Kane Simms: All right. All right, Ani. Welcome to the show. Appreciate you spending some time with me.
Ani Narayan: Yeah, again, thank you for having me and I love the branding. You got VUX on the back. You got it on the mic. That’s awesome.
Kane Simms: Absolutely. And normally, the mic would have a If I’m sitting at my desk, I’ve got a little yellow thing on the mic. But in the absence of that and in the absence of money on Sand being broke, I have spotted a slightly yellow tinged hoodie. So, there’s got to be a little bit of yellow around somewhere. Nice. So, yeah, cheers with that. Yeah, curious, I mean, there there might be a lot of people listening to this who are familiar with Lyzr. Might be some that are not. So, maybe it’d be a good place to start for you just to kind of give us the whistle-stop tour. What is Lyzr? You know, where does it originate from and what is it that you guys do?
Ani Narayan: Sweet. So, where it originates from actually, so Shiva, who is So, we’re three co-founders. Shiva started this company in August 2023. We were initially a data analyzer product, so hence the word Lyzr. The way it was is you had an SDK. You could just essentially with one line of code, you could be like, “Hey, I’m going to upload my Excel files or any data.” And then ask questions. And you had an like an LLM at at the uh kind of powering this underneath it all, right? And then what I think when Shiva was building this, he realized, “Man, it’s really hard to build this using existing platforms out there.” So, why not build a full stack agent builder at that point in time from scratch. Um so, I think 2023 and 2024 is when we were like, “Hey, we we were a low-code agent builder.” Uh but what’s happened over the last couple of years is today there’s a big need for “Hey, I don’t want to use 20 different platforms to build an agent and take it live into production. And I have multiple use cases where data is sitting in one place. I have I want to use one model or all want hosting of AWS versus, let’s say Google. There are a lot of these nitty-gritties that an enterprise cares about. So, you want a highly modular platform where you can plug-and-play some of these things. So, we decided to kind of build a full-stack agent infrastructure platform. And what does a full-stack agent infrastructure platform mean? When you want to deploy AI, you essentially need LLMs, you need data. You most probably want a deployment layer, which is essentially what we started with. You want a layer to observe what’s happening with all these agents. So, all of that is essentially contained within Lyzr. So, that was 2024, 2025. So, really good for, let’s say, CTOs or CIOs that are like, “Hey, I want to build these agents internally.” Then, over a period of time, I would say over the last year, year and a half, what’s become really hot is I need to manage all these agents that are being built. We It’s like it’s like, you know, like Instagram came out like, “Hey, people had like one or two like posts they’re putting up.” Now, it’s like, “I have all this content in the world. How do I manage it?” So, it’s become like, “Hey, agents have become easy to build. Now, it’s like, how do I manage all these agents?” So, we have something called OpenController, which is also like a platform where I can manage all these agents. But again, great for companies where like, “Hey, I’m building agents on Claude. I’m building agents on, let’s say, Gemini or like platforms like LangChain, CrewAI, Lyzr. I also have coding agents like, let’s say, Cursor. How do I make sure all that agents can all be maintained in one place? I mean, built and I can see the ROI. I can see if agents have gone rogue. I can be like, “Hey, there’s maybe an agent that’s running in loop, which is burning a lot of tokens.” So, I need an observability central control layer. That’s become our latest proposition.
Kane Simms: That makes sense. I’m definitely interested in exploring more about the the side of things and and what makes that successful. But I’ve also been interested in you know, where do you tend to focus or or do you notice any trends if you like in terms of where your clients are focusing as far as what type of agents are they building? Cuz I I I had a quick look on your website, it doesn’t seem as though you have any discrimination between various types of agents in various types of industries, whether it’s customer service or back office or whether it’s whatever, whatever, whatever. So, I wonder if you could share with me a little bit in terms of like far as your customers are concerned, where is it more common to be deploying agents and what are the sort of typical use cases that you’re seeing at the moment?
