A US company conducted a study which found that last year, AI created one million new jobs. In return, those who were fixing bugs, if there were any issues, AI agents would handle them. My engineers can focus more on creating new features, new products, new results. The biggest way to become an AI-driven company is to focus on the core objective. What are today's customers coming to Razorpay for? For me, complete more transactions. For me, reduce fraud or errors. For me, reduce mistakes or errors. That's the only thing they care about. If I can reduce that with AI, that's the biggest attack I can make. It feels a lot like Razorpay is transforming into an AI company.
In your opinion, what is the most ambitious AI product or project that Razorpay could get its hands on before we meet here next year?
Alright, hello everyone. I'm Akriti, back again with another episode of Signal Dialogues, where we talk about the AI landscape in India. I'm very excited about all the people I've had the chance to sit and talk with. And today, you know, I've been wanting to do this for a long time, and what could be a better occasion than the Global Fintech Fest. You might not notice that we are now at GFF, because it's impossible to get a spot there, so we've managed a spot close to the venue. And as I said, I'm very excited to talk to my friend Harshil about Fintech and AI at GFF.
Hey, great to be joining you. It's great to be talking about this.
Yes, Harshil, as you all know, he is the co-founder and CEO of Razorpay and has been doing some great work for the past few years, of course in Fintech payments, but recently in AI.
Okay, we were talking about this earlier. I think Harshil, you know, everyone knows a lot about Razorpay, right? Your payment business, your journey, how much you've achieved. I think not many people know about Razorpay's journey to becoming an AI company and how you are starting this journey from the very beginning. So I thought today we would focus on that. I'm saying I won't ask you about the IPO or things like that, but we will focus on what Razorpay is like as an AI company. So let's start with that. Tell us a bit from the beginning, maybe from the time GPT launched in 2022, when we were all in this ecosystem. How did you react?
I think it's one of the biggest transformative changes of our generation, and we are all seeing it happen. So it sometimes feels truly unbelievable. Of course, after GPT came. I was looking at Sam's direction and used it on the very first day, and I was mesmerized by it. And I knew Sam because we were at YC, and during our time, Sam was the head of YC. So, seeing it come and knowing Sam, I was sure it was going to be something huge, but I don't think anyone thought this change would be this big or transformative. So, personally, I'm trying to understand the implications of this change, and with every new version of GPT that comes...
But you adopted it very quickly. The thing I've noticed about you is that from that moment, you and many in your ecosystem adopted it quickly. Right? Technology adoption, right? And any new thing in technology excites me a lot. So when I saw GPT, I started reading the model. I started reading the research papers about it. Of course, the Transformer paper and other things, at that time a lot of it was theoretical. But there was a theoretical idea of how it could change the ability to predict and solve problems. I mean, such a capability in computer software, because those who work with technology know that creating software that can solve problems it wasn't trained or designed for is a huge step in computing. So, right? Until before GPT, we had never created software that could solve problems its creators didn't create a solution for. I think this is a huge transformative change in computing. And I think that's what mesmerized me, that we can now create software that people can use in various ways and solve problems we never planned for. And now that's exactly what's happening, large language models have been created and people are starting to apply them to assistants. People are starting to apply them to commerce, coding, and hundreds of use cases. These use cases were never there, it was never designed for that. I mean, I don't know if you saw yesterday's OpenAI news.
Every morning I wake up with a headache.
Yes, OpenAI can now solve mathematical problems.
And you've seen I launched it.
Yes. These mathematical problems, it wasn't trained to solve them, but it's now quite good at logical thinking. It can solve complex mathematical problems, right? It's truly incredible.
So how was the journey from the beginning? How was 2023? What things did you start with? I'm asking this because we meet many people and many of our audience are on this journey.
Yes, from the very beginning to large enterprises, right? So what were some of the early experiments?
Yes, the early experiments were very simple, like how we used language to create chatbots. How to create assistants. How to make customer support better. We started using it internally for our customer support team. For example, we have an internal agent that helps my customer support team. When a ticket comes, they get a pre-made response that says, you should answer this way, and then they can edit or change it. Of course, there were issues in the early versions and some things worked, some didn't. But I think the biggest change came when Opus arrived. Yes, I think that model changed the perspective that AI is no longer just an assistant, it's an agent, right? And I think you use Cursor and use GitHub Copilot for code completion. But in my opinion, Opus is a model that has moved from code completion to becoming a code writer. Meaning you tell it and it will write the code for you. You just have to review it. And when I personally saw Claude Code being used, I was truly amazed. Finally, we got a model that works as if it has its own personality. It's someone you can talk to and instruct to do a task. And then the whole ecosystem became agentic. Yes.
