Transcript
I am very happy and pleased to have both my partners in GI ventures Vijay Rajendran and Kushal Prakash. So, welcome guys. Every media outlet, Twitter, LinkedIn, wherever you go people are talking about AI. But, what people most people miss is that we are actually entering a completely different phase. There are about 1 billion users of ChatGPT, but only 9% of companies surveyed say they have put an agent into production. Since the margin for error itself is very less, it's a big trade-off for them.
If they adopt the LLM, they might save on time, but then they [singing] really can't compromise on accuracy. Let's drop a truth bomb that at least looking at 200 people who have been using Open Claw and Claude code and Claude co-work. One dirty secret nobody shares is that whenever they have to run even a medium reasoning task or medium complexity task or like a higher complexity task, even the latest models cannot give a one-shot answer. You'll see enterprises ask very critically is the juice worth the squeeze here? And is that worthwhile to be spending this much money on tokens? We need a lot of rule sets, lot of steps, processes defined for an LLM to work in a very accurate way and doing the exact set of actions which it should be doing.
Yeah, we're reminded of how in 1996 future-thinking CEOs heard about the World Wide Web and they told people in their organization, we're going to have a website. What's it for, boss? We don't know yet. And it turned out that they never thought about how the business model would need to be different. They never thought about who were the customers, how they connected to this. It was just a a screen for an attention and that wasn't monetizable.
And so, eyeballs turned into a bubble. Similarly, right now we have a lot of enterprises looking around and saying like, "Hey, I heard that we can get rid of people if we use agents." Or I heard that if we use LLMs, we're and we have an AI strategy and the board can't stop talking about the AI strategy, then other good things are going to to happen. Gosh, this can go like absurd if it's not controlled and tracked properly, especially while learning LLMs. A lot of it doesn't really require intelligence. The easiest of task people are trying to build agents for that can actually be codified. The mileage between codes and the LLM is what I believe has to be more disciplined.
In a way, I feel most of it would be hybrid in the future where LLMs are being used where necessary while in the other places the codes can actually do the heuristics and follow the set of steps. The next 12 months, people are going to talk a lot about AI revenue growth, cost management, observability, e-wells. There is a lot of value. There's a lot of opportunity and that's, you know, where Vertical is going to be successful in basically hyper-vertical pieces of the landscape. And just as a recap, we do not just invest in AI companies, we build them alongside the domain experts who know their industry better than anyone else. This is the phase one where the hype gets replaced by discipline because if the enterprise adoption has to happen, a lot of things have to change.
Hello everyone and welcome to Build AI podcast. Uh this is yet another episode where we go deep on what it actually takes to build and back vertical AI companies. Uh I'm very happy and pleased to have both my partners at NGI Ventures, Vijay Rajendran and Kushal Prakash. So, welcome, guys. I'm happy to be here. Let's do this.
Thanks so much. Yeah. And and just as a recap, we do not just invest in AI companies, we build them alongside the domain experts who know their industry industries better than anyone else. We basically have created a platform where we can take companies from minus one to one right from the idea stage to you know, when they have customers and they're ready to raise their next round. And we do pretty much all the product engineering AI engine work as well as help on the on the GTM side as well. And so with that recap, let's jump in.
As both of you know, I wrote an article in January that there's an enterprise AI adoption crisis. And since then we have seen even more hype on the consumer side, you know, with open claw and cloud code and cloud co-work. But everybody tells me and everybody keeps sharing the use cases on the personal productivity side of what they are doing if they're running an agency or their own personal work. Sometimes they also doing companies work, but really there's no control plane that a company has. They're they're just doing it on their own. So the big question remains that what's happening in the enterprises?
And so we thought like let's record this episode where we talk about why AI hype is ending and why that's the opportunity. And you know, obviously when we are talking about it, we mean more on the enterprise and SMB adoption side because personal productivity side like I think we are on our journey towards AGI. The the the SMB and enterprise side is like a very interesting question that where are we? Yeah, that's a a great observation, right? There are about 1 billion users of chat GPT, but only 9% of companies surveyed say they have put an agent into production. Where's you know, the the it's So where's the reason for this like big delta between what people feel empowered to do and where they have agency and where companies have really struggled with this.
