing this is is significantly greater than the short-term downside risk. And and look, we've certainly created something that's more valuable than it was on the first day, um and has continually gotten more valuable the longer we've been doing it. You know, I I think uh that path is I I always liked, I was I had this, I was doing another one of these, Chris, and I had somebody ask me the question if you're sitting on a plane next to a young, you know, young gun who's thinking about getting something started, what are you telling them? I'm telling them don't, don't do it. Agreed. It's an awful it's an awful chance. And and I tell them that because if I can if that's what takes if that talks them off the ledge, if they go, no, there's no way I'm doing this now.
SPEAKER_01Good point.
SPEAKER_00Then then you shouldn't have been doing it in the first place. Um, not because I I and I listen, I it is a terrible financial decision. If you have the ability to go get work somewhere else, um it's it's really hard. Um but for for myself, I I and I think Chris, you you would you and I know I I know we relate on this point, you just can't not do it. Yeah, you kind of have to like it's just buried in you. You've got you've got fire, you've got drive, you've got desire, you have unweird risk belief in yourself that you have this weird appetite for risk, um, and you have hunger. And when you have those things and you have the the you, you know, at my age, you know, I'm almost, you know, I'm just over 40. You know, at my age, uh, you know, you have enough life experience to be able to make real damage. I've got a great network. I've got people around me, everyone I've hired on my executive team, I've worked with before. Um, you know, I think you you have those assets. It's a great time to also, you know, I'd say reduce the risk. I'm not saying that that risk reduction is significant enough that you should go do this, but I think you know the calculus. It there's there's a you gotta you gotta think about yourself and think, do I have it in me to go believe in myself to go do this, to go ride what is gonna be an exceptionally hard ride? Um all the way to whatever end this meets. Um, and end might mean failure. And by the way, in nine 98% of cases, these things fail.
SPEAKER_01Yeah.
unknownYeah.
SPEAKER_00So you're odds, the odds are I I might have told everybody before they joined, like, this is like a terrible decision. Like if you just weigh the Vegas odds, you're weighing like bad odds right now because all of the people I have working for us are extremely qualified and are capable of making 10x what they're making with with us right now. But there's such strong belief in AI, I want to build, I want to create, I want to be a part of something real, I want to do it in the AI age, that they've decided that they want that that this is the the train they want to hitch their ride to. And and I I think which is so fun, and it's so fun to see that. Like you that that creates a culture where everybody's everybody wakes up every day saying, How are we gonna how are we gonna survive today? How are we gonna how are we gonna survive this week? How are we gonna survive this month?
SPEAKER_01Speaking of that, you know, obviously uh with the repetitions that you've had in the space, you've made a lot of mistakes. What are some uh some pick one maybe one of the issues or mistakes that happen with this business getting started that that might be specific to like the AI space environment? Yeah.
SPEAKER_00I don't know if this is a mistake, but maybe just uh, you know, I think we took a little bit longer than I felt comfortable with um to adopt modern AI coding tools. Um, you know, you know, I and look, I think it was because there was appropriate skepticism. It wrote pretty garbage code for two years. Sure. Um, and then, you know, come January of 25 uh 20 of this year, but really December of last year, there was a major switch. Um when we saw Opus come out, it really changed the ballgame. And Cloud Code completely upended what I mean. It started writing really good code. It started to be able to write infrastructure code as well. Um, you know, we we witnessed a major step forward in that. And I still think we were like three months or four months behind. Um and in startup time, that's an eternity in terms of getting now. I know, Chris, like this is an AI show, and I don't want to like discount the fact that we're talking about AI, but let me just tell you all the lessons that people have learned, all of the tropes that have been told and talked about extensively when it comes to startups have not shifted in the AI era. We are in a technology adoption cycle like any other. And while it feels unlike any other because it's so profoundly different than it's been before, because the technical barriers are lower than they've ever been in history. That part is very true. The inevitable go-to-market difficulty that the vast majority of companies will face is absolutely still present. And the adoption curve of AI in the enterprise is still wildly slower than it has been in the consumer space. Um, you gotta think about people were still using BlackBerries for years after the iPhone came out.