It's quite funny to say that. I remember talking to Tanay, the founder of Whisperflow, about this, where November 2022 was a significant moment in the history of AI. I often think maybe December 2025 was the biggest moment.
With Opus and some of the work done with it.
This agentic transformation has truly changed everything.
Finally, a model has become agentic, where instead of just using the model as an assistant, the model is now in the driver's seat and you just have to give input. You can start using it almost like a person. Like, hey, just do this for me, and do that for me. And I think, this is the change I define as the ascent to this agentic world. Which happened with their model launch.
Yes. And then, let's first discuss the thing I find most interesting about you, which is you've worked a lot on your personal AI usage, right? You've optimized a lot for your own productivity and work. Tell us a bit about that.
Yes, look, I think everything in the AI world is changing very fast. If you don't use it yourself, you will either overestimate or underestimate it. Right? Like, I also tell my leadership team that if you look at AI from the outside and just keep reading everything, you will fall into one of two situations. Suddenly you'll feel that everything can be done with AI, just like many leaders are trying to do now. Hey, just use AI for this problem, just use it, and AI works like this. So, I encourage my team and I am personally vocal about this. If you don't try it hands-on yourself, it's truly a big change. It's a lot like the 90s, when the world was going through the computer transformation and you're a leader who says, yes, computing is bad, but without seeing a computer, you don't know what this technology actually is. The same goes for agents, you have no idea, because there are many great opportunities, but there are also limitations. So, what have you built for yourself?
I'm using, like when OpenClaw came, I was busy with OpenClaw for two whole weeks. The team was complaining, where are you? Why aren't you focusing on meetings? And I couldn't answer, because it truly captured my mind. And not just because OpenClaw will become something huge, but it was for all of us. The same thing, it was a software or whatever you call it, with the ability to do something it was never asked to do before, right? What we read in theory became reality, where I told it to book a ticket on IRCTC and it completed the transaction. Where I told it to buy this thing for me and it completed the whole task. It had never seen the Swiggy website before. And it knows how to do it. It amazed me. So I've created, I mean, I still have three agents that manage my entire personal life. One is a fitness agent, which tracks everything from my sleep to my eating habits and my Whoop, and stays connected with it. And provides me with insights. Very unique insights. It has told me that when I eat Indian food in the evening, my recovery score is bad. When I eat Italian food in the evening, my recovery score is good. Like what time of day is best for me to eat, etc. It analyzes deeply. Then I have a business work agent that tracks all my meetings and creates pre-meeting notes. And sends them to me. And then tells me that I'm meeting this person and what was discussed last time we met, etc.—that's quite useful. And then I have a general personal agent. Which I instruct for any task. I don't look at travel websites for flights. I just tell it. I go to play pickleball every Monday. It suggests to me, should I book a pickleball court for this evening? It's in my friends' group. Great. People say yes, it books for them, and for however many said yes, it books the court. So, the main point is not that I can't do this.
No, 100%, you're absolutely right. Apart from leaders and founders, whether it's a 2-person startup or a 50,000-person bank—it doesn't matter, this is how the transformation will happen. You have to use it yourself.
It has nothing to do with whether it's very valuable or not. Maybe I could have a PA or someone else do it. The value is when it breaks, you understand the reasons for its failure. What mistakes it makes or when it makes a mistake, why does it do it? Because it's completely different from any other software you use. Most software fails predictably, meaning we know where it will fail and where it will work. Historically, if a software does a job well, you can do it 100 times and it will give the same result. Like Excel, if you add 2 + 3 and it gives 5, it will give 5 even if you do it 200 times. But this is not like that, it works on probability. And this is a very unique way for us to think about software. It's not normal for us. That it does something correctly 10 times. But on the 11th time it might fail. And it fails because its nature is probabilistic. And those complex or edge cases are very hard to understand. Until you use it yourself. So when I'm in meetings with my engineers and explain that let's do this or that, we discuss the framework for building it. If I don't know the reasons for failure, it's truly hard to design these systems correctly. Right? Personally, gaining this experience was very valuable for me. The entire leadership team. In our company, there are 200-300 people who are running their own Hermes agents within Razorpay. It's great, because they are doing the day-to-day work. The idea is, if we want to be a native company, we want to create agents for commerce, create agents for payments. My leadership team using agents themselves is the only way this is possible. They see what's failing, what's working. We recently started using Grok bot. All of them. For every change, you have to understand how to do it yourself. And yes. It might not be much. It might not be perfect today, because many of today's agents are very early stage. And I don't think we've seen its final stage yet. But the main point is, you're learning so you can contribute in the final stage, instead of just sitting outside and watching. 100%.