And it comes down to you know, a lot of the reasons why businesses at a certain scale will struggle. And that will be something interesting we can unpack, whether it's in information security or whether it's in the way they they manage their uh their data and who can have it. Right. And I I think everybody everyone knows AI is everywhere. Like every every media outlet, Twitter, LinkedIn, wherever you go, people are talking about AI. But what people most people miss is that you know, we have been in this space since 2024, and we are actually entering a completely different phase.
I I think this this is the phase, and I'm talking about from now and and forward, right? Like one where the hype gets replaced by discipline. Because if the enterprise adoption has to happen, a lot of things have to change. And for people who understand what it actually means, the timing has never been better. Question to you guys, right? Like this 2024-2025 AI hype cycle, you know, with chatbots and copilots and generic tools, why enterprise adoption is not did not happen, and why most enterprises are frustrated with the results?
That's a good question. So, if I can complete the thought of um a bit more from before, [clears throat] there is a learning gap. And this learning gap is is huge. And that learning gap it has to do with the workflows that people have. And that is really hard to take from one type of approach, which is puts in information and who gets it out, and then who do they send it to, let's say, to something very different, which is how can I develop and manage a an either an agent or instruct a you know, an LLM to be generative on my behalf. That is really really hard.
It's it's a big shift. And even in that 2024-2025 paradigm, you know, that was hard to do, and that's not going to get easier when we start asking people, "Okay, go create your own agents now." I do this point, enterprises have mostly been using deterministic legacy softwares for the last decades. Now, the shift has gone towards it started with 24, 23 where it was a Q&A chat with LLM where they could get the answers. That started changing to actionable. That's where enterprise adoptions opened up where they can now take actions which they were earlier doing with their deterministic softwares with LLMs. However, the shift hasn't been as smooth as it should have been maybe because they are not very trustworthy.
They were not very trustworthy back in 24 at least. Actions were not always deterministic like they were and for people who were sitting and spending time and doing those things, now to completely delegate to an AI was a complete behavioral shift and once the reliability was not as much as it should have been, that was another problem. So, that's where the adoption itself wasn't as good and it was hard for big enterprises to trust their data onto LLMs where the outputs were not always the same or not always 100% accurate. Since the margin for error itself is very less, it's a big trade-off for them. If they adopt the LLM, they might save on time but then they really can't compromise on accuracy. I think the shift has been very gradual.
Over time, things have improvised a lot on the LLM side as well and there's a lot of new developments like agents and security guardrails a lot of Right, right. So, okay, let's drop a truth bomb that at least looking at 200 people who have been using open claw and claw code and claw code work, one dirty secret nobody shares is that whenever they have to run even a medium reasoning task or medium complexity task and or like a higher complexity task, even the latest models cannot give a one-shot answer. Usually, what people are doing and they're not talking about it because they feel like they are less smarter than others if they talk about all the issues they face in getting work done with the AI. There is so much of back and forth that like for example, one of my friends was telling me that they he was trying to update a database of about 30,000 people uh you know, which probably had changed the jobs. Therefore, their emails were not delivering and stuff like that. And it took them like 3 days of figuring out what the right approach should be because every time they run an approach and clock ChatGPT, it will give like very bad results, lot of inaccuracy.
Also, seems like Kushal, maybe technically you can explain why this happens is that either the AI can handle breath or it can handle depth. It cannot seem to have you know, like be able to do both. For example, if I have to ask it to do research on 30, 40,000 database records database across like 20 fields, it will struggle. Yeah. Uh great point, actually. So, the way LLMs itself work, they are trained on billions of parameters, which is information across the world, which it should learn or it should have decent context about.
But when we are trying to make it more focused and trying to run either simpler or complex workflows in case of bigger enterprises, it's a bit harder because the context knowledge of LLM itself is bigger. Now, trying to focus that onto some parts only, let's say research on some individuals like you mentioned, there it's more action-oriented because it's very likely that that context is not already there. It has to go do some actions, figure out, and then give the inference back to the user. There, it tends to struggle a lot of time because the context knowledge itself is not that sufficient. And going and taking the relevant action requires a lot of training and a lot of guardrails and lot of processes as well. Like you mentioned, it takes a lot of back and forth and with some experience of working with LLMs, I can say that we need a lot of rule sets, lot of steps, processes defined for an LLM to work in a very accurate way and doing the exact set of actions which it should be doing.
