Do you sometimes think about that if you were to build a Meesho again today in this environment, how different would it be? Why are the next 500 million people not online? I know accessibility is a big problem and that's what I will go and solve, which is what we are doing. So we launched a new product called Vani. 75% of all our orders today come from recommendations and not from search. We started with something harder which is voice right, like voice has a lot of other problems that chat does not. But we said this is the vision that we imagine of what Meesho in 2030 should look like. Particularly on the one side and the conversational voice transaction. Are you starting to see gains beyond activation? Are you starting to see better conversion?
Consumer AI is right now out of fashion. I think it'll come in fashion maybe in the next 2-3 years when token cost goes to zero. What role should we play as a company? We have to play a role of an enabler where a lot of people want to move to whatever is happening on the frontier. You enable them with giving the right tools, the right technology. Everyone feels it every day that you're staying behind in AI. And if you stay behind in AI, if you're a technology company, which we are, things can go wrong in many ways.
Hi everyone. I'm Akrit and we are back today on Signal Dialogues on our next episode where we talk about what's happening in AI in India. So far, for those who've been following what we do, most of my conversations have been with what we would call native AI founders. So we've spent a lot of time with various AI companies both across the US and India. Recently, I think it was 6 weeks back or so, correct me on the timeline, I think six or eight weeks back, along with their quarterly earnings, this e-commerce company that I am today, Meesho, separately spoke about all their work in AI. And I started reading it and I started going down the rabbit hole and I realized that the definition of what a native AI company today is evolving. And there is enough reason for so many of us to think that a company like Meesho is now evolving into a native AI company. So that's what brings me here today. And I was very excited when I reached out to a friend Vidit, co-founder and CEO of Meesho, to say that everybody knows you obviously as your core e-commerce business, but incredible work that you've been doing within this new technology paradigm, and would love to have a chat and share with everybody this entire journey.
Thank you, thank you Akrit, glad to be doing this. And yeah, I think we've spoken about AI a couple of times before, but good to kind of do this chat together and learn. I also, for the past few months, you've been sending me some of the things that you see because you see a lot of new age companies, and I've been also finding patterns around AI. I think that's the same thing that we've been doing internally. We have a lot of ideas come up from all people within the company around how we can do different stuff for the customer, for the seller, for our logistics partner. How can we become very different teams? So we may not be an AI native company, but I think we are definitely on the path of getting there. And I sometimes think that it's much easier to be an entrepreneur who's just starting up in an era like this because then it's much easier to change. I think for us, we have to unlearn a lot of stuff and learn, but I think the entire company so far has been finding it a lot of fun. I think it has given us a lot of purpose and a lot of excitement. A lot of things which for a lot of times seemed very hard to solve now look solvable. So I think it's something that's very exciting for us.
Yeah. Amazing. So let me get to the one question I've been thinking about ever since we've been chatting offline as well. Do you think of Meesho as a company and what it will become differently now in the post-AI world compared to how you've been thinking about it for all these years?
I think so. Basically, what we did: we went public in December 2025, about 7 months ago, it's not too long ago. But just 2 months after that, the first thing we did because I always wanted people to have sight on the future rather than what's in the past. Everyone was celebrating. The first thing that everyone in the company did is we set a five-year vision for every team, every product, for the entire company. And like what something that I expected actually happened: AI was the center of most the vision that most teams created. So everyone now has a very strong shared vision, shared understanding of what Meesho would look like in an AI native world, and we are actually now making that change. In my head, there's a very clear vision of what that would entail: what do we need to change as a culture, as a team, as a product, as a business, and we are actually on that path. I think some things will work, some things may fail a few times before that work, but I think if you just think about long-term vision and clarity, I think we have as a company.
It's amazing, you know, fresh start in public and in this new technology world today. Do you sometimes think about, because you just said that, and I ask every founder at your stage this question: do you sometimes think about that if you were to build Meesho again today in this environment, how different would it be?