So let's talk about it properly, right? You started working on the technology yourself to learn it. You understood that the initial use cases were customer support, then Opus happened. And then last year we've obviously talked about, and everyone has observed it. It feels almost like Razorpay is transforming into an AI company. So help us understand last year, right? So after Opus, what did you do, where do you want to go with this? It's the first step to start coding, right? Like any coding, coding is about solving a problem that needs to be solved, and I think people think just give everyone a Claude Code license, and your job is done. And that's just the starting point. The challenge of engineering and coding is, usually large organizations like ours are built in a way where engineering is the slowest step, meaning engineering writing code. That process, right? Which is always a limitation. The slowest step. So, the rest of the ecosystem is actually not optimized for maximum speed. So the way engineering works, the way engineering organizations like ours are built, there you spend a lot of time ensuring that you don't ruin your most important thing, which is your engineering bandwidth. Your engineering bandwidth is very sacred. So you spend a lot of time creating product specs and tech specs. To figure out what needs to be built. To be completely sure that yes, this is what needs to be built. Because the value of those hours is the highest. Yes, because in an engineering company, engineering bandwidth is the most sacred thing. You don't want to ruin it for anything. Now suddenly that engineering bandwidth has become much, much faster. The person who used to take three weeks to build something, now takes one day. The person who used to take three months to build something, now takes one week. So suddenly it has become much faster and significantly faster. Because you have agentic coding. And now that takes time to learn. But once people learn to use agentic coding. That problem is solved. Suddenly you'll discover two big problems are being created. The first is you'll discover that the whole framework you built to optimize engineering productivity is actually backfiring because you've become much less experimental. You can now be much more experimental because I talk to many VCs, especially CEOs from the Valley, when I learn, from many of them, what they do is they create prototypes, like you yourself have given to the prototype team.
Yes, and you have to be very careful that it's a prototype, it doesn't mean it can be built in a day. Because you can't understand, especially in your business where you have to be very conscious about privacy. Yes, because often people look from the outside and say, hey, if I can do it in a day, why are you taking a month, right? So that part is still there. But then you started the product. The production part is still hard. The second big challenge is when your engineering speed suddenly increases. Your engineering development flow or other parts of the SDLC are not optimized for it. So, your DevOps, your security, your infrastructure, your testing, many such things. You never optimized for them. Because engineering was always so slow that you thought these tasks might take a week. You take a month. That's fine. For example, if you go to any pre-AI company, you'll see many issues filed in their repository. Because people keep filing issues and you keep filing issues and it's a lot. Any company that is becoming AI-native. You'll see there are no more issues in their git repository. There are many PRs or pull requests. Because many people have created something and submitted it as a PR.
You know, I was at an event where Satya Nadella was speaking. Sangeet.
Sangeet. Yes, we were there together.
And do you remember he said that in the last two years, compared to the last 20 years of history, they've seen more PRs submitted on GitHub.
Yes, so it's very easy. Because now you can tell an AI agent to pick up all these issues and write a PR. It has written a PR. But then you have to test it. You have to figure out canary deployment and do many more things to ensure this PR can be sent. So what you've changed is, you've moved this hard step of development from engineering to testing, deployment, and other things. Your overall outcome has truly increased. Because your net outcome might be 20%, right? So then suddenly you realize that's not enough. Even though your engineering will be 5x more. Your outcome won't be 5x. You're not deploying more code, not bringing more features. And this is a complex engineering problem that every company in the world is currently trying to figure out.
ROI outcome is everything.
Yes. How will you get to the ROI?