Ani Narayan: Great question. So, let’s let’s break that down depending on the customer, right? The most common use cases, if I had to give you a TLDR version, is you have sales, marketing, HR, customer support, IT IT ops, procurement, CFO, all functional stuff. Right? So, in marketing it it started off a lot around content generation, but today the most common use cases I want a set of agents that everyone in my organization can use. So, if if Kane’s running VUX, let’s say there are 100 employees, somebody might want to do content generation, somebody might want to be like, “Hey, I want to create emails. I want to do videos.” So, there’s a generation set, there may be also a lead nurturing side and an event side. So, all of that to have individual agents, right? So, let’s call it workbenches, that’s really hot. And people want that for sales, marketing, HR, all of that. So, people usually start off with that use case and then it becomes about moving to critical workflows. So, the other use cases that you’re seeing today is a lot of my information on legacy systems, a lot of information, let’s say, was in the physical, non-digital space. How do I convert that into the digital world? Right? So, I have agents that can let’s call it the OCR agent where which can look at all that data, convert it, and then I it’s like, “Hey, give me one place which is let’s say call it like the central knowledge brain where I can access all the information that’s within the company, and then I can search. I can do a layer of analysis over and above. Uh the third observation that we’re seeing is take any of these tools, again. Take a Claude co-work. Take uh Higgsfield. Take any of these Lovable Take any of these popular products, enterprises are like, “Hey, I want a private version of this.” Which is air-gapped. Which is I don’t want the frontier models to be touching it. I want to use maybe an open weight models, or I want to use uh all of this should kind of be within my environment. So
Kane Simms: And what is the main reason for that? Cuz I’ve spoke to a couple of companies who have essentially you know, using the same APIs, using the the the for example OpenAI’s APIs, just essentially built their own version of ChatGPT internally. And that was largely a cost kind of thing. Like it was cheaper for them to put some upfront cost into the development and pay for the API than it was to pay for a license to use to use ChatGPT Teams for the whole organization type of thing. What are some of the reasons that people have for doing that over and above just aside from cost, which is which is fair enough? What are some of the reasons people might have for doing that type of thing rather than just using Claude or OpenAI or whatever?
Ani Narayan: Good question. I think I mean cost is primary, but I’ll tell you what’s happening. Let’s say you have 100 people in your organization, and I ChatGPT makes me pay, let’s say, $20 to $40 per person. Right? Not everyone’s usage. You have like 5% of the users that are, let’s say, power users. Not everyone is using it at the same amount. So why can’t I figure a way out where I don’t have to pay that flat fee per month, and I pay per usage? So there is a costing angle to this. Second is the frontier models are actually damn good, obviously, right? But all the open weight models have actually caught on in terms of performance and output at 1/10 the cost. So, cost may be a like I would say 70 80% of the motivation. Uh we have customers which are like, “Man, we’re spending, I kid you not, $250,000 on Claude every month.” Uh there’s one customer which is in the FMCG space. They’re spending $3 million a month on Claude. Now, it’s like, “This is not sustainable because we’re still in the phase of I’m building all these agents, where’s my ROI?” So, while that’s happening, can I start switching to frontier models? So, there’s different ways of doing this. There’s concepts like shadow LLM, there’s obviously fine-tuning. Shadow LLM is like, I take repetitive tasks and I will slowly move like these frontier models to like open weight models. So, cost is one, definitely. Second is also flexibility, Kane. Like remember when uh My Tho’s came out with Fable 5 and then the government shut it down because they’re like this whole cybersecurity issue. Um they were like, “Man, like what if I was running on Fable 5 and I had and it’s some like a let’s say a critical workflow, I shouldn’t have to shut down my business or shut down that aspect of my business.” So, they want like almost a router which can switch between any model. Now, if I have to get ChatGPT OpenAI subscription, then you ask me to get a Claude subscription, then a Gemini subscription, it’s just like it’s an overkill, right? So, um give me one give me all these models that should route. If one falls, the other one should be able to be a backup. And then the final one is privacy. Third, everyone’s getting paranoid that uh OpenAI and Anthropic may be using my data to train. I don’t want that to happen. So, they started with virtual private deployments where you have think of this API key that’s connected to these models, but it sits all the data sits on my environment, doesn’t leave. Or it’s like F that. I’m going to go full bare metal and have a like fully enclosed all my models, all my data. It’s almost like me connecting my laptop to a CPU. And using it. So, I’d say those are three main reasons.