SPEAKER_01Yeah.
SPEAKER_00Um we're we're not what we're we're there. People were still people are we I remember reading a report like three years ago that we were still only 30% penetrated in cloud. 30%. That means 70% are still still operating a bunch of their own uh servers. And and I just what a what an absurd notion when you came in and worked at that, like the that there's a you know still multiple trillions of dollars available there. Um, and and TAM and upside for these companies, you can see why they're valued at what they are. Um so okay, as it relates back to the AI world, like still follow a lot of what people have said traditionally about how to build appropriate go-to-market motions. Those those have not gone away. Um, what has gone away, and what the best part of this is you can get to a yes faster than you've ever gotten to before, because you're able to build really quickly. You can go out and prototype something on your own, even if you're not an engineer, and bring that into a prospective customer and say, Hey, if I did this, would you buy this and go and play with this? Tell me what you would like and don't like, and um, and then all of a sudden they play with it and they go, Oh my God, this is really cool. Yeah, I'd love to keep this. Um, gosh, that barrier is uh would have been a uh hard stop. You would have had to hire someone, some technical team, or you would have had to find a technical co-founder to go do that. That doesn't exist anymore. Um, you would have had to have a designer mock something up. That that ability to go and talk to a hundred customers to figure out whether you've got PMF or not, product market fit or not, that is so much lower than it's ever been before. Um, so I guess in in some stages, like there's the kind of this a dichotomy of like I can move faster than I've ever come gone before, but I've also the the same like buying cycles, the same way to penetrate enterprise, all of those things still apply. Um, and and now you're up against more people than ever because the technical barrier is so low. So people are able to go do this. So it's it's it's a really fascinating new space, but like In terms of like uh mistakes made, I mean, dude, we made all of the trope mistakes. We did things that you we thought the market was gonna move faster. So we hired in anticipation, then we did things that we shouldn't. And we had to fire because we didn't do it. Those are things that we made. We made the mistakes thinking um this is gonna move faster than it actually has, and it and it actually has followed a totally traditional tech adoption cycle. Um Yeah, let's talk about that a little bit because the same thing.
SPEAKER_01Uh I launched uh my butt my my business, chiefaiofficer.com in March of 2023. Chat GPT had been out for four or five months. So early, by the way. I thought we were late because you know, I saw what it was, I saw what was possible even with 3.5, and I thought, oh my gosh, everybody knows this stuff. Everybody must be doing that. In that environment, what what advice would you give an executive who they're listening to this show, they're trying to catch up with things, they're doing the best that they can? Um, should they feel behind? Because you just mentioned that the adoption is um that that we're still early, even though we shouldn't be. So, like what advice would you give to somebody who said, Hey, we're we're we're paying attention to it, we want to do something with it, but what what do you think? Like, uh how quickly should we move? Uh, what should we be afraid of? Those sorts of things.