And again, the wrong or non-normal immediate reaction is for someone at a cost-conscious company to come and say, okay, now reduce this. Yes. Let's move on because I can't see any outcome. Outcome.
Yes, because you can't see any outcome.
You're not getting it because you're part of a relay race. One person on your team was running very slowly. So everyone adjusted to their pace. Now this person is running very fast, but everyone is still moving at that same pace. So you're still seeing the same result because someone is still the slowest. Suddenly they became fast, but no one else could keep up.
Yes. So, and I think that's the core problem we're emphasizing. How will you change that? Right? The second aspect is what we've started. Suddenly, you might have very strong engineers who can do a lot in every area and change the workflow. So, only in the field of engineering. Since we're moving towards workflows, and I was talking about coding, how is the organizational structure changing?
Yes, engineering. Yes, so, for engineering, we had to create a separate team whose job is now to focus on engineering productivity across the entire board. That means, how to solve the SDLC, software development life cycle? So, the core engineering is fine. But how to make testing easier for my engineers? How to make deployment easier for engineers? How to make PR reviews easier? How to create a system where my PR merge time decreases, etc. So, I think those parts, because we never had to optimize for them. Suddenly, we had to create a team, who are focused, 'Okay, look around coding and see, where are we wasting bandwidth or where are we taking too much time? How can we make this faster and how can we deploy agents to solve these?' As I said, this is not a fully solved problem yet, but is there a huge productivity gain there? No, it's, it's, it's, it's, it's, it's, it's not solved anywhere.
Do you think, especially in a business like yours, how are you institutionalizing domain knowledge or subject matter expertise among engineers?
Yes, and that's why it's a coordination, right? You have to work with those people who are experts. For example, when we're building customer support for engineering. It's not like, these engineers, you build customer support. Or other people, you build customer support and launch it. My operations team, my power team, and my engineering team are all working together to say, okay, year after year, this is what we've learned. This is what works, how you will train an agent. Because this team understands the agent. But it's a team that understands how customers behave. And how you put that knowledge to use in problem-solving. I think at the end of the day, the result will be that we'll have a human-agent workflow where agents do repetitive tasks and people do intellectual tasks. So, whatever new happens in the workflow will still have to be done by humans. But when it becomes repetitive, it will be done by an agent. I think that coordination is the real synergy, and it's still a hard problem to solve.
And I want to add one thing, many people talk about, does this mean that if engineering productivity or productivity increases, the number of engineers will decrease? I'll say the opposite, it's a huge paradox or puzzle. Yes. It's a huge increase in the engineering paradox, in other words, right? Today we have 1500 engineers in our company. Right? About 50-60... did you read this yesterday? Maybe you saw it, but I sent it to you. It's one of the most amazing reports that has come out. I think The Economist or some American organization did a study where it was seen that last year AI created a million new jobs.
It will be much more than that.
Yes. It will increase much more. Like, from my perspective, right? We have about 1500 engineers. Right? About 50-60 percent, on average at least 30-40 percent of engineers are actually not building new products or features, they are maintaining existing work. They are handling bugs, they are looking at issues, now an agent is taking that over. Right? My engineers' average productivity increases because most of my engineers can build new products. Building new features, creating new growth paths. So, the average result I get, if I invest X amount of dollars in engineering, I get Y amount of results in terms of revenue. So, it's increasing massively. It's a lot like a cake, right? If you get more results, you invest more. You hire more engineers. I will hire more engineers because suddenly engineers are bringing much more results for me. Instead of those handling bugs or issues, if agents handle that, then my engineers can focus more on new features, new products, and new results, and that's a huge thing for me. So, it will lead to many more engineering jobs than before.
From a customer perspective, tell me how AI will look at Razorpay.
Yes. So, I think from a customer perspective, we said that, okay, I think in January of this year, we said let's sit down and decide, we can apply AI here. Hey, it's happening, everyone knows, it's like the fourth conversation which was January was the deadline. Once that was over and this November came, everyone said, boss, this has to change. This has to change. So, yes, we can apply it here and say, hey, I will do customer support better, hey, I will do this better, I can give everyone code. But what are we actually doing, right? It's such a huge change that I don't want to be like many old-school companies who during digitalization said let's hire a digital team whose job is to do digitalization, Chief AI Officer.