Kane Simms: Yeah, interesting. And in in that situation, where does Lyzr fit there then? Is it that, you know, these things can be built in Lyzr? These kind of like, you know, flexible kind of model access, all that kind of stuff. Um or is it that you’re moving more into, you know, you go and build that with like LangChain or whatever, and Lyzr is going to sit in the middle and perform that kind of monitoring observability type of layer. Or is it both?
Ani Narayan: I’m hoping people think of us as When I think about agents, I should think about Lyzr. Or when I think about agents in enterprise, I should think about Lyzr. So, we actually today have seven offerings, but we don’t want to complicate it to the world and be like, “Hey, these are the seven things we offer.” If I had to simplify it, we do agentic workflow automation with an FDE model. Right? So, that’s our number one swim lane. We do that a lot. That’s our bread and butter. So, for that, it it’s a model where hey, we’ll build for you or take the platform and you build on you. But we also do fine-tuning. We also do, let’s say, we have a product just for banking. Uh which we call Madison. But then we have a physical product called Optimus, which is AI and hardware. Right? Then we have a control plane product. Now, they all hit different personas, but if I’m trying to move from, let’s say, uh a 20 million ARR to a 200 million ARR business, you can have one flagship that really blows it out of the park, or it’s like, “Hey, there are I need to figure out multiple streams of revenue. All of these streams of revenue are actually fundamental business problems that people are like it’s a pull right now. The way we just built like a like a nice brand around it and then we’re going to take it like
Kane Simms: Mhm, interesting. Um I I appreciate I kind of interjected when you were in the process of sharing kind of some use case examples. I don’t know if there was any more you wanted to share on that or whether we covered kind of most of it.
Ani Narayan: We could do more for sure. Um So like maybe give you an ex- So one more interesting thing right now that you’re seeing with system integrators which is Accenture, Deloitte, TCS, Wipro. I don’t know if you’ve seen the stock price. Everybody’s stock price is down, right? As soon as the news came out like oh, Claude is getting into service or Claude is going to have its own FDE program. I remember like the stock price of Accenture came down.
Kane Simms: OpenAI as well, to be fair. They’ve got the same sort of offering now, haven’t they? Process
Ani Narayan: Sorry, say that
Kane Simms: OpenAI have got the same offering now in terms of their pro services, acquisitions, and all that kind of stuff. They bought a device, yeah.
Ani Narayan: They and their goal is obviously going to be on like build on my ecosystem, use my models, right? Um what I think the use case that you’re seeing with big large consulting firms is we we work with a lot of large consulting firms. We work with Accenture. Accenture is one of our investors. We work with uh KPMG KPMG, Deloitte, Capgemini, the smaller ones like EXL and RGB. Um we do three things for them, which is really hot right now. One is we build internal solutions for them. So they all want to productize. They want to move from services to products that they can sell to their customers. So we build internal solutions for them. We do joint go-to-market for their customers in terms of agentic workflow automation. And then the third one is we white label our platform. So there’s sell with, sell to, sell through. And so those are the three motions. So we end up having some really cool use cases. One use case that I can tell you for a massive automobile manufacturer was they’re like we spent $30,000 on doing art direction and cinematography on a single day. Um is there a way that we could upload a lot of our car pictures and create like 20, 30 really good photo shoot images that we could use for marketing? So, that’s one use case. There’s another use case that we did for a really large makeup brand where they’re like, “Hey, let’s say Kim Kardashian came out with I don’t know, like an eyebrow cream. And they’re like, “Hey, can I take something that’s trending there and create a marketing campaign that I can push with my products or create a new product quickly and push it in my in emerging markets?” So, you’re seeing that as a marketing use case. Sales use case, which is really hot right now, is obviously like automate my outbound, which we do. But then a big aspect of it is also ABM. Which is “Hey, I want to land and expand. I want to go after these 50 accounts, do the research, send the email, map the org. How do I go about that?” So, um that’s one more use case. Then, what’s the other use cases I can tell you? Um we’re seeing a lot of customer support. And I think from customer support it is, “Hey, are there tickets, let’s say, which is very voice enabled? Um especially front facing. For them it’s a big success if from 100% human led it can move to even 40% AI led.” Uh which is always human in the loop. Oh, we’re seeing some sports use cases as well. Um
Kane Simms: Well, like sports teams, clubs, things like that?