SPEAKER_00Yeah. Gosh, I think, you know, it it it is a classic CIO dilemma that we run into over and over again, um, which is which camp do I belong in? Am I in camp A of like I just want to allow any tools so that people can be productive? Am I in camp B where I'm gonna bury my head in the sand and say we shouldn't do anything here until we feel like these tools are really up to speed? Um the answer is you can't be in either of those camps, but there is a really good crawl, walk, run approach here. And uh I think the crawl is you've got to get a tool in place. Okay. And while I would love to plug my tool as the only one out there um that that can meet the the data privacy and security demands you have, the reality is putting uh like Claude and ChatGPT or Grok or whatever it is that you're looking for, forget Grokbot for a second, forget Muse Spark. Don't don't think about those, those aren't even enterprise ready yet. But in terms of like getting your your your people using something, giving them access to something that you control, that you own, you cannot bury your head in the sand on that anymore. These tools are out. There are 1.2 billion users of these tools every week. You cannot afford to allow them to do things outside of your organization's control. And so I would be giving them a tool, whatever that looks like to you, in order to facilitate what is an increasingly needed skill internally, which is how do I employ these tools to do my job better, faster, of higher quality. You cannot afford to not do that. Um, you will be dusted by your competitor who is absolutely leveraging these tools, um, or certainly will be. Now, are you behind? The answer is absolutely not. In fact, I'd say 80% of the conversations that we have today are still people in the, hey, we we know we need to deploy something, we just don't need to, we don't know what it is. Um, so that and listen, we're having 50 conversations a week right now. So I just clear on like no one's behind here. You're you're right where you should be.
SPEAKER_01That's why I was particularly interested with this concept of wait a minute, something sits between my people hitting enter on Chat GPT and that information going out into really as an executive, you don't you don't know where it goes, right? So um that's why I thought that this this conversation about this mechanism that that's employed uh when you use liminal was really important. Having some sort of a bridge that look, your people aren't trying to screw your company over. No, they got to get home by five. And or they really hate doing that thing. And boy, it'd be a lot faster if I could just do this. And, you know, they don't really pay attention to the AI use policy, and they slept during the governance conversation. And this is a mechanism that allows for the business owner to know that I don't care if they read it or not, it's not going out there. It's gonna get caught, it's gonna get redacted, it's gonna get anonymized. We'll still get the great output from the models, but it's getting anonymized before it can turn into this big pool of training data.
SPEAKER_00That's right. And and big pool of regulatory compliance fines. Agreed. And so I I yeah, I completely agree. And I I think the I can I Chris, I just want to keen on one thing you said, which was, you know, uh that that folks are uh folks are not trying to to to harm their companies. Yeah. Um I I I can't emphasize that enough. Um people are just trying to get their jobs done. And the person who's taking their mobile device and snapping a picture of something on their their desktop here, and you know, to they're not trying to sell that to China. They're they're trying to do their job. And uh I I think the the reality is uh you know, as a CIO, as a business owner, it is incumbent upon you to help them do their job better. And if you see them doing this on their mobile device, you shouldn't be punishing them. You should be giving them the tools to do it on network. Um, people aren't trying to screw you, they're trying to win, they're trying to do well at their job. People want to do well at their jobs. Um, nobody shows up to work hoping that they they they mess up or get yelled at because they've done something inappropriate. Uh one of our healthcare customers, Northern Hazel and Healthcare, he said, I've had employees come up to me and tell me, if you ever take liminal away from me, I will leave the company and go find a company that has liminal to go use it. And that was a real threat. He's like, oh, that kind of scared me a little bit. Like, uh, you know, that that's the reality of when you provide something at someone a tool that helps them get their job done that much better to strip that away from them is uh uh it's it's not great.
SPEAKER_01So and you know, one of the things let's compare this to anything else that they've been using. They've been using email or they've been using Google Drive or Microsoft. And there's never had to be this conversation really about like you need to be careful about what you what you attach to an email. Well, we've got the disclaimer at the bottom. Okay, sure. Nobody ever says, hey, be careful what you upload to Google Drive or to SharePoint or anything like that. And breaches occur regularly. So they there hasn't been this practice by anybody of let me let me think about what I'm doing here before I send the attachment or before I you know create the the slide deck and and upload it to Google Drive. But now you're in a situation to where you've got this temptation to use this tool that, wow, that used to take me two or three hours, and now it only took me 20 minutes. As you know, somebody that doesn't want to work on weekends or has to get the kids to the game or whatever the thing might be, like that's the lifesaver, but there hasn't been more than maybe a conversation or some scary talk about, you know, you have to be careful about what you put up in there. So I think that this is um Go ahead.
