Join us on this call led by Insignia Ventures Partners venture partner Timothy Lee, which features an exploration into the realities, possibilities, and limitations of integrating AI in various consumer experiences from buying cars, to buying homes, from buying FMCGs and cosmetics to buying stocks.

With him are the CEOs of leading consumer platforms in the region: ⁠Carro⁠ CEO and co-founder ⁠Aaron Tan⁠, ⁠Ajaib⁠ CEO and co-founder ⁠Anderson Sumarli⁠, ⁠Pinhome⁠ CEO and co-founder ⁠Dayu Dara Permata⁠, and ⁠Super⁠ CEO and co-founder⁠ Steven Wongsoredjo⁠.

Editor’s Note: This episode is a edited recording of a panel at the 6th Annual Summit of Insignia Ventures Partners

Highlights / Summary

Five things you need to know about the reality of leveraging AI for consumer experiences in Southeast Asia

1/ As a consumer platform, it’s not enough to just have the right data to build AI applications. There needs to be a way for the market to “agree” with your data. 

In the case of Carro, this factored into their marketplace approach to auto retail from early on. As CEO Aaron shares, “While machine learning can be extremely accurate, the truth is that the consumers will never agree with your prices. So we scrapped the plan of starting off with a data approach first, and we started off by selling the cars first, and embedding this within the banks. So today, most of the banks have an MOU with us to find out the pricing on vehicles. And that sets up the benchmarks for this. So we are able to say that our prices are accurate and people actually follow it.”

2/ AI has the potential to drive down costs to serve consumers, which does not only benefit the end-consumer but also the platforms with better margins. 

For example, it can cut down 1000 chats, 15 agents, and more than 50 listings in a typical home buyer’s journey in Indonesia to a much more streamlined process saving around US$100 per transaction. 

As prop tech Pinhome CEO Dara shares, “As we are a three, four-year-old company, there has been some evolution in how we leverage AI. We first started with rule-based predictive AI, but now that we have so much development and advancement in the areas of LLM, Natural Language Processing, that provides a very good advantage for us…From our side, as a platform that really reduces a lot of cost. So that $130 per transaction now can be reduced to about $20 to $30 per transaction for a hundred thousand dollar property.”

3/ Even if AI is operating in the background, it can come at the cost of user experiences, especially in rural areas where digital adoption is not as advanced. At the end of the day, all these innovations are for the consumer.  

As Super CEO Steven shares, “It’s very important to manage the balance between innovation and product-market fit. If you’re being too innovative in the rural areas, then no one’s gonna use it. Your AI is just gonna become a trophy in the cabinet. And for business decisions, it’s also the same thing as well. Based on my understanding, building an AI after rural areas is like building a skyscraper. You need a solid foundation before you get there.”

4/ At present, market opacity still requires human presence to supplement AI output, especially when it comes to decision making. Such was the conclusion for Ajaib’s investment platform, pairing up their AI-generated content with licensed investment advisors. 

As CEO Anderson shares, “Actually today, more than half of that content is generated by our AI…and we’re increasing the proportion of it, going from 50 percent hopefully by next year going up to 90 percent…But one of the things that we also need to be careful of is not to go ahead of ourselves. I think that we have held ourselves back from going to recommendations, even if there are experiments that we’re running on it. It has to be paired with our human registered wealth advisors and so forth like that.”

5/ Externalities like consumer behavior (e.g., their propensity to follow through on commitments like loan repayments) and regulatory changes (e.g., taxes on property driving down prices) need to be factored in when deciding to what extent AI should be implemented. 

As Dara shares, “So AI is really good at estimating intrinsic value, if, let’s say external factors are all constant and it’s not behaving erratically. But when you add in extrinsic factors…AI is not able to capture and those are so completely unpredictable.”

As CEO Aaron shares, “AI can do a lot. But do consider this whole concept of whether there is that similarity between markets. You can draw that and say, “Oh, credit scores are great.” But the truth is, for these people, they do not care because there is zero switching costs.”

Timestamps and Highlights

Highlights;

Mis-Introductions (by Claude, ChatGPT, Bing, Bard);

Carro was initially an AI company for testing sound engines? And how it’s still an AI company today;

Real estate pain points that can be solved by AI and Pinhome’s cost reduction with AI;

What AI means to rural Indonesia and Super’s approach to digital adoption;

AI reducing cost to serve and the value of still having humans in the process for Indonesia’s largest digital investment platform Ajaib;

Where have you tried to apply AI and it hasn’t panned out as expected?;

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.

