Welcome to the inaugural edition of AI Chronicles by GIA (Global Indian Alpha) by Leo Capital. In this series, we explore how C-suite executives across functions—CTOs, CIOs, CMOs, CISOs, and more—are leveraging AI to transform their organizations and, by extension, their industries.
For our first interview, we sit down with Arun Rajkumar, Co-founder and CTO of Atoa, to explore how AI is revolutionizing software development, hiring practices, and organizational culture in Atoa. For the uninitiated, Atoa is helping 1000s of businesses across the UK receive fairer payments in just a few taps.
Q: How did your AI journey begin? Was it a planned strategy or experimental?
Arun: ML isn’t new to me. When I started my career in 2006, I was doing ML, neural networks and other stuff. My first startup was actually something similar to what GitHub Copilot is now, but I was too early in the market.
With that being said, coming to Atoa. When ChatGPT was launched, there was a lot of buzz, but the models weren’t up to the mark initially. Many thought leaders like us believed AI was still too far, it wasn’t able to get enough context for coding. But to our surprise, within 6 months, a lot had changed. Then came the Anthropic models, and that was a game changer, really.
Also, just for some context, we’re growing pretty fast. The product is UK-based, but everybody, product, tech, even some sales folks and content marketing, etc., are working out of India.
Initially, there were lots of questions and concerns. What happens to our source code? Is it used for training? Since we’re FCA licensed in fintech, security is paramount. People were also worried about job security, what happens if AI starts coding? Will my job be secure?
There were a lot of assumptions and scariness. But I believed AI was going to be a real game changer and help us accelerate development. For example, we had a big project coming in and I was supposed to hire 5-6 people.
Instead of investing in hiring, I thought, let me give AI a try. That actually helped us stop the hiring and increase productivity by almost 70%.
Q: That’s a significant productivity gain. How has this changed your approach to hiring?
Arun: Yeah, we will not hire unless it’s very essential. The typical process now is that the only main thing is that you have to just define what you want to do and get the prompts ready. The rest, everything, is what AI is able to do.
Now with agents, it has become super powerful, it even does unit testing, runs, and executes everything. The only place where we invest time is in code review.
We have to step in with business knowledge to ensure AI isn’t hallucinating or doing incorrect things. That’s the only place where we spend more time. So we’ve optimized hiring. The engineering team is now flat, it’s not like when you raise funding, you immediately hire 20 people.
Q: How has your interview process evolved with AI?
Arun: There’s only one thing that’s changed significantly. Now we’re looking for more experienced people who know how to architect systems and design solutions. Previously, we used to check if they knew data structures, basics of coding, etc.
Now we focus on: Can they solve problems? Can they think outside the box? Can they design systems and know what’s right and what’s not? Because AI is going to do the groundwork.
The process is the same, but we don’t check if they can type fast or code quickly, or if they know data structures extensively. We only care about whether they have enough problem-solving experience and architectural experience. At junior levels, we think that engineers who can give very good prompts can be effective.
Q: How do you decide which AI tools to adopt across your development lifecycle?
Arun: When I wanted to drive adoption, I first told my team to try ChatGPT. After a week, they said it wasn’t useful, too vague, giving random responses. But I strongly believed it had the capability, so I took up coding tasks myself for a week or two. I tried almost 5-6 tools to see which model worked best and where it was lacking.
I noticed that if you just randomly ask without specifying exactly what you want, it fails. You have to specifically optimize your prompts, that’s why they call it prompt engineering now. If you give a prompt exactly specifying what you want, it works like magic.
We use Cursor primarily. We tried multiple tools, including Windsurf, and we’re adopting some of those as well. I completed a big module without writing a single line of code even with testing and deployment, and it produced excellent code. Then I shared all the learnings and trained the team.
Q: Can you walk us through your complete AI-powered development cycle?
Arun: Starting with product ideation, our product managers use a mix of ChatGPT and v0.dev. Nowadays, product managers even create wireframes to give to designers, showing what they want to achieve. They build complete prototypes rather than typing big specifications and having meetings to explain them.
Even competitor analysis is done by ChatGPT. We’ve given it all context as in the required company knowledge, and we have an agent called “auto product analyst” that does product research. It has all the context of what we do and don’t do, so PMs use that effectively for feature development, pricing, market research, etc.
For design, it’s the one area still not fully automated. We’re experimenting with Figma tools, but design-wise, we want to stay very focused. Design is still the area we’re experimenting with AI, but haven’t achieved full automation.
The real fun happens with the engineering team.
We use Cursor or Windsurf. We have three different approaches. Sometimes we run local models that are fine-tuned for specific tasks like log processing. Cursor is most used, and we subscribe to it. The specs are directly converted with AI doing most of the heavy lifting.
Unit testing also happens in Cursor, and then the module goes to QA. QA automation also happens through Cursor. Then deployment, the DevOps team also uses Cursor to deploy and automate certain things.
