Welcome back to On Call with Insignia Ventures! In today’s episode, we continue our insightful conversation with Zi Yong Chua, Carro COO, Southeast Asia’s largest used car marketplace.

In Part 1, we explore Zi Yong’s fascinating journey from founding his own mobile payment company to roles at PayPal and Alipay, before joining Carro. We learn how he tackled his first order of business as COO by refining Carro’s internal data processes to drive productivity gains, and how he linked these improvements to tangible business results. Most importantly, Zi Yong shares his three core principles for leveraging Generative AI effectively.

In Part 2, we dive deeper into how these principles have shaped Carro’s Gen AI journey over the past three years. Zi Yong reveals how Carro has translated theory into practice, implementing targeted AI solutions that deliver real business value rather than pursuing AI for its own sake.

We’ll also explore what’s next for Carro in 2025 as they develop agentic workflows and experiment with smaller, domain-specific AI models. Finally, Zi Yong breaks down how he communicates ROI for Gen AI transformation to both management and frontline employees—a crucial skill for any leader driving technological change.

Whether you’re just beginning your Gen AI journey or looking to refine your approach, this conversation offers practical insights from someone who’s been experimenting, learning, and implementing for years. Let’s join Paulo and Zi Yong for the second part of this must-listen call before you undergo your own Gen AI transformation.

Part 1: Principles for Generative AI Transformation

Timestamps

(01:19) How Zi Yong joined Carro after PayPal, Alipay, and his own ventures;

(04:41) Zi Yong’s first order of business as COO, refining Carro’s internal data processes for better productivity;

(08:20) How Zi Yong linked productivity gains to end business results; 

(12:59) Zi Yong’s three principles on leveraging Gen AI;

Part 2: ROI for Generative AI Transformation

Timestamps

(00:00) Translating Zi Yong’s principles on Gen AI Transformation (in Part 1) to Carro’s journey the last three years;

(06:55) What’s next for Carro when it comes to Gen AI;

(10:46) Communicating ROI for Gen AI Transformation to both management and frontline employees;

About who you are on call with

Zi Yong is the Chief Operating Officer at Carro. He drives the “People, Process, Technology” transformation within Carro, ensuring that Asia Pacific’s fastest growing automotive marketplace remains at operational excellence as it continues to grow rapidly. He is focused on digitising workflows, enabling automation and data-driven insights within the organisation.

Zi Yong brings more than 15 years of experience from the Internet and e-payments industry, with roles in global giants like PayPal and Ant Group. In his 5 years in Ant Group, he led Product teams and projects with Paytm and Touch N Go Digital, helping them become ‘super app’ in their respective markets. He was also the principal inventor of a patent for payment system decision making during his time in Ant. Back in 2009 while still in university, he founded his own venture-backed mobile payment company, and was one of the early drivers of the Android developer scene in Singapore.

Connect with Zi Yong

Follow us on LinkedIn for more updates

Subscribe to our monthly newsletter for all the news and resources

Directed by Paulo Joquiño

Produced by Paulo Joquiño

The content of this podcast is for informational purposes only, should not be taken as legal, tax, or business advice or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any ⁠⁠⁠⁠⁠⁠Insignia Ventures⁠⁠⁠⁠⁠⁠ fund. Any and all opinions shared in this episode are solely personal thoughts and reflections of the guest and the host.

Transcript

Paulo: Thanks again for taking the time to come on. As you’ve seen on our show, we’ve had a couple of your colleagues on already. Obviously, Aaron, and then Ernest as well—twice, I think. And then Bryan—we even did a whole documentary. I flew over to Carro Indonesia, and filmed him going around. I’m happy to have another member of the management team here to get another perspective on the company. I think it was just a week or two ago when I reached out, and we were able to set this up—really because of that LinkedIn post I saw. I had just come from an AI event with some of our other portfolio companies, so the topic was fresh in my mind.

Then I saw your post, the Google Cloud post, and how you reposted it. That’s when I thought, “Oh, here’s someone I can talk to who has actually gone through this whole journey—not just someone who started using Gen AI last month, but someone who has been experimenting, failing, and learning over several years.” I think it would be really interesting to hear more about that on the show.