SPEAKER_00When you're doing your when you're doing your courses on on training organizations on how to use these tools, how many of that con how how many people are saying, hey, this sounds like governance is needed? Um and I'm I'm actually thinking that number's very low, by the way, just in case you were gonna say you were gonna couch and uh well, you know.
SPEAKER_01It is. I was and I'll give you a perfect example. I was speaking to a group of um presidents of an or of organizations uh in Nashville earlier this year. And before I got on stage, I was chattering and and they were all interested in like open claw, right? And I was like, Oh, that's you're pretty advanced, that sort of thing. But then I started asking them questions like, well, well, what are you using in the company right now? And they're like, Well, we got some people with some AI licenses. Well, are they your licenses or theirs? And they're like, I don't know. And I was like, Okay. So they're more interested in this was presidents, you know, they're uh hired guns in some cases, but they were more interested at as probably most a lot of the listeners are. It's hard to not look at the shiny object and to think about, you know, the boring conversation about governance. I mean, just the word alone, it's like uh like AI gives me freedom. So it's awful. Um most of them uh you know, they're starting to, but but the the risk is just it's a this nebulous topic. They they don't know what they don't know about the risk. And when you mention like an AI use policy to them, they're like, oh damn, that's a good idea. But it's not something that they've thought about prior. That's probably one of the worst.
SPEAKER_00I guarantee. And like you're focused on all the value that they can achieve from this without thinking about governance. And and the reason I asked that question, by the way, is it shouldn't surprise anyone that that that people aren't asking, but that's not the cool stuff that AI is known for. Um, and so what does that mean? It means organizations can race into deploying these tools without thinking about the regulatory consequence of the of how data is being used in exchange with them. And and I I'm not here to be the harbinger of negativity. I'm here to be the opposite and say, like, you can achieve all of the cool things Chris and his team will train you on how to do um while still being able to provide an effective governance layer on the on the topic. That's that's the piece that like I think is you're gonna check, you know, your CISO, who's your you know, typically either reports side by side to the CIO or reports to the CIO, and who's always the the arbiter of no gets to be the arbiter of yes. Like we want to turn them into champions. They don't have to be the guy saying, no, we can't use these tools because there's too much of a data risk here. You get to say, hey, I I I actually want you to use these tools and you should deploy them everywhere, um, give people access to them. Um we just need to make sure that we have the appropriate governance and that person's job is taken care of. While of course the CIO gets the victory of being able to provide his end users exactly what they're looking for. So that's again, that's the promise that we've always tried to say. We always try to say, can we get people, the CISO and the CIO, to get to W's, to get the wins that they're looking for, to get to yes, to make them the arbiters of yes, and that's what we hope we can do. Um and we've certainly been able to do that over the last couple couple years.
SPEAKER_01If I'm working with a tool like a liminal, that, you know, can I still use automations that allow for when the email comes in that it extracts the PDF and it reviews the PDF, and then if there's missing fields, it sends the email back to the sender and and you know, because AI is having to in an automation or an agent doing this, AI is the the intelligence, the brain behind the evaluation and the determination of next step of action. Can I still do those things with liminal?