This is the full cycle. Plus, there are other features, even in sales, we use AI. Our cold calling to some verticals happens through AI.
Q: What about security concerns, especially in fintech?
Arun: Code, generally in our architecture, is not coupled with any sensitive information. We haven’t exposed database access, so that’s one critical part. For business analytics on data, we may use our own self-hosted AI, not public models.
Coming to code security, I think every company founder has this fear: what happens to my code? Does someone steal it? But ideas, even if unique, depend on execution.
Code has become just lines of text because AI has been trained with all the GitHub code in the world. It already has all possible codes.
Your code will be a subset of existing code. Even if you upload your full source code, it won’t be useful to anyone. AI isn’t going to learn anything new from you. It’s already giving you others’ code and making your code faster and better, increasing productivity and identifying performance improvements.
Code is not proprietary anymore, it has become too commoditized. The only thing proprietary is your execution style or customer data, which is protected. If we process customer data, we do it with our own in-house deployed models.
Q: Do you have specific ROI targets or metrics you’re tracking?
Arun: Yeah, I have this ambitious target: when specs come in, within T+4 days, be it any big or small feature, I want it fully automated until deployment without any human intervention. With some automated checks when the product manager releases specs, we want to eliminate sprints entirely.
There’s no concept of sprints in my vision. The feature comes, then agents pick it up with minimal human help, and it gets shipped immediately. You can see the velocity this can give, one human handling 3-4 agents and ensuring features get shipped fast. That’s my ideal goal where software development is just an agent doing the job, and your only focus is selling.
Building should be an autonomous task.
Q: How do you manage change management and keep engineers motivated in this AI-driven environment?
Arun: This challenge comes often. People who are futuristic and excited about AI fit in well because they’re exploring new things and building innovative solutions. The people who are ‘hardcore’ developers always have this fear. When I tell them to ask AI to debug instead of spending time manually debugging, they think, “Why should I ask AI when I’m here? What’s my importance?”
We ensure that credit is not given to AI; instead, credit is given to the person who owns the work.
All our product features don’t just come with a list of “do X things”, they always come with a business context. When you do this, how will it help the business? We highlight this connection.
When someone finishes a feature, we share customer feedback for that module and publish the new revenue generated from that feature. Unlike other companies where founders keep revenue information very strict, in our company it’s pretty open- which customers are ongoing, what volume they’ll do, and product growth metrics.
That’s why our attrition is very low. In the last 3 years, I’ve had only 2 resignations.
Q: How do you stay updated with the rapidly evolving AI tool landscape?
Arun: The primary training focus is on prompting, specifically determining which prompts to use and what context to provide.
Training is very dynamic. Every week, somebody is launching a new model, new context changes, new tools. We train every week.
We teach specific techniques: Don’t ask AI to understand everything, be very clear. Tell AI not to touch other parts of the code or optimize other functions. Only change this specific part, don’t overdo it. Some engineers give everything to AI, and it changes 20-30 files, then they come back confused.
We emphasize that code review is very important. You cannot just keep AI in agent mode working crazy, changing everything, and breaking things. We train on what not to do and what to rely on AI for. Use it as your buddy, not as something you give tasks to while sleeping.
Q: Do you prefer platform tools like Cursor or specialized vertical tools?
Arun: For sales and marketing, we use vertical-specific AI tools because they’re fine-tuned for those contexts with purpose-built agents.
For engineering, tools like Cursor might be enough. These platforms can even write, say, end-to-end Playwright test cases beautifully, so verticalization is less necessary for technical teams.
Q: Where do you see the biggest AI impact in your engineering process?
Arun: Development and coding is where there’s the biggest impact is. QA automation works to an extent, but there are corner cases where, as a QA person, you have functional knowledge that you need to apply specifically.
QA is evolving, and I think reasoning models will improve this. Even SDR (sales development) and sales calls still have quality gaps.
Q: Any advice for entrepreneurs starting their AI journey?
Arun: Entrepreneurs should first start believing in AI and focus on business. For coding, find someone who believes in AI coding and can promise that they’ll use prompts effectively with AI. If you have someone like that, blindly hire them. Don’t hire people who say “I’m a hard worker” or focus on traditional coding skills while dismissing AI as unreliable or prone to hallucinations.
I’d encourage founders to try building their prototype or MVP themselves using AI, then give it to engineers. You’ll understand the power of it.
AI in engineering should be highly prioritized. If you’re not doing it, you’ll be left behind and definitely lose to competition.
Q: What’s next on your AI roadmap?
Arun: Outside engineering, there are lots of spaces we still need to automate. For example, our cold campaigning, digital marketing, sales, and analytics. These functions are evolving slower and need to be tailor-made because every company has a different approach. These are the areas we’re now heavily investing in to automate and scale faster.