So I thought we’d start off by giving some context on how you even ended up pushing this initiative within Carro. Maybe first, let’s talk about how you joined Carro—how you found out about the company, given that you’ve been in several tech companies before. What was that journey like for you?

How Zi Yong joined Carro after PayPal, Alipay, and his own ventures

Zi Yong: Thanks for having me. It’s a pleasure to be here with you. As you mentioned, many of my colleagues have been on this podcast before.

I don’t tend to post much on LinkedIn, but quite a number of things have happened in the last year in terms of solutions being rolled out into production. Google Cloud wanted to work with us to showcase some of these solutions, which is why I posted that article two weeks ago. I think that’s how you picked it up.

Happy to share from an operator’s perspective how we built our Gen AI journey from the ground up. We’re not trying to build AI for AI’s sake—we’re really trying to get business value out of it. It stems from my role because I’m in charge of operations and processes. It’s really about how we streamline our people and processes, using Gen AI to facilitate that.

Thanks for the quick introduction. Before joining Carro, I actually spent more than ten years in internet companies. In fact, I started my own company about 15 years ago.

Back when Android had just been released. In fact, I developed the first Android geolocation app for Singapore back in December 2008. After that, I ran a mobile payment company, then moved on to PayPal and later Alipay, where I was leading various regional projects and digital initiatives.

At Alipay, I was supporting Paytm and Touch ‘n Go, helping develop their super apps. You probably know that Paytm is the largest wallet in India, and TNG is the largest wallet in Malaysia. I was part of the product team driving digital transformation for those companies in those markets.

I joined Carro because I knew Aaron back when I was running my own company. At the time, he was with Singtel Innov8, which was also one of my investors. I was always impressed by Aaron as a person, and we interacted frequently because of our investor relationship.

When I left, I told him that whatever he started in the future, I’d be more than happy to invest in it. So I actually personally invested in Carro’s Series A. And by the way, I didn’t even know what the company did! It didn’t matter—I was betting on Aaron.

Fast forward to 2020, at the height of COVID. Aaron and I caught up again, and he told me he needed someone to help him at Carro. By that time, I had spent seven or eight years in the corporate world and wanted to go back to startups. So I signed up and joined Carro.

Paulo: That’s an interesting journey. Typically, people join big tech before they start their own company, but you did it in reverse. I love the story about him investing in your company, and then you eventually investing in Carro early on.

You mentioned that you’re in the COO role, and I really like how the PPT framework sums up what you do in a concrete way. Maybe you can take us on this journey—you’ve been with Carro since 2021, right? That’s already four years now.

Quite some time in startup years. So there’s definitely a lot that you’ve seen and shaped within the company. Maybe you can walk us through that journey. Before this call, you mentioned splitting it into two parts. The first part was about “cleaning house”—maybe you can speak more about that?

Zi Yong’s first order of business as COO, refining Carro’s internal data processes for better productivity

Zi Yong: When I first joined, my focus was on cleaning house and driving digital transformation. Carro had a lot of systems and tools—our own in-house ERP, our own CRM. Brian, whom you’ve met, built the AI pricing and inspection app that we use.

But truth be told, there were still many gaps in the workflow. Ground staff were using Google Sheets and Excel to plug the various workflow gaps. As a result, data going into our systems—whether ERP or CRM—was often inconsistent.

I remember having to do a lot of reconciliation for analysis and reporting. It was a nightmare—one single analysis would take me four weeks because I had to reconcile data across four or five different sheets, cleaning definitions across all business lines. Revenue in Malaysia was defined differently from revenue in Singapore, or in insurance, or in financing.

Back then, every country spent at least one day a week just generating PowerPoint reports. It was incredibly unproductive, and the numbers weren’t always reliable. My first priority was to clean house—move all workflows into ERP and CRM, standardize reporting across the group, and implement a governed data definition framework.

That meant clearly defining revenue with and without commissions, COGS, OPEX—everything had to be properly structured. This took about two years, going business unit by business unit, country by country.

Today, I can go into my dashboard and instantly get detailed slices of inventory, cohort profitability, and sales performance down to the individual salesperson in a specific branch in Malaysia.