Transcript

Note: Transcript has been edited for concision and clarity.

Carro was initially an AI company for engine sound testing?

Timothy Lee: I hope you have a feel for the types of founders here who are creating these consumer experiences in Southeast Asia. It takes a special type of character, and now we know them a bit better than the general collective AI.

I want to start by asking a few questions about those consumer experiences, but first, a little bit of a story. 

Aaron, I believe when you started Carro, you actually were intending to start as an AI company testing engine sounds. Can you tell us how that came about?

Aaron Tan: The engine sound story is always an interesting story. So my chief scientist Bryan, — I’ve known him for like almost 20 years — one day I was in the park with him, and then I asked him some random questions that I always interview my data scientists for, and I always ask them sometimes, “How do you know whether the engine sound is actually okay?”

In other words, when you just record the sound, how do you know whether the engine is okay? And most of the data scientists can never actually answer me that question.

One day, I learned how. Again, I was in a park with Bryan, and then I asked him the exact same question, thinking that he doesn’t know the answer. And I asked him, and he told me, “Oh, all you need to do is just record the sound, and then you convert the sound from a three dimensional domain to a two-dimensional domain using this technique through signaling that’s called the Fourier transformation. 

And then after that, you just use time machine learning on it, and then as a result, you are able to tell whether or not the sound is okay, if it sounds normal or abnormal.”

So when he first answered that question, I was thinking, how does he know all these things because this is some guy and his PhD, which is from Cambridge, is on nuclear science. It has totally nothing to do with data science or AI or computing or machine learning.

And he told me this is because sound is physics. So this is a domain under a physicist, and as a result, he knows all these things. And when he was doing his PhD thesis, it was actually to check whether or not the nuclear reactors [were operating] properly.

That was really how we started this whole conversation. And that’s how today as much as possible, every car that we buy in, we try to make sure that we listen to the engine sound to ensure that the car actually is functioning properly or will function properly moving forward.

And that’s one application of what we do using AI across our business.

Timothy Lee: So it really works. Bryan was right.

Aaron Tan: Bryan was correct. And he actually did the model for it.

Timothy Lee: That’s pretty fascinating. Of course, now Carro is not known as just an engine analysis company, but the full service from buying a car to owning one and financing one.

So in that customer journey, where is AI applied?

Aaron Tan: So we try to look at this really as a tech company. 

The backstory here is that one day in the very early days of this business, I asked one of my engineers to return back to the office — he was from Malaysia — and then he said, “Oh, I went back home, and then it was Chinese New Year. And I told my friends what I was doing.”

And then I asked him, “So what do you tell your friends about Carro? How did you describe the company?” 

He said, “Oh, that I work for a car dealer.” 

And from then onwards, every single town hall, — this was very early on in this business, around 2016 — I’m going to remind everybody we’re a tech company first, and that’s the truth.

So when we started this business, we thought of this as like a Carfax / KBB business. In fact, we have owned this domain called graph.co for the longest period of time. And we wanted to just try to understand the pricing of vehicles, the condition of vehicles, et cetera.

But we just couldn’t get it done because at that point in time when we started, we made the worst of mistakes, which is we didn’t have the data. Not only do we not have the data, we didn’t have data that was considered editorial grade. And while machine learning can be extremely accurate, the truth is that the consumers will never agree with your prices. 

So we scrapped the plan of starting off with a data approach first, and we started off by selling the cars first, and embedding this within the banks. So today, most of the banks have an MOU with us to find out the pricing on vehicles. And that sets up the benchmarks for this. So we are able to say that our prices are accurate and people actually follow it. 

We really think about the whole journey and think about how we use data to better the whole process. It may be machine learning-driven, it may not be machine learning-driven, but the whole emphasis for us, in at least the past last one and a half, two years, has been how do we use data to better every single step of the process and how do we step change all the functions that we have one by one, from the start of buying a car, pricing the car, reconditioning the car, to selling the car, financing the car, doing insurance after, the whole spectrum of it. 

How do we put data to work? That’s high-level what we do at Carro.