In terms of intelligence, it's really great because it offers its perspective and that comes from the billions of parameters which it is trained on. But when you give it a set of actions and ask it to do over a large data set, it might deviate across a few of them because it's not 30,000 is a huge data set and expecting it to do the exact same set of actions over all of the 30,000, that requires a lot of steps to be defined and that becomes a workflow which needs to be much more focused and built in a specific way to handle those actions. Yeah. And and that is like I think at the core of enterprise adoption problem. Like when we are searching for, okay, you know, build an itinerary for me, you know, I'm trying I'm going to Austria or some other place and it will do a great job, right? Or I ask it to find out information on Kushal and Vijay and you know, what they have done in the past, it will give like the whole encyclopedia on the two of you.
Mhm. But then [clears throat] enterprise use cases are actually on 30,000 or 300,000 records and you have to do the same level of accuracy and breadth of parameters on the depth of data which is like problem number one. Problem number two is that you will end up spending so many tokens on doing that job uh which is inaccurate at the end of the day at a 30,000 or 50,000 record level that cost management itself is going to become a very big issue. And then of course, you know, Kushal, as you talked about like, you know, the the drift of the agents, the accuracy, the the inaccuracy problem because of that. I I think these are some of the problems that that we are facing. Is Is there anything any other problems that you guys see?
I think your point about token cost and how that will catch up with us is a very important one and we will we will see enterprises ask very critically, you know, is the juice worth the squeeze here? Is it worthwhile to be spending this much money on tokens? And in some cases, the answer might be, well, maybe we actually want to use a previous generation of the model as opposed to the latest one. In other and in some other instances, it may be for what we're trying to do and using more internal data sets, rag and other types of little language models can be more effective and more cost efficient. But right now we're in the exciting phase of what can we do with this stuff and not in the penny-pinching phase where people really care that much about, you know, what all this this compute is up to. Very cool.
Just [clears throat] to add and and you know, go to the next point, there have been studies like, you know, the the famous or infamous MIT study where they said 95% of the companies report that you know, generative AI implementations are falling short. There have been a bunch of other studies after that, different numbers, but the bottom line is that there's something wrong with the with like, you know, how enterprise can prove ROI. And as as this year we spent 2.5 trillion dollars in right in terms of investments and and projects and so on. Just like if if the ROI is not coming, I think there will be questions. While on the other hand, what we actually see is Jensen Huang saying that if I have a engineer whom I'm paying half a million dollars and if they're not burning quarter of a million of tokens, there's something wrong. Right?
So there's a there's like such a big gap between hype and reality sometimes. What Like how do we close this gap? Yeah, both these things can be true. Uh in terms of what Jensen is trying to run towards and build out as quickly as possible. And then on the other side, I think expectations that businesses have but the lack of enablement that they provided. So, what does that mean?
You know, we're reminded of how in 1996, you know, future-thinking CEOs heard about the World Wide Web and told people in their organization, "We're going to have a website." What's it for, boss? We don't know yet, but you know, they spent time and attention and consultants and designers like thinking, "Oh, okay. Well, if we're a media company, then we will have an online magazine or we'll try and do the online this or the online that or the e-thing." And it turned out that they never thought about how the business model would need to be different. They never thought about who were the customers, how were they connected uh to this thing. It was just a a screen for attention, and that was monetizable. And so, eyeballs turned into a bubble at that first.
Similarly, right now, we have a lot of enterprises looking around and saying like, "Hey, I heard that we can get rid of people if we use agents." Or I heard that if we use LLMs, um we're and we have an AI strategy and the board can't stop talking about AI strategy, then other good things are are going to to happen in terms of greater efficiency in using our data and how we're going to just do all these things really fast. And whether it's financial analysis or doing legal contracts or something like that, the sort of magical moment of I prompted and I got something back does not translate into we're going to change the way the business operates fundamentally because the old way of making money and serving customers didn't change. I I really like the how you termed it is the juice worth the squeeze. I think this there's too much effort going into the squeeze today to to sort of you know run anything reliably on the enterprise side. Yeah. I I something to add there.