Ani Narayan: Yeah, we were I mean, we we sponsored the San Francisco Unicorns earlier this year. We We met the folks from Miami Autocars. Uh we spoke to the folks from McLaren earlier this year. I mean, you’re you’re I’m assuming in the UK, so I would maybe I’ll tell you more about McLaren. So, their use case or they were saying is which is interesting is there’s not much flexibility on the engineering side cuz it’s only like in the winter and summer they get like 4 weeks and 6 weeks to make those changes and the engineering budget’s actually part of the engineering budget. So, it’s harder to do things there, whereas on the operational side, you want to do things on the marketing, you want to automate my HR workflow, all of that you could do. Uh so, they’re more flexible to work on that. I think engineering will take some time unless they want to build something on their own. Uh
Kane Simms: How How do you kind of um How do you approach I suppose like the What am I trying to say? How do you approach like either demand generation, marketing, that type of thing when you know, the All All of these use cases are like quite vast, aren’t they? It’s almost like the the your imagination is You can build whatever you can imagine, you know what I’m saying you know what I’m saying, which is fantastic. And if you’ve got clients who are saying, “Look, I’ve got all this data of all of the images of all of our sports cars and we want to be able to use that to then create marketing campaigns.” If they come to you with the idea, that’s brilliant. Um and so, how much of it is the clients and prospects come to you with these ideas of stuff that they want to do and they just want to see if it’s possible to do with AI versus how much do you have to try and like inspire them to think broader and to imagine differently because I can imagine that, you know, as you said as we’ve said like your imagination is the biggest limitation here. Um so, how much of that is them coming to you like we want to do do brilliant idea versus you having to go to them and like look, we can do things a lot differently.
Ani Narayan: Yeah. So, it comes down to their know-how about agents and where they are in their kind of knowledge cycle of bringing agents into the organization. I was talking to uh one of the founders of a massive healthcare company. Like we’re talking about they do more than $2 billion a year. And they’re like hey, we’re not tech-savvy. We sell like pharma products. But he was quite clear about hey, maybe there’s one use case that we can start with which is hey, we’re seven companies that keep making these decks. And I want to figure a way out where I should be able to be like hey, I’m meeting this person. This is all the information. Pick and choose and figure out what deck to make. But while making sure it’s private. That’s the thing. The the privacy aspect is so big because they’re highly regulated, they’re public. They will not do this on a cloud. So, um so, he was clear about a use case. But then when I spoke to him and said, okay, this is these are all the use cases that are possible. Hey, we’ve done this in HR. We’ve done this in sales. They go, oh my god, that is cool. Can we do a session where we educate the next layer of leaders on what is possible? And then you can either so, there are two ways in you can help them prioritize. One is you map out and say, you know what, this will have the maximum ROI. Like and and the other one is you actually crowdsource all the ideas and then you use a tool like what we have called architect where you can put all these ideas in and because we’ve done a lot of these implementations on the back end, it will show you hey, this is a high, medium, low priority because this will have high business impact, but oh maybe it’s easier to implement, it’s not that expensive. Or this is massive tech debt. So, we were talking to a huge I can actually tell you this. We’re talking to Old Mutual from South Africa. They came to the office like a couple of weeks back. Big, large, 200-year-old insurance company. They realized that they said 5% of the GDP is spent on funeral services. And they wanted to make sure they the agentic side was not going to be on here as soon as someone passes away. It’s like we should speak to an agent. It should be to speak to a human. But everything behind the scenes, behind the concierge person, can we make that agentic? Where getting this information, filling that out, uh getting a loan if needed. So, all of that, uh so I think there’s always a layer of education, which they’re happy to We’re still in that era, you can say. And then you can have multiple use cases which can go on in parallel as well.