“So when we started this business…and we wanted to just try to understand the pricing of vehicles, the condition of vehicles, et cetera. But we just couldn’t get it done because at that point in time when we started, we made the worst of mistakes, which is we didn’t have the data. Not only do we not have the data, we didn’t have data that was considered editorial grade. And while machine learning can be extremely accurate, the truth is that the consumers will never agree with your prices. So we scrapped the plan of starting off with a data approach first, and we started off by selling the cars first, and embedding this within the banks.”

Real estate pain points that can be solved by AI and Pinhome’s cost reduction with AI;

Timothy Lee: Now, Dara, I want to ask you a little bit about, from buying cars to buying a home, and maybe even financing a home, there are also some similarities. And one of the things about buying a home is that it’s very conversational. Usually, you have a lot of back and forth. I’m curious about that process. How are you applying conversational AI in your journey?

Dayu Dara Permata: Typically for a renter to commit to a transaction, they would converse 1000 chats with 15 different agents, and for a buyer, they have to converse 3300 lines of chats with agents, more than 15 agents for Indonesia’s standard. 

So what we do is we are building a platform for home seeking, home financing, and home services. We serve first-time homebuyers.

How the journey started is on the supply side you have owners, property agents, and property developers uploading their listings. And then on the consumer side, they will discover those listings through our mobile app and website.

And how AI is basically being leveraged is, in the process of renting or transacting, there are six steps usually to get to closing the transaction. 

You first need to discover the listing, then you have to connect with the agent. You will then schedule viewing. After that, you will explore financing, negotiate pricing, and then at the end, it’s payment.

And every step of the way, if all these processes are done manually, and before our platform, it was the case, this is a three-month process, and it would cost about $150. to process from inquiry all the way to closing. It would involve more than 15 agents, more than 50 listings. It’s extremely complex, time-consuming, and also very resource-consuming for consumers. 

So we’re embedding AI — and of course, as we are a three, four-year-old company, there has been some evolution in how we leverage AI. We first started with rule-based predictive AI, but now that we have so much development and advancement in the areas of LLM, Natural Language Processing, that provides a very good advantage for us. 

We don’t have to spend like hundreds of millions of dollars to train our large data sets, and then we use that readily available technology, and instead we focus on optimizing and enhancing and fine-tuning that technology so that it can fit real estate use cases. 

So when users are looking for listings, we build capabilities with AI to allow for accurate property recommendations, enhancing listing quality by, let’s say, removing watermarks automatically. Or generating descriptions of listings automatically that really helps with supply liquidity connecting with the agents that automated allocation started with, of course, rule-based predictive AI. 

And now we’re venturing into really leveraging data sets, understanding where conversion happens the best, what kind of archetype of user versus agents, and that also will enhance that. And then in financing, allocating to the bank that gives the highest likelihood for approval, and lastly, negotiating home value estimation. 

So that’s on the consumer side. How can we make the search for property shorter, more delightful, more efficient? From our side, as a platform that really reduces a lot of cost. So that $130 per transaction now can be reduced to about $20 to $30 per transaction for a hundred thousand dollar property. 

“As we are a three, four-year-old company, there has been some evolution in how we leverage AI. We first started with rule-based predictive AI, but now that we have so much development and advancement in the areas of LLM, Natural Language Processing, that provides a very good advantage for us…From our side, as a platform that really reduces a lot of cost. So that $130 per transaction now can be reduced to about $20 to $30 per transaction for a hundred thousand dollar property.” 

What AI means to rural Indonesia and Super’s approach to digital adoption

Timothy Lee: Thank you for sharing the data, and I did not realize there were that many text conversations in the process of buying a house, and also that cost reduction is pretty stunning. 

Steven, I know that a lot of what Super does — and thank you for clarifying exactly what you do earlier — is when you apply AI and data science to it, it’s in the background. But can you explain how you are using AI in that customer journey as well?

Steven Wongsoredjo: So there are two things that AI could be applicable in our business, but the key of executing business in the rural market is simplicity. There is the customer-facing aspect and there is the business decision aspect.

For the customer-facing aspect, it has to be very simple, meaning that you have to understand the cell phone that they are using are the second hand devices that the first-tier cities used. So if you’re going to the second tier, they’ll be using one cell phone behind us. 

Back in the days, there was the BlackBerry. While the BlackBerry was being used in the rural areas, we started using an S1, S2 Samsung, or back then iPhone, the early smartphone. And then it has always been like that. So that’s number one. 