So, I think one of another problem is it's hard to measure success in AI runs. At least that's what I believe is one of the primary cause of low adoption rate of taking time to adopt the AI systems. The way AIs are expected to do expected to work is not just not just taking data and do something. It's about running workflows, taking actions, doing some complex things which a human would do otherwise. The problem here is over time as in when you train the models or add more context, AIs also tend to make mistakes. There's a problem of over training or data leakage, data trafficking which is which has become common among AI LLMs.
For this to be solved, again, that has been the bigger issue wherein we have to have the whole AI self-learning based on what's happening and if there's any change that also has to be learned by it and the action has to change accordingly. And it has to ensure that there's not much changes which are happening over time which is one of the reasons why I believe businesses, enterprises are taking time to adopt because they don't see how the whole success itself can be defined here and how to exactly measure and ensure it's learning on its own. There was a recent report by cio.com called master of board ROI report and it said 56% of the CEOs in enterprise say that AI has produced neither revenue growth nor cost savings over the last past 12 months. Uh seen adoption in some of the personal use cases and then at small and medium enterprises, the numbers are much higher, right? So, all is not bad. I think it's just it requires a lot of planning and lot of architectural decisions to figure out the right approach to use AI to solve a problem.
And in some cases today, it may not be the best solution, but in many cases it is. So, I think you need unlike what people are saying, anybody can write code and you know, like create a solution. I think it requires, you know, people who have actually experimented and have muscle memory with it and you know, like know a lot of different approaches to take depending on the situation. But that is something I feel like it's like underrepresented today in in the narrative. Yeah, I think that's a fair fair assessment. Yeah, there is a big difference between, you know, I I have something that's a solution and it's a horizontal solution or it's a GPT wrapper and you know, people feeling like this really was built for our industry and for our problem and it was well designed with domain expertise.
And and so, in those few cases when that does emerge, those companies have turned out to be very valuable and and successful. I can think of, you know, Harvey in the legal space is like one prominent example, but there are others, too. Uh and so, you know, that's that's an important criteria when we think about how to build for a particular space and how to invest in the right types of founders. So, there's this ongoing debate with regard to that point, right? That the shift like some people say there's a shift from horizontal to vertical and the others say there's a there will be a shift from vertical to AGI or like, you know, Claude gets so good that, you know, because Claude is there for legal, Claude is there for Excel, Claude is there for a bunch of things and that caused like $250 billion of market eruption disruption earlier this year. Which direction do do do you guys see it going?
Is it are we going more vertical or are we consolidating in in in Anthropic's stability? Anthropic this week and you know, some other model next week, right? Like at the rate at which they are all getting better. You know, I I think we were joking earlier this week that two months ago in every WhatsApp group there were people like asking if somebody had an allocation of OpenAI shares and right now, you know, the the same conversation's going on. Do you have an allocation of Anthropic of Anthropic shares. And then who knows?
Maybe in six months it'll be, you know, LLM X or company X. So, you know, I think the the horizontal game is is on and those things will get better. Um, but much in the same way that, you know, in in the web uh and and internet 2.0, there were there were ways to uh you know, maybe do word processing with uh with Google Docs or, you know, or you could use Microsoft. Um, and one of them is free and one of them is paid. Uh it turns out that, you know, there's there's a lot of like room across like a massive application like that. But alternatively, if you are building something that is just for doing a certain type of of work and it needs to be secure and it needs to feel like trusted and and and for not everybody, but just for you.
But then, you know, in those particular niches there is a lot of value. There's a lot of opportunity and that's, you know, where vertical is going to be successful in basically hyper vertical pieces of the landscape. Yeah. I think it'll be very industry specific as well. Like health care, financial services, that's where vertical AIs would win. While a lot of other industries, there there there are still horizontal AI players who are going to dominate.