I think it'll be very different. So the first thing which is very obvious, it will be AI native. So I'll maybe start with one story. When we started the company in 2015, we started to solve problems of those days with the technology available in that day. So when we started, the big problem was e-commerce was limited to big cities and was limited to branded fashion, branded electronics, period, nothing else. So we said let's just go deep into the problem and understand what's happening. The biggest problem was most people used to have one app on their phone and that was WhatsApp, because the cost of data was very high, no one wanted to download these shopping apps with heavy images that consume a lot of data. So we created a way for people to start buying on WhatsApp, and for the first time people across the country who never bought anything online started to buy online. So leverage the technology of that day in solving a problem. Right? We are actually now doing the exact same thing but solving a similar problem today. Like when we now go to rural areas, we see that our penetration there is still very, very low. Actually, it's almost untouched yet. And when we go deep, the biggest problems are again problems of accessibility. The same way we had to solve a problem in 2015 of accessibility. A lot of these customers really struggle to read on the app, to type on the app, to understand what clicking a particular button means on the app.
Imagine if you're voicing at that time.
Yes. Like we could have just done a very different thing. Right? So now you think, now finally we have AI, we have voice AI available in the native language of that person. Yes, I don't need to force them to read on the app, I don't need to force them to write a few keywords that come to their head when they think about a product. None of that needs to be there anymore. So we can basically, as if we start the company today, I would just do the same thing: I will just go out and say, okay, why are the next 500 million people not online? I know accessibility is a big problem, and that's what I will go and solve, which is what we are doing. So we launched a new product called Vani, and it came basically exactly from this insight. Every single time you go and speak to rural customers, accessibility is still as big a problem as it was in 2015 for mass India. So we are very confident that the future of using internet all kinds of internet services for mass India, 10 years from now, is going to be voice.
Do you know, it's so fascinating, right? And that's why I really wanted to have this conversation. Do you think that because what you do as a business, do you think that AI and the timing and where it's come is a massive acceleration? And let me ask the reverse: how are you thinking about it if this didn't happen?
No, which is what: you use the technology available to you to solve these problems. Right? So far we were doing, I would say, very incremental stuff. We were going and trying to find colloquial language related to e-commerce and shopping and putting those words and going local, changing the product photograph so it's more relatable. Every time you do any of these experiments, you'll see some conversion, but it's not the same. Like now you go, I can just leapfrog from offline to directly voice e-commerce for a customer base who has not seen anything else. You don't need to do these incremental gains that we were trying to do because some of these technologies were not available. So you're right, if this was not there, maybe what we will now do in the next 2 years we would have done in the next 10 years.
Right, like you never know.
The good thing about technology is that it surprises you. It just keeps coming, and if you are on the cutting edge and you are experimenting and you are exploring enough, you can keep finding this, and then as you said, you can just keep changing the inflection on the graph. And we have seen that. For example, the first version of Vani was very crude. The first version, because it was not an ideal phone, all our systems like search system, order system on the app is still not built in a way that AI agents can do everything. So it's very crude. V1, the first time we launched this product with that rural customer, we saw activation go by 30-40%, which is crazy. And we're like, with a sub-optimal product we see such gains. When we really improve this product with everything possible, actually the next 500 million people I think will come online much, much faster. So yeah, I think that's why you asked me the question if I would start today what will I do. I think I will basically do the same thing: I will go and solve the same problem. I just feel that we're lucky to have more options in terms of what we can go out and solve and how we will solve them. So that's why we have a lot of fun stuff to do, which is what we are trying to solve. And I think this is just the beginning. It just gets sweeter and sweeter from here.
Amazing man. I can sense the excitement. And we've obviously been chatting, but separately, given the work that I do and we do at Activate today, we've also spoken with a lot of your leadership and I can sense the same degree of excitement whether it is Sedat and Dud and Sanjiv and a bunch of others. Now let's go to the next level where I wanted to use the remaining time to actually...
We've all, I think every one of us has been to hundreds of conferences, seminars and events and even interviews which are very surface level about what people are doing with AI. I think you guys, in what we observe, are very deep. So let's first start with on the customer experience side. You spoke about Vani which is the voice engine to be able to transact end to end. Tell us something more what else is happening on the customer side with AI that wasn't possible before. I'll just start with the journey. I think we over time have traveled a lot of what most people have done. Our first product was use it in customer support, which is what we scaled first. We have a lot of users, so that's a large problem for us. So we went out...
That now is like some serious scale and outcome. Like I think the last time I remember you said close to 50% plus is...