Kane Simms: Yeah. And what what are your thoughts on I think that the example you gave there of the insurance company wants to keep the human agent front and center, but wants to automate the back office side of things. I can imagine that being a type of use case where very specific, very dependent on their data and their systems and their process, and therefore some kind of custom agent would be absolutely required. I’m curious to sort of get your thoughts on where you think businesses sort of will go for for certain types of use cases. So, I’ll give you an example. Um ChatGPT this week, I literally We just came back from CCW last week, and as a bit of an experiment, I thought, I want to just give ChatGPT all of the videos that we took. We took about, you know, five different interviews, and and some sort of like medium-form interviews, like 10 minutes each. And I thought, I’m not going to put them I’m not going I’m going team to do this. I’m I’m not going to do it. I’m not going to get it into Premiere Pro and start chopping it up. I’m going to just give put them in a folder, give it a ChatGPT, tell it what the what we did, what we want, and all this kind of stuff. And I thought, I don’t think it’s going to be able to do it, you know? I thought there’s no way that ChatGPT is going to be able to just edit all these videos. Anyway, it did. It took about 15 videos, top and tailed them. So, top top took the beginning off, took the end off, made it tight so that that’s just the content, put lower thirds on it, and then translated it into clips. So, I was like, “Right, give me like, you know, three clips from each from each interview, portrait, captions burnt on. And if anyone’s moving like this, I want you to track them.” You know what I mean? And and make it seem as though the camera’s following them. And I thought there’s no way it’s going to be able to do that. And sooner or later, you know, literally 5 minutes later, I’ve got like 32 different clips from these interviews. And I’ve got captions up for for TikTok and Instagram and LinkedIn, and I’ve got thumbnails for YouTube, and I’ve got everything. And it literally took me, I think probably maybe elapsed time, half an hour. So, it’s like for something like that, it that’s a perfect use case, brilliantly executed. Obviously, the question is how do you make that repeatable and templateable so that anyone can else can sort of do it. But the reason why I’m sharing that story is I kind of want to get your perspectives on for something like that, you know, I didn’t need to go and build my own custom agent. ChatGPT has that capability natively. But for this insurance company, which is very specific specifically tied to their data and their systems and what triggers the agent is going to be unique to them as well. So, that seems to be obviously a custom build. I just wonder if you have any thoughts in terms of where businesses can land, do land, or should land as it relates to these different types of use cases. Like what’s worth building custom versus when might you may as well just use the frontier models that are out there already.
Ani Narayan: Yeah. I mean, it’s fascinating to hear what you said. I’m almost like I need to go to my video team like dude, let’s try this now.
Kane Simms: Oh, 100% if you’re you need to get them on it because it is honestly I’ve we’ve got more content from last week than we can do anything with. It’s crazy. Yeah.
Ani Narayan: Nice. Uh because they’re like always going to take 2 weeks. We have a dedicated video editor. I’m just going to go and be like guys, I heard from Kane that this is possible in 30 2 minutes. It’s
Kane Simms: And I’ll I’ll I’ll tell you also what is what is even better than that. I didn’t ask for this. It just did it as part of what it did. It built me a HTML page in our branding, yellow banners, all of our fonts. And on that page, it embedded all of the videos that it made. So, it had the first interview. I just press play. I can watch the whole interview. Next to it, I can just press the button to expand on the description for YouTube, timestamps for Spotify, everything. And then underneath the video, it’s got the four portrait videos that it’s cut from that with all the captions on it and stuff like that. Titles, captions, all that underneath that. And it just literally on its own without me asking it, it just built this HTML page for me to proof everything from. Unbelievable. Yeah. Yeah. Anyway. We were digressing.
Ani Narayan: But no, no, I think that’s that’s awesome. I think your point though, Kane, is so you build custom tooling when you’re dealing with multiple systems and you need custom code to connect some of these things. Right? Like if I have let’s say a legacy system which is built in the ’80s or ’90s. Let’s say especially like mainframes. There’s some let’s say there’s Tally which is in retail. It’s not easy to have like today when you’re using Claude Co-work, I just need to authorize my username and password for HubSpot or a Slack and then I am connected to that tool. Whereas it’s not easy to do that in the enterprise side at all. AI is a layer that’s being built on top of legacy. So, there’s also like where is all my data? Who are all the people that I need to speak to? It’s a massive ecosystem. So, in that sense, it makes sense to do custom. Mhm. And what your example, what you did, you should if you’re an SMB startup and you’re you want to get stuff done I mean massive cycles of uh you want to save time in in terms of cycles like this, you should use a tool. You should find as many agent SaaS tools um and then like not hire as many people. I think like that’s why you’re seeing this where you have 30 people organization which are worth billion dollars, right? Which is It’s a $19 product. It’s a $99 product. So, um Yeah. So, it comes but all these players when they go to larger legacy systems, they’ll have to build custom.