Timothy Lee: Your customers are largely in tier 2, tier 3 cities, not in Jakarta. So we have to make sure we’re in the right mindset.

Steven Wongsoredjo: Exactly. In the rural areas, the GDP per capita is lower than $5,000 USD. The more things you put in the app, the harder it is for you to get to product-market fit. So it’s very important to manage the balance between innovation and product-market fit. 

If you’re being too innovative in the rural areas, then no one’s gonna use it. Your AI is just gonna become a trophy in the cabinet.

And for business decisions, it’s also the same thing as well. Based on my understanding, building an AI after rural areas is like building a skyscraper. You need a solid foundation before you get there. 

The first stage is building a bank of data. The second is deep learning, where this deep learning will assess the data and interpret, extract, and give you some advice. 

Super right now is building that bank of data and in the inception of deep learning. So we’re not yet fully implementing AI. Why is that? Same thing. People who are executing things in the rural areas are not as advanced as in the capital cities. 

So even if you integrate AI throughout your system, it won’t work if they don’t even understand how to bank the data properly. So that’s the thing that we need to tackle before moving forward. 

But then, if you’re asking me about the future, I think AI will help us in making business decisions. Let’s say, it’s summer, and in the eastern area, they like spicier food. AI would then help us find out what kind of spices that they would like to eat, and give us a much better predictable distribution towards the SKUs that we need to sell during those seasons, and so on. 

But at this point, we’re still building towards it. I think we’re not there yet, but we’re at the foundations of it, and it has to be properly addressed before we fully implement AI so the team can execute well and use it properly to interpret, extract, and decide for making better business decisions. 

And at the same time, we have to make sure the device is suitable, and has the capacity to load because the more you put AI-driven functionality in an app, it’s going to slow down the app loading. And if you’ve seen the data, users often drop after 10 seconds. And so once we put one AI, loading can take up 15 seconds or 20 seconds. That’s not the trade-off we’re trying to chase.

“It’s very important to manage the balance between innovation and product-market fit. If you’re being too innovative in the rural areas, then no one’s gonna use it. Your AI is just gonna become a trophy in the cabinet. And for business decisions, it’s also the same thing as well. Based on my understanding, building an AI after rural areas is like building a skyscraper. You need a solid foundation before you get there.”

AI reducing cost to serve and the value of still having humans in the process for Indonesia’s largest digital investment platform Ajaib

Timothy Lee: I’m glad that you provided insight into the consumer mindset, especially in those Tier 2 and Tier 3 cities. Understanding where to apply these types of technologies to enhance their experience is crucial. It doesn’t necessarily mean that they’re asking for it just because we see it in some of the capital cities.

However, one of the areas where we often hear questions about whether AI can be applied is in making money. As investors, this is something we think about. 

Anderson, I know you’ve given a lot of thought to both content generation and investment recommendations as a stock brokerage. Could you please tell us about how you’ve been applying AI in different areas of your business?

Anderson Sumarli: First of all, I just want to echo what Steven and Dara mentioned. One of the big challenges of running a consumer tech company in Indonesia is that you’re going to have to interact with masses, millions and millions of people from the mass retail segment. You’re going to have such a great distribution to reach them with low CAC, and the cost to serve needs to be incredibly low as well.

Now, that’s not always easy because when you go to the second tier or when you go to interact with complex products like property or in our case money, they want to interact with you in person, they want to have more information, they want to have more knowledge. 

So one of our challenges running a stock brokerage or crypto exchange or even the bank is that we gotta find a way to serve the customers in a low-cost way.

And this is where we found opportunities using AI, leveraging it to help us serve folks with lower costs, similar to Dara mentioning about the cost to serve them. One of the ways that we’re doing it is that in Indonesia, because you have 98 percent of the population that’s never invested in capital markets before, the way that they start learning about it is that they have to talk to the community, they have to learn with one another. 

Ajaib was born with a Reddit-style community page already embedded within it, and millions of interactions happen every week, every month. Right now, we’ve been using generative AI to absorb all the complex articles and news pieces, thousands and thousands of them that’s going all over the place online. And we summarize them into pieces that’s very easy for users to digest and interact with.