SPEAKER_00We are not a coding harness. So liminal does not replace Cloud Code or Codecs or any IDE with integrated models or coding agents in them. Um so I want to make that kind of crystal clear. Like we are for all of the kind of uh alternative productive uh productivity use cases. The way that agents are done today, outside of the coding use case, um, is much like I don't know if Chris, you ever played with OpenClaw, you mentioned that earlier. Um I, you know, I deployed it when it was ClaudeBot before they had to change that. And um, you know, when that was you know all a rage going around on Twitter and I put it on a machine luckily here, and I had it like searching through, sifting through a bunch of like LinkedIn product candidates, and um, we were searching for a product manager and um, you know, kind of surfacing people at the top and doing a scoring mechanism. And it was fine. Like it was it was insanely token uh token hungry. Uh it was like I can't remember, I spent so much money. I think I like ran up a $150 claude bill um just trying to configure Claudot to go do something very simple. But uh more importantly, like there wasn't a lot of governance or understanding, and and kind of most importantly, you myself, uh others that are maybe even listening to this call, we are advanced users of of AI. Yeah. Um the reality is that represents less than five percent of an organization. I can tell you that with impunity, with certainty, because I deploy to large organizations all the time. Um in fact, probably closer to less than two percent of an organization of any kind of real rich depth. Okay. But like the notion of being able to deploy genetic frameworks in order to accomplish some multi-chain of action automation is really cool. The reality is uh it's a lot harder than people want to give it credit for. Um, it's a lot more technically complex than people want to give it credit for. And uh it requires people to want to lean into that. And for myself, for you, for power users, they want to lean into that. We're all curious, we want to go do that. But for the average worker who's showing up at nine and leaving at five and has a family, they don't have time to do that. So when we think about agentic, we think about how can I make that user's life simple. The power user, give them power user tools. I want you to have codecs, I want you to have cloud um and and uh and co-work. I think those are really incredibly powerful tools that that will get put to use with the right, it is categorical overkill for the vast majority of end users. Yeah. Um and so you're you would be paying $200 a month for a Cloud Max account or a GPT Pro account when what you really needed was something like Liminal. And the reason I plug in Liminal here is because we are really built for that 95% of the organization that sits below the power user threshold that wants to experience automation. And the way that we think about it is uniquely we sit on, I mentioned earlier that we have observability baked into our governance layer. And observability is the ability for us to see what users are putting in and what the prompt outputs are being driven from the models themselves. Um, that's of course for uh the purposes of audibility, but what we've been able to do is uh is run that through some models of our own creation inside of your tenants still to be able to understand patterns of behavior that exist in your workflow, your everyday life. Because again, you're sitting on this incredibly rich data set of how people are using AI, what questions they're asking them on a repeatable basis, what they're trying to accomplish in their job using AI, and how they're leveraging their internal data to accomplish that job. So now we have this great data set that we can go and say, hey, Chris, I noticed that you've been doing this thing every Tuesday at 8 a.m. Would you like me to do that for you? You're accessing these five different connected MCP servers, you're accessing your enterprise search loop, you're taking information from the web. Um I've watched you do this thing repeatable multiple times. It's repeatable. I can do that for you. Would you like me to deliver that thing to you every day at 8 a.m.? What we call that is behavioral agentic automation, where we are looking at uh you know your pattern of behavior, identifying high value workflows, and then self-deploying and self-improving agents on your behalf. So I don't need you to be a technical end user because use the end user don't want to have to build out an agentic framework. You just want to go, I'm an accountant and I do a variance analysis on Tuesday of every month, and I from I create a deck for my CFO. I sh I did that in Liminal once. I pulled data from a bunch of different systems, had it output it for me in a deck and a spreadsheet, um, and then I deliver that to my CFO. Well, the second time on the next, on the 14th of the next month, I went in and I tried to do that, and Liminal goes, Hey, I noticed that you do this on the 14th of every month. Would you like me to do that for you? And the accountant goes, Yeah, that'd be great. Um, in fact, give me a five-day buffer on that so I have time to work on the draft with it. Great. I'll do that for you on the 9th of every month now. Um that's that's how simple automation is in our world. Now that doesn't mean, again, I'm not beating up agentic frameworks, I'm not beating up the lang chains of the world, I'm not beating up the I think they're it ain't easy and it's oh and it's costly and it's um often overkill for end users where where I can say, hey, I I know how you work, I know how you operate, okay. Detect those behavioral patterns, and then we can deploy those agents on your behalf.