Now I think it's almost all. I last saw it was like 70% plus of all our customer interactions are AI interactions. And that number continues to increase. So I think that's now scaled up. But that's the first thing that we built. And because of that, we became a sizable user for a lot of voice products globally. So that's what we started with. We did that, we learned something, and then we started to use it in marketing. You look at creatives across the board. The beautiful thing is most of these technologies work really well for us because we have a very heterogeneous base, a user base who speaks lots of languages, have very different kinds of needs. So every single time we use AI in marketing, the biggest use cases you can now personalize your creative for everyone. It's amazing for us. We don't have a cookie cutter customer base in a big city. We have very different kinds of customers. So we started to use it in marketing. We've seen a lot of gains there. We started to use it internally for all interactions with our sellers, our logistics partners. We use a lot of our fraud models that historically used to work on ML models but were never as effective. Because a customer wants to return a product which is different than what was delivered to them. All your image models earlier, people could figure out over time how to game you. Use some of these frontier models, you can really say that the customer is trying to game the system and basically return something. So we've seen lots of gains in all our trust systems like fraud control end to end. And we again continue to use more and more. So in general, big focus was we run a lot of operations, a lot of touch points. Can you really bring efficiency? Just go out and bring efficiency in your operations. So customer support, marketing, I would say fall in that bucket where we are not trying to basically reimagine the experience as much as we were doing something, can we do it better? I think the second stage came in where we said can we do things that we could not do before. And I would say one was not the first thing. The first thing that we've been trying to do is again, we are a large platform, hundreds of millions of products, very different kind of consumers. These products are uploaded by businesses of all size and scale. Some of them are tech-savvy, some of them are not. They don't know how to put attributes. So I think the biggest use of AI has been in our recommendation systems.
Our cataloging systems, all the systems that we run at scale. Like when you come to the app, you start to basically click on something.
Have you seen that? Because we were talking to somebody else about this. Have you seen that recommendations have gotten significantly better with LLMs compared to you've been using recommendation systems ever since?
So at different stages. One is just getting attributes from products, we use LLMs. Then even the architecture of all of our recommendation systems in the company are all transformer based today. So even if you feed the same cataloging information, you get benefits. But now you put the new information that comes a lot from LLMs that are much better. So I think across the board, and I think we put out in the same letter that you were talking about, 75% of all our orders today come from recommendations and not from search. And the reason that happens is that over time as we start to learn about a customer, we start to see very, very relevant recommendations. It's fascinating.
Which is quite interesting and again something that goes back to how you guys are thinking of very first principles.
You have not heard chatbot yet?
No, we've not heard that. The first thing I did was I opened this chatbot for people to ask what should I buy, what can you know, the classic like can I get a red colored shirt and give me some answer. Because that feels like a very incremental, frivolous problem to solve.
Absolutely. And I tell you why. In our case, even earlier, a large percentage of our users would never search. Less than 50% of searches were text searches. More than 50% searches are either voice or image search. So basically they take a photograph and say give me something like this because they cannot use words to articulate what they want. So we knew that this is not a text native customer anyways. So you have to solve this problem. That's why our first chatbot was Vani, which is voice only. Exactly. It was not like voice or chat. We were very clear that if you have to build something that scales for a customer, I think that's true for anyone, because that chat interface is not very native for shopping anyways. A lot of people have tried and failed. But I would say even otherwise. So I think that was very clear to us. We don't want to build something that definitely fails. So we never ever built a chatbot. It's not like we built something and it failed. We never built a chatbot because we were very sure that for the customer we want to build, I don't think that's the right answer. We started with something harder which is voice. Voice has a lot of other problems that chat does not have: latency, you have people speak very different kinds of languages, the models are still... Also people lose a little bit of discipline versus text, they speak longer, they speak more unstructured. When you have to type, it's effort, so you are generally more structured. But when you speak, you're very unstructured. So we knew that it's a harder problem, but we said this is the vision that we imagine of what Meesho in 2030 should look like. The vision is that in the long run, a customer should never have to read, type or click a button. They will basically speak to an AI and they will start to see only visual information, maybe images, maybe other information that's required to move forward, make payments, everything through voice. And how do people move to that experience is our biggest focus area.