Kane Simms: Yeah. Yeah, interesting. And I suppose that as I said, you know, like if it’s just me and three other team members that are doing it, it’s different to having an organization where your marketing team is, you know, 150 people and all of them need to be able to use it with the same level of consistency, same quality of output. Uh so, there’s a lot of other stuff to consider there, I think, as well.
Ani Narayan: Actually, one more thing, Kane, which might be interesting is see you’ll you’ll realize this about Claude, it doesn’t have sustained long-term memory when you cross multiple threads. And what if Kane wants to use this across three other people and they all want to collaborate on the same thing? That’s hard.
Kane Simms: Yeah, it is, yeah. I mean, ChatGPT is actually a bit better than Claude at that. It will pick up on thing I think it’s a bit better at getting stuff into its memory, but sharing that with the rest of the organization is I think definitely a bottleneck. Like, how do I know that John was working on this and that Suzy was working on that and actually there’s a lot of shared kind of like information and knowledge that’s being sort of done there or they’re both working on the same thing or whatever. Yeah, it’s it’s interesting. I’m curious into kind of alluded to it at the beginning in terms of how you approach this kind of observability and monitoring type of work because, you know, the picture you described at the beginning where someone’s using, you know, LangChain and someone’s using this and that and they all got different models behind the scenes. Some of them might even have built some stuff in Lyzr and stuff like that and it’s like how do you observe what’s going on? How do you spot rogue agents and and this that and the other? So, what are the sort of key things to consider, you know, as you’re deploying these agents? What are the key things to consider as far as observability is concerned?
Ani Narayan: I’m going to do it. Can I share my screen and maybe show you something or is or is it going to be an audio predominantly?
Kane Simms: Uh it is audio, but there are there is also a video. If I just allow that, should work for you now.
Ani Narayan: So, one is this, right? Like when you look at this, this is what Okay, so one is the how, right? Now, what are we tracking? When you talk about observability, you’re tracking You should be able to first of all figure out who are all the players that you want to track, right? So, there’s there’s models, there’s platforms, there’s frameworks, there’s coding agents, and then there’s agents that you build. So, Databricks, there may be data coming in from there, but there’s also agents which are sitting within Databricks or Salesforce or SAP. So, you need one kind of layer to bring all of that in. So, that’s the first part, right? So, identifying and coding agents is there’s token consumption on Claude or Cortex or Cursor or GitHub Copilot. How do I make sure all of that’s happening? Now, we have built something called a control plane where imagine you have a readme file that reads through all of this, reads through all of these agents, and let me see if let me show you the next one, which is this. Right? We call it the control plane. But basically what we have done is you have, let’s say, like a read me that follows it discovers all the agents that are out there. And once it discovers the agents, it converts it into a specific format. Now, that’s the important part, Kane, like imagine um these are all players. Right, these are all basketball players. You’re measuring this Let’s say somebody’s height is measured in centimeters, somebody’s height is in like feet and inches. The registry makes sure all of it’s in one format. Everybody is in feet and inches format. Everybody’s in kgs instead of pounds, right? But for agents, that’s So, what are you capturing? You’re capturing uh role, description, soul, what is the what are the tools it’s connected to, uh what are its features, all of that’s done. Then the next layer kicks in in terms of your identifying what these agents do, what are their activities like. And then there’s evaluation. So, they you you run these evaluations on these agents. Uh and once you’re then you build the platform from there, which is how do you how do I make When I say running evaluations, it is I can have test cases and I can be like, “Hey, this is the current output. Against multiple test cases, is the agent still reliable?” Right? And then from there I kind of move to how do I take this into a live environment? How do I make sure this is deployed? The rest of the stuff is more once it’s live, like oh, agent gateway is not everybody should have access to the same sort of agents or there’s limit limited access. LLM gateway might be is like, oh, make sure there’s Remember I told you like if Anthropic fails, go to OpenAI. That’s gateway deciding. You can for multiple tasks. You can decide maybe there’s certain reasoning tasks that need better models. Maybe simpler simple tasks need small models, so you could do that. And then it’s making sure there are guardrails and observability. And this the observability is basically telling you what all is available on the public side, internal, and private side. So, some of this gets even technical for me, to be honest. Uh but it starts telling you what’s happening at the app level, at the user level, at the agent level. Uh so, that’s that’s the big observability play. I think a lot of it first is just making sure Let me see if I can give you another good image which could be interesting.