SPEAKER_01For the listeners, what's what's one thing that you have been able to hand off to generative AI and whatever that tool may be, that has been a huge win for you? Because one of the other reasons that people don't use AI more is I'd love to, but I don't know where to use it, right? And I I I love getting these ideas from power users, folks who are capable uh with these tools in an operational environment to say, oh man, I did this, this, and this, and that saved me a lot of time, or I get better results or whatever. So what's what's something you'd handle?
SPEAKER_00So many things. I'm like, I'm trying to kind of wrap my head. Okay, so my favorite app that I built um uh using Cloud Code uh was a uh I call it the pocket sales engineer. Um our sales team is um, like many sales teams, are not particularly technical. They're not engineers by trade. Yeah. And I don't want them to be. They're phenomenal at their job at helping customers understand the problem and where we can fit and using us as a solution. They're they're sales folks. Um but they're not technical. So I built an application um using Claude Code and exclusively clock code that um listens to calls in real time. So they're talking to a customer and it listens to questions from a call and it sends that question when it detects a question, um, it sends that uh question over to uh a knowledge base and hunts for the answer to that. So based on you know, 350 different calls that we've done in the transcripts are we built this knowledge base that has the appropriate answers to those technical questions, and then surface it on the screen and for them to answer in real time in language that they can alter. So like the they can become it's listening and says, Oh, that's a technical question. Here's the answer that surface it on the screen to them, and they can say, Oh, the answer to that question is. Um, and it gives you like an accuracy metric on how accurate it is based on the knowledge base, how many times that's beautiful. Um, that has been a really fun application, and it's been one that I've like considered can commercializing. Like we were running a company already, so it's I'm it's really cool because it it turns our salespeople into weaponized sales engineers, um, where they actually have the technical knowledge and they only need to answer it once before the next time it's asked, they've got that answer in their head. But then this the loop becomes self-reinforcing because the more calls they do, the better the knowledge base improves. Um, and the more accurate these answers become over time. Um, so it's been it's been really cool. Um, I've got a bunch of these. So I did uh another one where um I took all of our call transcripts and uh plugged them into because we don't have a product leader right now. We want one, but we haven't found the right fit yet. Um so I and my CTO work on product together and we kind of uh prioritize a roadmap based on customer demand and then features that we're coming up with. And um I uh took all of our call transcripts from the last 120 days, and I do this every two weeks, and I run it through um uh a cloud model, then the application then surfaces the top 15 features requested on calls with the number of customers that requested with it, a semblance of understanding how urgent that request is, so that we can prioritize a roadmap based on customer input. Um and it does that without me having to, it's so much easier than me having to collate that data on my own. Um, I just have it, you know, listen to all those call transcripts and it extracts those transcripts from a service we use called Gong for our sales team. Um and uh and it just I just have to click a button and I get a feature request update prioritized by urgency, prioritized by size of customer, prioritized by any number of different parameters. So those are my two favorites. Um I also just rebuilt our entire partner portal. Um, so I didn't have to pay somebody else to do that. And I was paying, you know, whatever, 20 grand a year for our partners to be able to register deals to some other company. Um, now that that whole deal registration process, which is more deeply integrated with our systems. Dude, I've been on a build frenzy. I've got so many of these things. Um my favorite one is definitely that pocket sales engineer because it's been so helpful for onboarding our sales teams to be smarter, faster, um, and get the answers that they need.
SPEAKER_01So for the listener, uh maybe you're not gonna be able to build your own tool, but but uh what I'm looking for when I ask this question is something where you can go, oh, A, I didn't know I could do that, or I don't do that, but it that becomes like the thread that you get to pull. But we're doing this thing, yeah, and it's kind of like triggering something for folks.
SPEAKER_00So let me just tell you guys, I didn't write a single line of code in any of those applications I just discussed. That's the reason why I brought that up. So if you're wondering the power of AI, all of those are hosted. Every one of them is SOC2 compliant, every single one of them is fully functional and hosted. What'd you use? Uh I use cloud code for all of them. Uh and codex for the last one because I just wanted to see how powerful Astro was. Yeah. Um but uh but uh uh cloud code um exclusively. So uh I I I say that to go, I didn't know I could do that with AI. Guys, you don't need to be an engineer, you just need to be able to capably communicate your vision.