Are you yet, I'm guessing it's early, but particularly on the one side and the conversational voice transaction, are you starting to see gains beyond activation? Are you starting to see better conversion?
So actually we haven't scaled it beyond activation yet. I'll tell you, it's multiple reasons. First is cost. The token cost on server on activation, you anyways are spending on that customer. There's a customer cost of acquisition, your CAC, so that can justify like I'm going to use tokens for this, it's fine. And as long as a better conversion makes up for it. But if the customer starts to come and use it again and again, your cost start to go up and you have to get a very strong payoff for it to make sense. So if we go out and just scale this product, it is not viable. We cannot scale this to everyone. I think the second thing is as the customer speaks to this agent again and again, we realize the customer expects that you remember everything. So as you remember everything, this customer will say whatever I showed you yesterday or whatever I searched yesterday, can you show something like that at a 100 rupees lower price point? We don't still have a great infrastructure around memory so that the retention of that is as good. So I think we're still on that journey. I would say it was the easiest for a customer in the activation stage. That person has never used most on our platform. Someone who is pre-activation most likely has never used any e-commerce before ever. So for them to be introduced to a voice interface that may not be the most optimized yet, but they don't know anything else. So they don't expect you to be really fast in responding. But for someone who's using the app, if you try to use any voice app, you will realize you have to keep waiting for the response from this agent and it just feels inefficient. People have much higher expectations on latency, much higher expectations on context understanding. So I think that would require some time to happen, but we believe it will happen over time. I think we all think about the cost. Another thing that we are investing in heavily is building more edgy AI experiences. So we're experimenting with all kinds of models. We're experimenting with open source models. So I think we have to figure that out.
Do you have a dedicated AI R&D team?
We do. We do have a team. That team is growing now. We still do not train LLMs in house. So we do a lot of infrastructure work, but we don't go out and train LLMs in house. I don't know, maybe I think that would change at some point in time where some of the problems we have cannot be solved elsewhere. But I think a lot of pieces that you need to build in the stack we'll have to build it ourselves. That's why a lot of infrastructure that is required for this to scale, like around having context, having memory, this agentic marketplace so that an agent exists for all the problems that a customer has over time, I think we're building that out. So I think over time that will change. I am very confident that cost will get solved at some point of time. The token cost will get to zero at some point in time. These models may not be as good as the frontier model, but they'll be 80% there and that would be good enough, but it'll be free for that customer. So that customer can come and use it more and more. If you see globally, no one has been able to offer a subscription-free consumer experience even in the western markets because these token costs compound like crazy.
It just, if you work backwards from the bottom of the stack, that's the problem. It starts from chip and goes through everything.
Absolutely. So I don't think it's possible for you to run this on the cloud and offer this for free. And I don't think we'll have to go and offer a subscription experience. Hey, you pay us per month to use voice. I don't think for a market like India it's ever going to escape your business. So I think AI is the only answer. I think it will happen at some point in time. I was hoping Apple would do something now because they have the best chip in general on phone. It did not happen this year, but I'm quite confident if not this year, next year or the year after. But AI will come and that's when consumer AI will start to scale. If you see globally, AI right now is all enterprise.
All revenue is enterprise though. The technology is so powerful and so many people can do so many things, but I don't think as much is happening in consumer.
Yeah, it's all. The only consumer AI revenue is whatever ChatGPT, and they're also not happy about it anymore. So it's not like it's Google kind of revenue with 90% margin. So consumer AI is right now out of fashion. I think it'll come in fashion maybe in the next 2-3 years when token cost goes to zero. At some point in time, token cost goes to zero and suddenly we are trying to do it for commerce. This customer who does not use any kind of internet services today because accessibility is a big problem. Imagine you can reinvent everything for them with voice and AI. So I think we are very excited. It's just the beginning. I am very confident that AI will come, if not in one year, two years, three years. We are patient anyways. We're building it for the long term. So it'll happen at some point. And when that happens, then we can offer it to every customer.
Are you imagining any entirely net new products or experiences today which may not even sit directly within the core Meesho app?
I think it will organically happen over time. As I was saying, I started with this. A lot of team members in the company keep bringing ideas to me: 'Hey, can we do this? Can we do that?' I don't think it's part of some intelligent design, but I feel it'll happen automatically. It'll happen because teams who are experimenting on a daily basis, who are spending time with the customer on the ground, they continue to hear these...