Kane Simms: I’ll stick this link. I’ve I’ve got actually got this web page up in here, actually. I’ll stick I’ll stick this link into the uh into the show notes cuz I think it does a good job of of kind of visualizing uh all those different all those different components. Um L- Fi- final question before we sort of wrap up. Wha- What are the in your experience in terms of, you know, the customers and prospects you speak to, what are some of the things that businesses are getting right today as far as their deployment of of AI agents, and agentic AI, and what are some of the things that they are getting wrong that should change?
Ani Narayan: Even us, I would say what we are getting right is we are making sure we’re at least us, I can speak for I don’t know enough about them, but everyone’s doing the FDE model. Which is it’s not like, “Hey, here’s a platform and LEGO blocks you build on your It’s not It’s not IKEA style. It’s Bro, I’ll come to your house. I’ll fix all the nuts and bolts. Here’s a chair. Sit on it. Is it working, right? So, um players that are owning the outcomes with your customer are winning. So, they’re doing that right. Um The ones that are doing it wrong, the ones that are doing it wrong, I feel are building products where Claude can very easily start doing what you’re doing. You know, like especially if it’s a $20 product. Because even if let’s say I’m building an agentic workbench, which is like multiple agents. Today, it’s like I find it difficult to tell people to get out of Claude if everybody has a Claude subscription. If they all have Claude subscription, it’s multi-usage, as you know, like they’re using a videos, content generation, you name it. So, uh The smart thing that you should do is actually move to like a fully even other players, actually Harvey is doing this very well. Move to a very verticalized play, like build a brand around a specific industry, build a wedge, which is really capturing the nuance of a workflow. Be a le- if you’re a legal person talking to another partner or an associate, more resonance and I want to buy from you. Like Harvey, for example, has 350 legal product managers. Right? So, like so, they’re doing much better than OpenAI coming out and saying, “Hey, I have a legal offer.”
Kane Simms: Yeah, absolutely. Yeah. What about the the end clients as well then? So, you know, of the brands that you work with that are deploying agentic AI in their organizations, what are they getting right and what what do they need to to improve on, do you think?
Ani Narayan: So, all the service players that I know uh productizing some of their workflows, converting it into a brand, and selling it to their existing distribution. That’s that’s super super smart. I think everyone’s doing that well. Uh every large consulting firm is doing that well. So, I’m still surprised. I feel like some of these consulting firms are undervalued in my opinion as of today as a result. Um What are they not doing well? the experience needs to be agentic end to end. I think some of them aren’t making sure like you know if I showed you a SaaS dashboard and I said it’s agentic you’d be like bro, how is how is it different from like Claude experience gives It’s like it’s nothing, right? It’s a simple voice input. It shows me how the agents are working, then gives me an output. But if you give me a dashboard where I have to click on things, people feel like I’m regressing in some sense. So, you want to make you want to create the experience as much as what Claude is doing. One thing though which I’m still surprised that Claude hasn’t done, Kane, is why is my output not multimodal? Why do I not get Like if I’m learning something, it should give me a YouTube video with that specific indexed part that I can go watch the video directly.
Kane Simms: Yeah. I mean OpenAI’s kind of launched something a bit like I forget what they called it now, but um they sort of announced something a bit more like that where like you know if you were asking, you know, how does what’s what’s the difference between a muscle and a tendon? Then it creates like an animation, it shows you how it all connects together and stuff like that. That’s a that’s a great kind of feature that like.
Ani Narayan: Beautiful. Somebody did that show it on an Astra 5.2, I think, which is all the bones and muscles in the body, and then you could like kind of move it in three dimensions, click on it, see what your muscle stands for. Yeah. That was sick. That was really beautiful.