So that set the stage for us to leverage this data and integrate it into Gen AI, allowing us to automate parts of our workflows as well as our alerts. Since we had already cleaned our data structure and built semantic mappings across all our metrics and dashboards, we could now focus on AI-driven automation.

How Zi Yong linked productivity gains to end business results

Paulo: Data is really the foundation for any kind of AI adoption. After two years of cleaning house and addressing data complexity and granularity, what came next? What was the second phase?

Zi Yong: By the end of my first year at Carro, I was tasked with turning around one of our biggest business units. At the same time, I was still in the midst of digital transformation—aligning all the systems, making sure every policy definition was clear. And suddenly, I had to step in to fix this BU.

But what was great about this challenge was that it gave me a proof point—showing that digital transformation actually brings tangible business results. Many organizations struggle with digital transformation because they don’t see the immediate purpose of data cleaning or having a well-governed data structure. CIOs will emphasize its importance, but at the end of the day, business priorities often take precedence.

For me, this was an opportunity to run both initiatives in parallel—driving transformation while also executing a business turnaround. My approach was to go into every single department within that BU, redefine people’s roles and responsibilities, and optimize the processes.

As I mentioned earlier, PPT—People, Process, and Technology—is the foundation. Ideally, I always start with the first P, which is People. It’s about ensuring that we have the right people in place, that they understand their roles, and that they know the KPIs they need to deliver. Once that foundation is in place, we move on to Process—ensuring that departments work seamlessly together, SOPs are clear, and everyone understands their handoff points.

It took one or two months to figure all of this out. Within a month, we were already able to track the performance of every salesperson, which ties back to what I mentioned earlier about productivity and tracking. By being able to measure performance at the individual, branch, and BU level, we could take immediate actions and make more informed decisions in real time.

For a company like Carro, this is absolutely critical. While we often talk about AI, retail experience, and delivering a seamless car ownership experience, the core of our business actually lies in supply chain management.

How efficient are we at procuring the right cars at the right price? How well do we manage costs across logistics, maintenance, and reconditioning? How fast can we sell these cars?

The more effectively we operate—whether in terms of time or cost—the greater our competitive advantage when going to market to procure cars. Because we can analyze what we buy, how fast a car turns over in our ecosystem, and what pricing strategies maximize profitability.

Paulo: So it’s about optimizing the car’s turnaround within the ecosystem.

Zi Yong: Exactly. As I always say, we are not in the business of selling one product a million times. We are in the business of selling a million products one time. Each car is unique. Okay, maybe not just one time—some cars might be resold two or three times in their lifetime—but you get the idea.

The real challenge is: How do we best manage all these cars?

This also ties back to Gen AI, because at any given time, Carro has tens of thousands of cars across our leasing books, marketplace inventory, and auction yards. We also manage cars that we procure for auction from banks.

How do we optimize inventory management in the most timely and cost-effective way to maximize value? This is a massive challenge, and it’s impossible for a human to manually analyze dashboards and make the best decisions for thousands of cars.

That’s where Gen AI comes in.

Zi Yong’s three principles on leveraging Gen AI

Paulo: The way you structured Part 1 and Part 2 really highlights how you linked efficiency in input metrics to efficiency in output metrics—from improving internal productivity to directly impacting the business’s bottom line.

That sets the stage perfectly for our discussion on Carro’s Gen AI journey. But before we dive into that, I think it’s important to understand your own views on Gen AI—because those perspectives shape how you lead your team, implement experiments, and drive adoption.

Zi Yong: To qualify myself—I’ve never been an engineer. In fact, my only background in programming was an introduction to programming course in C.

So, when I first built my Android app, it was in Java. I also wrote several applications for Google Glass. But here’s the thing—I built my apps using code snippets from Stack Overflow.

My superpower was knowing how to find the right functions, put them together, and make them work.

So when Gen AI came along, it was like a dream come true for me. Instead of me searching Stack Overflow for the right code snippets, Gen AI can now generate the code directly for me.

If Gen AI had existed 10 years ago, I would have been able to do so much more.

Now, looking at the broader picture, Gen AI is going to fundamentally change how work is done.