Actually today, more than half of that content is generated by our AI. It’s no longer user-generated content or UGC. That’s a key success factor for people to start interacting with AI-GC, and we’re increasing the proportion of it, going from 50 percent hopefully by next year going up to 90 percent. 

So people are just chatting and interacting with content, fresh content about news on stocks and so forth, all made by AI. But one of the things that we also need to be careful of is not to go ahead of ourselves. I think that we have held ourselves back from going to recommendations, even if there are experiments that we’re running on it. It has to be paired with our human registered wealth advisors and so forth like that.

But what we’re really excited by is to see how AI would reduce the cost to serve in all the financial services so we can go into the tier to go in with a tier three with a very low-cost model.

Timothy Lee: I’m quite curious about that human in the loop that you’ve kept there, especially for the investment recommendations and also the research. Can you talk a little bit about why you feel you still have to have a human?

Anderson Sumarli: One of the things that we’ve done is that we’re very big on retail customers, but we also have a segment of affluent and also institutions. 

We figured out that people want research reports, equity research reports, market research reports, and actually if you look deeper into those research reports, a lot of it is just a function of summarizing multiple things, whether it’s investor calls, or whether it’s their latest financials, or whether it’s whatever announcement they made.

We actually now generate equity reports very quickly because whenever there’s an earnings call, immediately we take it in, we digest it, and using AI we immediately produce a document. But the only portion of those equity research that we cannot generate fully yet using AI is the recommendation, is the price point, buy, sell, hold, etc. 

Of course, you have things like consensus direction. You can get it from Bloomberg and get it from many places. You can leverage other analysts’ ratings and you can do some technical analysis there. 

But the issue is that in the Indonesian stock market, it’s not a perfect market. So even if you go and look at fundamentals or even if you do some basic technical analysis, it just doesn’t work. So that’s why we have to pair it up with our licensed investment advisors, because there’s no substitute with them sitting in front of a CEO of a public company, breathing the same oxygen, and asking are you credible or not, when you’re telling me about the plans of your company? 

So until you can substitute that, right now, we are pairing recommendations on investments or money decisions with a registered wealth advisor or analyst.

“Actually today, more than half of that content is generated by our AI…and we’re increasing the proportion of it, going from 50 percent hopefully by next year going up to 90 percent…But one of the things that we also need to be careful of is not to go ahead of ourselves. I think that we have held ourselves back from going to recommendations, even if there are experiments that we’re running on it. It has to be paired with our human registered wealth advisors and so forth like that.”

Where have you tried to apply AI and it hasn’t panned out as expected?;

Timothy Lee:  Thanks for sharing that. Now, I’m sure there’s a number of areas — and I think you’ve touched on it in this case — where you’ve tried things with AI. 

And I want to make sure the panel has a chance to talk a little bit about what hasn’t worked too because a lot of this is about experimentation. 

Dara, I know that you’ve also, in the same vein as Anderson, tried to estimate home prices. You’ve got a home value estimator and so on. Can you tell us about some of the areas which have been challenging or where you’ve decided maybe it is not the best place to apply AI? 

Dayu Dara Permata: So I’ve talked earlier about how AI can use and be used to kind of power conversations, right? And there are three types of conversation. There is informational conversation. There’s consultational and transactional conversation. 

And right now, the biggest degree of AI implementation is in informational and conversational. I would say the highest is in informational, asking questions about, “Hey, is it located near certain POIs? Is it in the flooded areas?” 

Consultational would be like, “Hey, can you give me more options with this kind of filtering criteria or characteristic? Or could you exclude properties, let’s say, in a one-lane road because I only want two-lane roads, right?” 

That’s in the consultational, and then in the transactional is, “What is the value of the property? What is the market estimate?” And we’ve tried all those, and I think all of those have worked, and then increasingly now we are improving the degree of usage of AI across those three areas. 

But where AI has not worked [consistently] for us is using that information, for example, for home value estimates, as a basis for acquiring property or investing in property. And the reason for that is because while AI is good at estimating the value of a property, it will give you a price range. 

And that price range really depends on the supply-demand liquidity you have for a certain area where supply and liquidity is high. You have a high number of listings and that number of listings are being transacted, and you see how the market responds to that price. Then the range becomes very narrow to a point where you can be very confident. This is the price for that property and I’m willing to bet for it. 