SPEAKER_01And I can tell you, as somebody who's done those types of builds, the first ones are gonna be a couple kind of messy, but you're gonna learn. And then from there on out. And here's what here's what you're doing. You're giving it natural language instruction. I want something that can do XYZ, and the models figure it out.
SPEAKER_00So Stephen and Ken, if I can give you guys, like real quickly, Chris, just a little bit of advice on how best to do that. Um, what you want to do is clarify your call your thoughts in exchange with the model. So, like before you go into Cloud Code to have it write anything, what you want to do is I I chat with Opus on in Cloud and I just go, hey, I've got an idea for this thing. Let's work on this together. Be a creative partner with me, be a thought partner, let's work together on refining this idea. And we go through it and we refine it, and then I say, great, build the requirements for Cloud Code to go and build that. It builds this markdown file. I take that markdown file, I plug it into Cloud Code, and I tell it to give me the prompt for it as well. It gives me the prompt, and then Cloud Code goes and runs with it. That's how easy it is. Like That's coding these days. It's wild. Yeah. So just so you guys are aware, like you don't have to be an engineer to be able to see that thing that's in your head come true. And that gets really exciting when you start to realize just how capable you can be, even if you're not a software engineer.
SPEAKER_01Yeah, and I'll tell you, you know, for the listener, this functionality, what he's talking about, it exists in your $25 a month or $200 a month plan with Chat GPT or Claude right now. So the only thing stopping you from doing that is you just haven't clicked on that navigation link and said, uh, let's build something. So that's great advice. So, Steven, um, outside of going to liminal.ai to find out more about the company, are you producing, I don't know, uh position papers or thought leadership content in any of the platforms?
SPEAKER_00LinkedIn is where we tend to broadcast all of our content. Um, we're not really big on Twitter uh or um the other socials yet. Uh, I think it's an expansion opportunity for us this year, but um, we just haven't had that as like a primary lead mechanism for us. And um of course you can go to our blog and subscribe to that on our website. But if there are any of you out there that would think, hey, I've got customers of mine that might be really benefit from something like Liminal. Yeah. We go to market almost entirely through channel, so resellers and um and uh distributors as well. So if you're one of those and fit that, we would love to speak with you about what whether or not that could be an opportunity for you.
SPEAKER_01If you are in an industry or your this position is particularly sensitive to I don't, I don't want to trust them. I want I want that extra layer of insurance. Uh, liminal is something that I would encourage you to look at because this concept it will help your people from making the mistake and having to learn from the mistake, it'll catch it in advance. So if everybody thinks thank you again for listening to uh using AI at work. We're now in episode 120 something, which is uh pretty incredible. If you're getting value out of these conversations, which I hope you are, if you're, you know, I our listenership continues to grow, um, and there's other people in your organization that you want to get them on board or you want to get them to be able to understand it enough to go, hey Chris, I think that's a great idea. Let's go talk to the boss or let's bring it to the board or whatever. Um, please make sure that you're letting them know that you're listening to this, share it. And uh if you want to, um we'd love it if you'd leave a review. Um, they always help other people kind of determine is this thing for real? And we are for real. So uh again, thank you everybody for listening, and uh, we'll see you next week for another fascinating episode of Using AI at Work. Thanks for tuning in to Using AI at Work. Don't forget to subscribe for more conversations about how to use AI at work. And a special thank you to our sponsor, Chief AI Officer, for empowering businesses with AI education and training. Visit their website for a free AI readiness assessment and AI strategy guide to help you get started using AI at work. That's www.chiefaiofficer.com. Follow us on Twitter at the handle UsingAI at work and visit www.usingai at work.com for free resources to help you harness AI in your role.