1. Smaller teams with cross-domain expertise.

Instead of large teams of specialized experts, we’ll see small, agile teams with broad knowledge across multiple domains.

The key will be integrating solutions rather than just specializing in one area.

2. Bottom-up innovation instead of top-down mandates.

In the past, AI projects were often driven by data scientists or engineers, trying to build solutions and then hoping the business would use them.

With Gen AI, the model flips—the best solutions will come from business owners who have full accountability for results.

Instead of AI teams dictating strategy, they become partners, helping business teams bring ideas to life.

3. Quick wins over big, complex projects.

Rather than spending years building a massive AI-powered system, companies should focus on small, high-impact tools that deliver immediate value.

Let me give you an example—document processing for loan approvals.

At Carro, we want to provide the best experience for customers applying for financing. The process involves assessing a customer’s credit score and determining whether to approve their loan.

Unlike Singapore, many countries don’t have SingPass, so a lot of documents come in paper form. This means huge teams of people manually entering data into systems—whether at Carro or at a bank.

Instead of building a massive AI system, we focused on solving the front-end problem:

• How can we convert documents into a machine-readable format (JSON)?

• How can we automate data entry to speed up credit approvals?

Once we solved that, the credit team could process applications in real time instead of taking days.

This simple solution reduced loan application time from 40 minutes to just 5 minutes.

We built this in less than a week—one day to prototype, five days to refine the UI, and we shipped it immediately. Our users have been using it ever since.

This is the power of quick wins—small tools that deliver immediate impact, rather than waiting years for a large-scale solution.

Paulo: That’s fascinating. We’re going to see more and more leaders like you—people without an engineering background—being able to build these kinds of AI-powered solutions without a massive engineering team.

Translating Zi Yong’s principles on Gen AI Transformation (in Part 1) to Carro’s journey the last three years

Paulo: You’ve shared some of these principles that you’ve taken from your own background being a kind of quote unquote commoner as you’ve mentioned. 

And I think while that’s very interesting to hear, I think we’ll see more and more leaders like yourself who may not have engineering background but are able to build these kinds of solutions without a huge engineering team or even without engineers necessarily. How did those principles that formed you as a leader translate into the principles of Carro when it came to its own Gen AI journey?

Zi Yong: So maybe to give you a very quick three points. So when I first started off this journey, that was back in 2023. Right now it’s, yeah, back in 2023. Yes, early 2023.

Back then, I think that was when ChatGPT first came out. But we were seeing a lot of projects coming up like LangChain. For those who are familiar with LangChain, it is a project whereby you can chain multiple prompts [00:01:00] together to make it into a workflow.

There were also talks about AGI: SuperAGI, BabyAGI, AutoGPT. Quite a number of tools came up, so it was very clear to me at that point in time that if Carro is to invest in Gen AI, we shouldn’t be building everything ourselves, right? In fact, we should only build things that are Carro specific.

Because even if I were to build a tool myself today, this exact same tool will be available one week later. And it will be even faster than how it will, even faster than myself shipping it.

So instead of trying to do everything ourselves, we want to leverage the community also to do this at the same time build what is Carro specific. To give you some example, when we first started on LangChain, we needed to build a lot of different capabilities together, including things like how do you run calculation on language, how do you do text to speech, speech to text, and also image recognition because we are dealing with multi-format, multi-modal types of conversations within Carro.

Is there something better at that time? I think that is a no. So hence, instead of building all of that, we shifted in terms of building endpoints such as Carro search that goes into our ERP, right?

Why is that the case? Because at the end of the day, if I want to get a task done, or let’s say if I was sales, I want to know that this customer probably can only afford a car with maybe $1,000 to $2,000 of installment per month.

What kind of cars does Carro have in inventory today that will suit the needs of the customer? So we basically built our own tools that serve functions like that so I could query this, and then the sales team will get five or six different inventory available and then recommend to the customer.

The second point is that we prioritize building agent windows over big general chatbots. So this is not to say that we didn’t build those big general chatbots before. We had some company sponsoring us to do it.