Problems that they did not know how to solve, now they have a new toolkit. So I think it will automatically happen. We would have things that we would have never imagined two years, few years from now.
So, great segue into the next part, which is, and you started out the conversation with this right, which is unlike a two-person startup out of a house in Indagar, for who building AI and native AI today is a lot easier. You know, Meesho is now a listed company, for arguably Meesho itself as an enterprise, one of the largest buyers of things like voice AI today, but there's a lot of people also.
Right. And culturally having them change, adapt, move, bottoms up, come up with a lot of these things. How have you done it so far? And how are you thinking about it? Because it really feels like in all my conversations with a lot of founders or even CEOs of large companies, it feels like that is the majority of the problem statement. Once you get that to happen, everything happens from there.
You're absolutely correct. I think it's a hard problem. I would say we're somewhere on that journey now. To start with, we've been in a better place for a few reasons. One is our workforce is still young. So a lot of these people come to me and say, 'Hey, I want to use this tool and I will learn, I want to change how we've been doing this.' So the agility to learn on new stuff is much higher. So I am generally very hopeful. We've been trying to be an enabler right now. So let's say what role should we play as a company? We have to play a role of an enabler where a lot of people want to move to whatever is happening on the frontier. You enable them with giving the right tools, the right technology. Now, how do you change the entire platform? Because if you start today, I don't think the company would have the same composition that we have today. So I think we have to still make this transition, but because we have a younger company and second, our headcount is 2,000 people, it's not very large. So if you see, for a company of our size, many companies have close to 10,000 people. Our entire headcount is about 2,000, and majority of that is in technology. So tech R&D, engineering, product, that's where a lot of these changes are happening. So that team only comes to me and says, 'Hey, I think we need to change.' So I told you that in January we basically did our five-year planning. The first thing that the technology team comes to me and says, 'Hey, we want to go and become a completely autonomous SDLC, entire software development life cycle, in the next one year. In one year, no engineer will write a code to build a feature. Some journalist would be doing it using a coding agent. We'll just go out and build infrastructure that will enable them to move forward without any friction.' So they came up with that, and then basically started with what is all needed in making this happen. We now track this every month basis, what milestones we have put, are we getting there, not getting there. But we have not said that we will get there in five years, we said that we will get there to autonomous SDLC. I don't think it's happening anywhere yet, but the goal is to get there in one year. And it was all bottoms up. So far, it's okay, we haven't reached a point where we have to push something top down yet, because we've been seeing a lot of people themselves say, 'Hey, things are changing, and if we have to stay ahead, we have to keep moving fast.' So I think we are making that transition. The question is, can we do it faster?
Do you think about the org structure differently?
We are. I think the big change that we are doing, I think that's the biggest change. So today, the way we think about a pod. We have our basic unit, the atomic unit in Meesho is a pod, which is an independent team that has an independent P&L, independent budget, independent goal, and they are measured on the achievement of only that team. So it's an entirely self-sufficient independent team. An average pod size in the company is about 8 to 10 people, which is composed of some generalists, maybe a strategy person or an operations person or a product person, and some specialists. That's just out of curiosity.
They're organized by like what?
Shopping category function. So generally they are organized by problems. So we say a customer problem. So they take a customer problem. The customer problem is, for example, like we were discussing one. So there's a customer problem: my rural customers' activation is only 10%, I want to take it to 20%. So they will take this. So that pod now owns this problem. Whatever support they need, they can ask for it, but independently they go out and achieve it. So this is like an independent startup in the company. So 8 to 10 people, few generalists and few specialists, few engineers, maybe a data science person, maybe an analyst, and so on. So what we believe will happen in a few years is none of the specialists would be part of the pod. These specialists would rather work centrally and build great infrastructure, build great platforms, and would provide agents to generalists who are part of the pod. And these generalists would be leveraging these agents to basically move fast and execute whatever. So I need to code, I can do it. I need to train a model, I can do it. I need to ship a model, I can do it. I need to do very deep analysis, I can do it. I need to design a new experience, I can do it. But these generalists can do it using great agents that we have built over time.