Maybe in a more developed market, like the United States, where data is very structured and more transparent, it could work, but even then there are still some areas where AI wouldn’t be able to capture and embed that in the model. 

Timothy Lee: So it’s quite interesting because you have the examples of the iBuyer types of companies like Zillow or Opendoor in the US. And what I think I hear you saying is that it doesn’t work great when there’s no structured data.

 Dayu Dara Permata: Exactly. And there’s this lack of transparency or lack of liquidity, whether that’s on the demand side or the supply side. 

So AI is really good at estimating intrinsic value, if, let’s say external factors are all constant and it’s not behaving erratically. But when you add in extrinsic factors, for example, the government just, let’s say, increased the property tax. There’ll be higher supply liquidity that would reduce price because there’s higher competition. 

The government just increased, let’s say, interest rates, core interest rates. That would affect the demand side because then all of a sudden the disposable income and the debt-to-income ratio then becomes more problematic. Hence, less demand. Hence, property prices tend to drop. 

So all these external extrinsic factors AI is not able to capture are so completely unpredictable. So we’ve actually experimented with secondary iBuying — purchasing a property in the resale or secondary property market, and we use AI in that exercise. And what AI was able to give us is a wide range, plus-minus 20% standard deviation rate, which is very risky, plus-minus 20% is scary. 

But then in the primary property market, that could work though because the estimate is a bit more controlled. The cost plus price, and usually developers wouldn’t go beyond a certain point. And that is how we are using, for example, AI, of course, with calibrations and some operational approach as well, to make investments into real estate property development projects. 

And it’s still at a pilot stage, but that could work just because the extrinsic factors could be calibrated well, because we know the cost to build those properties and the land price range could be estimated better for emerging markets.

Timothy Lee: I think that’s a helpful lesson for all of the founders here that there are aspects and areas of the market where applying AI can be risky when you don’t have the right data. How about you, Aaron? Are there areas where you have tried to apply AI and found it hasn’t worked?

Aaron Tan: So we employ AI techniques, AI strategies on various things ranging from computer vision, sound to pricing. We do actually less from a pricing standpoint, but more on credit. 

Because we do have quite a sizable loan book within the company itself. And we spend a lot of time trying to figure out who are the best customers to lend to. And I’m happy to say that our NPLs across the region is about 1% or just under it. So honestly, we do have a very healthy NPL ratio compared to any of our peers. 

But that said, I feel that the whole process of doing AI on credit analysis is not exactly the greatest and easiest, especially in the emerging markets. And I’m not saying that because I feel that there’s a lack of data or lack of historical data. 

I feel that there is a difference in understanding of switching costs. So in the United States, if you ever default on your credit, or in China, if you ever default on your credit, it’s going to be tough to get a credit card moving forward. It’s going to be tough to get a loan moving forward. Life is going to be hard, especially in China. You pretty much get locked out of everything. 

The same cannot be said for Indonesia, and it cannot be said for Thailand or other developed markets. It’s tough. And if you look at first-tier cities, maybe there’s still some switching costs. But if you move out further to fourth-tier cities, it becomes very apparent that the switching cost is not high.

I’m just using the word loosely, switching costs, but it’s conceptually what I’m trying to say, right?

Because you can use all the AI, all the data, all the back-testing, the regressions that you can do and say, “Oh, you know what? I think this guy, I can lend him money, right? Because my model says so.” 

The truth is that this cannot apply to the third or fourth-tier city people whereby the switching cost is entirely different. They can just decide to default the next day. 

So I think coming back to this, we have tried to do AI on many things and I do have to say that there is a higher correlation of defaults for customers, even though the AI might have screened and said otherwise. But there is a high correlation with people in the third, fourth-tier cities simply because of that. 

And honestly, I don’t have a good way to resolve this issue. And in the end, we decided that the best way to do this is to expand the financing business in the hub-and-spoke model. So you have to have someone within 10 kilometers from where the borrower stays. 

And I think that has, at least from my standpoint as CEO of the business, controlled the default rates. And I think that’s something that we always tell people externally. AI can do a lot. But do consider this whole concept of whether there is that similarity between markets. You can draw that and say, “Oh, credit scores are great.” But the truth is, for these people, they do not care because there is zero switching costs.

Timothy Lee: Thanks for that. We’ve come to the end of time unless there’s any other burning questions. I’d like to thank all our panelists here for your time.