Yeah. Yeah. I think regulated industries where there's a lot of process heavy, compliance heavy, data rich environment, there I think vertical AI will make sense for some of the reasons that we discussed earlier. And you know, in in our company Faster Capital AI, we're just focusing on a very complex workflow called transitions for RIAs. Today I can tell you with guarantee that you can have the max planning cloud and whatever like you know, model that you use 4.7 that they launched today and you try it out for like even 3 weeks, I don't think that transition problem can be solved. The workflow is so complex.
It involves looking into dozens of Excel sheets and input from various softwares from financial planning to CRM. And then choosing from 65 different forms and filling them and you know, 10 other steps. Uh long horizon agents, you know, as I was saying, can do a bunch of simple steps on a small database or like you know, some breadth and maybe a little bit of depth or or vice versa. Um okay, let's talk about the AI discipline because I think the next 12 months people are going to talk a lot about uh measurable ROI, revenue growth, cost management, observability, e-wells. How do you guys think about these things? The discussion right now is quite less to be honest.
I So far I think in the last 12 to 18 months what we have seen is can we do this with AI? And the answer seems to be more and more yes. But then once you start doing it AI, like in the case of one interesting use case, somebody actually deployed AI for a use case and then they measured after you know like a month and they said like amount of cost involved in all the time we spend on building it back back and all the back and forth and the token cost and the cost of cloud and everything put together is actually higher than what we used to do manually earlier. So, that's the discussion I think we'll be having. Yeah, exactly. We we we don't know because we haven't probably got the full accounting.
I think the full accounting for this. I think costs can go like absurd if it's not controlled and tracked properly, especially while running LLMs. I think people have burned thousands of dollars overnight using open clock just for very simple tasks as well because sometimes it tends to get confused and it runs over and over again and in no time you would have spent a lot because as and when you add up more context the tokens are increasing and then that's also adding up to the costs. So, it's important for especially businesses to be very wary of these expenditures, where what models are being used and for what reason. Like for instance, smaller tasks are some easier executions, easier workflows doesn't require an Opus 4.6, Opus 4.7 which are very expensive. They can always use open source models, hosted themselves or go for a cheaper APIs as well which can control some costs, but more than all of it it's very important to ensure that what is being used and how much is being used of the LLMs is also controlled.
A lot of it doesn't really require intelligence. Like the easiest of task people are trying to build agents for while that was that can actually be codified. Just create a code for it. Where there's intelligence required to take difficult actions, that's where LLMs AI should be used. So, the mileage between the codes and the LLM is what I believe has to be more disciplined. In a way, I feel most of it would be hybrid in the future, where LLMs are being used where necessary and their tokens and the expenditures are fine.
While in the other places, the codes can actually do the heuristics and follow the set of steps to then later on push the outputs onto the LLMs, which will take the intelligent actions. Cool. So, what we have summarized so far is that between vertical and horizontal, we don't have to pick one because both of them will survive and will exist and thrive, just like how even after the success of Apple's iPhone and then Android launched, people said like, "Not sure whether this will take off." And today, Android is such a big ecosystem. And same thing happened between AWS and, you know, a bunch of things which came later on like Azure and Google Cloud and so on. Everybody coexist and everybody has their own specialty. Similarly, I I think the AI discipline will get in and all of these questions that we were discussing earlier around cost management, observability, eval's are all becoming very very important.
Before we go ahead, I wanted to ask you, Vijay, that from a GA Ventures perspective and as heading investments, can you talk a little bit about our fund and how are we helping these entrepreneurs, domain experts, uh you know, build and fund their companies? An important thing that we bring together is is talent and then, you know, engineering and product and also capital. And that's important because we want to launch companies that have working product, leadership in place, an initial team sometimes, and and cash in the bank. So, they have some some runway. And that means, you know, we're typically investing, you know, up to uh $250,000 in each company uh when we uh incubate them in in a couple waves. And that establishes us as early investors who can get, you know, some meaningful ownership.