Our realization is at the end of the day, it’s nothing more than a glorified FAQ. For the big general chatbot, the best function you can do is just to repeat whatever information you have in your documents: your policies, your SOP docs and things like that.

But there’s a lot of nuances that you will not understand, and all this will be lost in between. Instead of trying to use that and confuse my workforce, we focus on agent windows, which are easier to deploy for us to fine-tune it and make it more accurate.

So some of these examples, just now an example I mentioned — the automated application — was actually built on our own busy backend. So we have a simple backend whereby I can create agents. I’ll write my [00:04:00] prompt into the agents and then all the user needs to do is to upload a document and click a button, and then it returns a JSON result right from the document.

We have done that for numerous different things. We have done that for the application forms I mentioned just now. We have also done that for even our standup meetings.

So imagine you have standup meetings with our sales team twice a day, morning and evening. So how do I govern the team to make sure that all the various customer leads are being taken care of?

We have built a meeting summarizer in which they just need to drop the meeting recording into the folder, and automatically we return the key points. Of course, we have to sanitize for other key points, like how we format the key, and that will automatically be generated on a daily basis.

In the conversation of accuracy, what’s the alternative? You have to review minutes from 10 branches, 10 times two, 20 recordings a day, 30 minutes per recording.

That’s probably one full-time person’s job. Instead of doing that, now, everything is done using RAG and using Gen AI. I can build workflows subsequent to that. So that’s the second part. Prioritizing building agent windows.

The third and last thing is, as I mentioned to you just now, I focus on small and quick wins instead of overly ambitious projects. In Carro, I have to tell you, our time works very differently.

Our time works very differently. We always want everything yesterday, so much so that we think one month always felt like an eternity.

Instead of trying to scope the project so big, and it takes weeks to months to develop, we focus on quick wins and making sure that with each of these quick wins, we can answer to our boss, myself to Aaron, why I get my engineers to spend two hours to solve this problem that we have in this department because we can confirm how much time saving, how much cost efficiency or productivity that we can save.

To get to an agent workflow, we need to make sure all these small processes are there because eventually you need to chain them all. So that one agent workflow can cover for the whole process end to end. It is part of the journey anyway, so that’s how we went ahead and developed our strategy towards Gen AI.

What’s next for Carro when it comes to Gen AI

Paulo: I love that you contextualized the principles within the last three years that you’ve been exploring Gen AI from LangChain and targeting specific workflows. That could immediately show you like real-world results.

You talked about multi-modal models popping up the year after or several months after, and you’re now able to take in more types of input and process that. Now in 2025, what’s exciting for you guys in terms of the Gen AI journey?

Zi Yong: There’s a few things that we’re working on. One thing is an agentic workflow, which I mentioned just now. It is about how you carry the whole workflow from end to end.

How do I automate the whole entire process? Let’s say, for example, I have a customer who wants to buy a car.

Today to carry this customer all the way from buying, starting the transaction all the way to payments to transfer, in between, there’ll probably be four or five different teams that will be involved.

Someone processing the transaction document. Someone who will be handling the loan application and chasing the banks for disbursement. Someone will be responsible for the transfer of the car. Someone would be responsible for the logistics.

Of course, on the physical world side of things, I cannot replace that with AI, at least not yet.

But do I really need so many people in the process to handle documents? So the thing is about how we think about and reinvent how the workflow would be in the future so that the sales team basically has a virtual assistant in front of them to handle all this process for them.

Imagine you have a sales rep directly in front of a customer, instead of telling a customer, “Oh give me a while, let me check with this other colleague in terms of where your loan application is” and things like that. On the dashboard, you would be able to immediately see what’s the progress and which stage we are at.

All these are all handled programmatically, because I have an AI agent that’s managing all this.

This is not to say that we are eliminating jobs. It is just that the nature of the jobs would change in the sense that the person in the process who is handling documents, instead of doing all the manual typing themselves, the manual submission themselves, will probably be the governor of the process.

Paulo: You still need the domain expert to guide that. 

Zi Yong: The domain expert is there to govern the process, making sure that the AI is not making any mistake, making sure that it is going through smoothly. Obviously the most common issue is typos. 