Kane Simms: Yeah, it is. Yeah, yeah. The whole kind of concept of the generative interface I’m kind of really interested in. Um and also how that you know, how how do you kind of build them generative interface type applications and and when how how what type of UI components do you use at what point in the experience? I think is a is an interesting sort of question. Um but yeah, no, this has been brilliant. I really appreciate it. I think it’s um yeah, it’s really interesting to We spent a lot of time talking about kind of customer experience and and conversational-based sort of uh AI. I think this conversation’s been really interesting to get a look at the other side in terms of, you know, how a business is approaching the harder challenge really, which is automating the back office processes, you know, cuz that’s where a lot of the the challenges are. Data, as you mentioned, legacy systems, stuff like that. And also the things that they have to do manually, like the video editing and like all that kind of stuff. Uh how AI is being used to sort of improve that. So, it’s been been really interesting conversation. Appreciate it.
Ani Narayan: Yeah. And and since you since you run a podcast, right, Kane? Like I think there’s such a good offering here, which is imagine I’m on Slack and I’m like, “Hey, I Kane, it’s like hey, what other podcasts should I be looking to get into?” And it like finds all that information, automatically sends the email, like, “Hey, this is good. I want to come live.” Schedule like what if the entire podcasting workflow productized and then you give it out to people as a as a flat fee, right? Because is better than I mean, is they might be in That might be an interesting space for you to explore also.
Kane Simms: Yeah. Yeah. I think that’s a good idea. And and also the the things that I’m sort of working at the moment is once this podcast is done, I’d be I’d worked a lot in the last few years of getting it to a point where I can start a podcast in 5 minutes. So, for this podcast, literally, it was 5 to 4. I’ve got the cameras are on rigs that go up to the ceiling. I pull the cameras down, turn them on, open Zoom, and all the gear is ready to go. I I press go and it’s done. So, it literally takes 5 minutes to set up. I’ve kind of tried to lower the barrier or the time it takes to start. What I’m sort of working on now is decreasing the time it takes to finish. So, once this once you press stop, what happens then? You know, get the asset, get the audio, edit the audio, edit the videos, chop up the short clips, get it all distributed, all that sort of back off back end stuff. Then I think there’s there’s a quite a lot of merit to, you know, like I mean, it’s something that we’ve been trying and not not with any degree of intensity, but something that’s been sort of on the on the radar for a while is how to streamline that workflow. And I think there’s value in that. I think the only thing is with stuff like that is that, you know, I can imagine that that’s this tools like Munch and these other sort of tools that do those like identify the clips and all that kind of stuff. So, there’s already sort of things out there that do a bit of that. But when you package it, I think as a podcast end-to-end workflow, I think that’s where it gets quite interesting, you know. You could almost have it as a you could have potentially even host the podcast through it, you know.
Ani Narayan: The podcast hosting and also even the setup, right? Like what you have as a mic and this background and stuff, like now it’s almost podcast is a good way to the way I look at it is it’s high effort, but you can really squeeze the juice on repurposing. The the way we use the podcast is we record these clips, amplify using ads, right? For like specific audiences. Even if you were like enterprises like us might or not So, that’s a startup like us might be like, “Dude, I need someone to help me sell this space.” So, it end-to-end could also be like the physical experience, too.
Kane Simms: Yeah, yeah. That’s true, that. That’s true. Yeah, yeah, yeah. Good point, yeah. Nice one. Nice one. Well, I appreciate it. Uh it’s been really good, hasn’t it? Yeah, it’s been really interesting. Uh we’ll stick that link to the OpenController and the Lyzr website uh in the show notes. We’ll put your LinkedIn there as well if people want to reach out to you and connect. And uh yeah, definitely keen to to keep in touch and and see where this kind of stuff goes. I think you’ve got a really interesting proposition there. Um very diverse in terms of the kind of offering. So, it’s kind of interesting you know, you’re kind of involved in in all parts of AI really, which is really interesting. So, yeah, be definitely keen to to see how it goes.
Ani Narayan: Yeah, and thank you for having me. That was really kind of you to to schedule time with us. And yeah, happy to help, happy to push push the word out whenever you need help.
Kane Simms: Nice one. Appreciate it. Thank you so much, and thank you all for tuning in. We’ll see you again on the next one. Thanks now. Bye.