Not too aggressive, but I I think among the best terms that I've seen in the venture landscape. So, that is how we want to support and invest in these companies, and that's why our startup fund continues to grow in supporting, you know, the entrepreneurs that are out there and that we get to partner with. Well, so let's let's go to the next part of the discussion. I just have two more topics to discuss. This is Kushal, this is especially a question for you that we have seen a shift from chatbots and especially the infamous co-pilots to agents and, you know, agentic systems. We ourselves, right right from the beginning, we were very focused on action-oriented like agents instead of just, you know, answering questions.
Um and so, what what's happening now and what do you expect to happen going forward in in AI? Yeah, great question. From the time LLMs have launched, it it had it has been mostly a chatbot where people can go ask questions and get their answers. Over time, it first started hitting the text space where coders who used to like write lines of code sitting and just coding an API for three, four days now can just do it in a couple of minutes. So, it transcended onto those places where things used to be written in a similar way like Q&A here, I just ask a cursor or Claude code to write something for me and it does. While LLMs itself used to do the Q&As, now it has shifted onto agentic systems where it's it's it's not just answering to a question.
It's more of also looking at different aspects of it, wherein the room for error is very less. So, agents are expected to see multiple things, look at what's running, if required, do internet search, or see another code base, and then generate something. Now, it is going into actionables as well, like we already discussed. Well, it can go and run some APIs, based on the API outcomes, decide on what to do later. So, agents take the decisions by themselves, based on how they are built. I believe the next road to this would be a whole system where everything is end-to-end.
Like, agent does something, and then there's another agent learning from the actions taken, see what are the mistakes, and also there's a security agent which is in place to see where is the draft, and if there's any PII data which was looked at, or if there was some there's room for prompt injection, or some data leakage. And also, yeah, telling the agent that this was the mistake, now learn from it. So, I believe there would be an autonomous system where the agents are going to keep learning by itself, and this will be by agents talking to each other on what performed well, and what really didn't go well. Now, it has to learn. I think we are seeing this thing in a very decent way already, wherein the initial agents built are usually not that good. As and when you're making it learn from the actions taken, it gets better.
This is happening because agents are communicating with each other, be it payments, or be taking actions. They are seeing what's happening, and learning. How Cloud has come up with skills, skills itself can be learned. In a way, it works like an agent, which is looking at the steps, and based on that, taking a set of actions. So, I believe a self-learning system which works autonomously and also defines the guardrails by itself and ensures the whole systems are very secure. That's where the whole AI landscape is leading towards.
Yeah, that's very interesting and also agree with you. I was reading somewhere that Anthropic has not written a single line of code since December. Their models are writing the next models. Um and to some extent seeing it in couple of other places as well. What might be also interesting is that, you know, each one of you talking about, you know, some of the other work that we are doing at GI Ventures and the companies we are building. I think we spoke earlier a little bit about Fasttrack AI.
Uh Vijay, if you can talk a little bit about what we are doing with Swiggy AI as you are, you know, pretty close to that. And then Kushal, if you can talk about Turtle AI that uh you are focusing a lot on. Yeah, absolutely. Just to to follow up on your earlier point, there are companies like Anthropic that are not writing much code. And then you have Microsoft and Google where maybe 40 or 50% of code is being uh written by um by by AI. And then you go to legacy companies and it's between 5 and 10%.
So, when we start talking about like AI-first versus AI-guided, like this is like what you can see in engineering teams. Now, how do we then think about building ourselves? Well, as AI-first folks, we have Fasttrack uh which you mentioned in the RIA space uh focusing on transitions. You know, a business we we thought about how do we build not just a product at a feature level, but how do we build something that lets people feel confident about what they're they're doing. So, that means having good integrations. It means having ability to establish more trust in the client outcomes.
So, if you're a investment advisor, if if you're a a uh or an IRA, what what you care about is that someone has to really good client experience because when you promise you'll do something and it doesn't happen and you and you thought they were on board and 2 weeks later you're still asking for some information, there's that erosion of trust. So, in any business that is trust-based, you know, you've got to uh you know, either create more of it or you lose it. So, uh that's actually the thing that FastTrack really does. Is it helps these people run their firm and deliver on what's promised to the end client in a way that creates more trust. And in a way that's what all great vertically eye products should do is they should allow people to feel more confident, not less confident, not more anxious about like is this non-deterministic answer going to be a hallucination hallucination that embarrasses me or my team, my department, my company, right? You want confidence.