The domain experts do that, but instead of all this process being manual, the operator becomes a governor of sorts to govern that process.

That’s one thing that we are looking at related to this year so that we can better streamline our whole processes as far as our workforce. That’s the first part.

The second part we are looking at is really to get into self-sufficiency when it comes to AI. For Carro, actually, we have a wealth of data points throughout the years that we have been operating.

As mentioned at the very beginning, we always have AI tools in Carro. We have AI auctions, we have AI pricing, et cetera.

We actually also reside on labeled data that we have accumulated over the years. But it’s very interesting, especially with DeepSeek’s whitepaper in January.

We are also experimenting to see how we can take this data that we have and train them into smaller models in which we could deploy, especially offline to our agents on the ground to get things done, whether it is on the inspection front or whether it is in the customer service front.

Doing some experiments in this front to develop highly specific domains and use case specific models to solve the problems that we have. That’s something that’s coming out for Carro in 2025.

Communicating ROI for Gen AI Transformation to both management and frontline employees

Paulo: I think it’s quite interesting that from experimenting and looking at what kind of technologies would work for Carro, it’s shifted to how then can you build the right technologies for Carro’s specific functions and trying to build your own small models.

I think one thing that our audience would definitely be interested in is, after all of these experiments and all these kind of efforts that you try to do to implement Gen AI and see how it can be valuable for your organization, at the end of the day, it’s all about ROI and how you’re, how you communicate that to stakeholders and especially for those who are like middle management. 

How do you communicate that to your higher-ups? How did you approach it? How did you think about ROI and communicate it to the organization?

Zi Yong: There are two parts. One is to the business stakeholders, the management, the board, what Gen AI actually brings to them.

Two is what’s the value of Gen AI that brings to our working level people. I’ll dive on both sides.

So probably to first address the management point of view, I think after rolling out Gen AI, there’s a few other solutions that we have that you mentioned earlier, such as our ticket QA bot.

We actually audit all our agents’ communication with the customer, to make sure that they do not say the wrong thing or perform the wrong action with the customer.

The ROI that we get at the end of the day from a business standpoint is much better conversion rates, because we can identify problems early, we can take preemptive actions to train our agents that are making all the mistakes, and we can do service recovery.

In a sense, if there is a problem that our agent has generated for a customer, we can do quick service recovery for the agent. At the end of the day, with the same spend that we have in marketing to acquire customers, I’m getting almost double in terms of the conversion rate.

It is all down to governance, down to people management, and down to mindset shift. Do I actually need Gen AI to be a hundred percent accurate in this case, do I need Gen AI to have guardrails because they may say the wrong thing to the customer?

I don’t need that because it’s internal facing. It is about changing people’s behavior, changing management of the company.

With that, we have seen that kind of improvement in terms of results. And of course, I mentioned just now about cost savings.

The RAG workflow is saving us easily 2000 man-hours per month, with just a few hours of work from an engineer’s perspective.

So that’s pretty much how we manage the ROI from a management standpoint. And these are numbers that I can quote you and say, this is what we have achieved in the last one year.

On the people front, the ROI is about how I can help them to minimize the mistakes at their work? For most companies, when you make mistakes at work, you get penalized.

Garbage in, garbage out. Going back to their point earlier, every input into ERP, if the input is wrong, then naturally the rest of the process will get wrong and it could have a financial impact on the company.

So as per any company, we are also very strict and disciplined in terms of data cleanliness and making sure that our team members are on the ball. With the workflow I mentioned to you earlier, we are able to help them reduce their error rates.

They are spending so much time, or rather they have eye fatigue from eyeballing every single thing and making sure that they do the right data entry. All these things get automated so that they can focus on the real value creation in their job.

That’s the ROI that we bring to the team for them so that they are able to do their work better and more effectively, which in turn helps them to get whatever incentive that they get from their productivity. The output is there.

Paulo: I really like the word that you used earlier, governance, like individual kinds of workers having governance over their workflow. And in a pre-AI world, there’s a lot of factors outside of your control that might impact your own kind of results and accuracy, even as you mentioned. But I guess with more kinds of automated workflows and Gen AI, you can have, I guess, greater ownership over these workflows.