And you want that to be a be based on foundation of trust. And the same is uh true in many regards for uh Suite AI, which is focusing on this this huge opportunity of uh of lending technology beginning, you know, first with you know, what are the opportunities for, for example, single invoice factoring and things like that where uh the the product has to build trust first in internal systems that uh we can do this step of of underwriting and and and and funding a business in a way that gives us the same results and same confidence in very quickly and also expands the landscape of businesses that we could lend to. So, it's not just an efficiency play. It's it's also, you know, it's speaking to the top line. Um and ultimately there has to be a good experience for people because if it's a bad one, then um they're not interested in you know, what AI uh is involved. They just, you know, see the the real disappointment of something promised and something not earned.
So, the theme of of trust and building that trust is very important in all B2B sales and all enterprise businesses, but it's got to be really part of how we we build it into the DNA of the business from the beginning as we're investing in those companies. Right. Yeah, I think over the last few months we have spoken to a lot of business folks and companies and one thing which we saw was they were they even found it hard to even understand how cloud code works, how to actually use it, and same goes with lot of other platforms where they can automate their processes. Our hypothesis behind Turtle was to give wings to these business folks because each of them work differently and a lot of it is maneuvering across different tools and doing different stuff. So, thought process was what if they can record everything and just drag and drop it. Can can the AI learn from it and do that later for them whenever they want it.
And not just that, can we make it into an end-to-end system which is what we see the world is leading towards where security guardrails, everything is defined, and these people who are the business folks need not worry about the smaller stuff of the issues which could happen otherwise. That has been the biggest, I believe, thing which holds them back, worrying what if something goes out loose. I don't know how to fix it. So, Turtle is our attempt towards ensuring that the whole end-to-end systems are handled and business folks can do anything which they they they want to do or automate their daily routine, their workflows, and keep running it on our on the orchestrator on Turtle. And another, since we discussed about cost management earlier, that's that's a very tricky thing because Open Cloud has been run by a few business folks who found it hard to understand, but when they started using they saw that, "Oh, I'm losing money all of a sudden, and that's a lot of money." So, here we are trying to ensure that the best models are chosen based on what and the complexity of the actions to be taken. So, any video drag and dropped will be created into a skill and business folks can run it on the orchestrator or even schedule it while not worrying about the security aspects or if there's something going wrong, and they can also monitor every runs which are happening.
They can see if there's a drift or if there's something which can be better, and also give feedback so the skill can be self-taught. So, we are giving complete control to these guys to manage their workflows, run anything, or even make the AI think in their personas. So, there's a way to define personas, so they can ensure that, let's say, a payroll processing needs to be done, that needs to be looked at in the lens of an HR because they know the intricacies of it. So, those personas can be defined and workflows can be run in the point of view of those personas just to ensure that the relevant actions are taken and not just in a generic fashion. Turtle is a platform which we believe will open up the doors for business folks to automate their workflows, do any set of actions which they otherwise would be taking a lot of time on, and not worry about the technical complexities of cloud code or any other tools which they have been finding hard to understand. More worried about what else could happen, and will I be able to fix it.
Right. So, bunch of problems that we discussed in the beginning of this episode and with that I think I would like to sort of, you know, wrap up the the discussion today. It was wonderful and you know, like we will continue doing this more often. If today's episode resonated with you, share it with someone building in the space or someone who invests early and, you know, especially if you're a domain expert who knows your industry cold and want to build a company with us, do reach out to us at guide.ventures. There's a form and we review every application. So, do not worry about that and uh another sort of, you know, thing I wanted to just share is that for accredited investors in the United States would like to learn more about our start fund.
Uh you can reach out to us. You can write to Vijay Vijay@g Of course, important to note that all investment is subject to accreditation verification and this podcast is not an offer to sell securities and we are very happy to sort of talk to you if you have interest. So, with that I would like to sort of end this episode. Thank you so much Kushal and Vijay and let's do it more often. Absolutely. Thank you.
Great talk. Thank you.
End of episode · Ep #18