The Company Predicting Whether You Will Return Something | Seel on Lightwork

0:00Lightwork Host: You know, people talk about how Vibe coding enables, you know, this one person billion dollar software companies. We build the infrastructure that enable one person billion dollar brand.

0:08Zack Peng: Today, returns are one of the biggest pain points in e-commerce. By SEAL's math, up to a quarter of everything merchants sell. And Zack decided that it wasn't a logistics problem, but an underwriting one.

0:20Lightwork Host: I was a philosophy and physics major in college, and that experience is where I learned how to code.

0:25Zack Peng: He left Orbital Insight to found Seal in 2019, starting with Return Assurance and built towards an AI operations platform for commerce.

0:34Lightwork Host: Seal is very deliberate about building agencies and not just agents. The vision of the business is to power the global GMB with this digital AI workforce.

0:45Zack Peng: Today, the platform runs support, resale and compliance for thousands of brands and marketplaces, including TikTok Shop and Poshmark.

0:53Lightwork Host: I saw this really interesting stats a couple of days ago that the agentic traffic has exceeded human traffic. We talk about it on the show all the time.

1:01Zack Peng: Zach Peng, welcome to Lightwork. Zach, it is so great to have you on here. We're very excited to discuss SEAL. It is quietly powering something that I think most consumers don't necessarily think about day to day. And as someone who does a lot of online shopping, as we dug a bit more into SEAL, what you all were powering, I just realized there is so much behind the scenes work that goes into powering our shopping economy. And I can only imagine how this continues to evolve with all things agentic commerce. So really excited to have you on.

1:35Lightwork Host: Yeah, no, thanks for having me. I'm excited for this.

1:37Zack Peng: I guess even to take a step back before we dive into your founder journey, how you think about the ecosystem, it might be helpful to just give the audience a sense of what SEAL does and the problem that you're tackling and maybe make it very tangible of like an example that, you know, if I'm buying something on Hoshmark or a TikTok shop, for example.

1:59Lightwork Host: Yeah, absolutely. SEAL is an AI operations platform for commerce. When a consumer makes a purchase from a brand, a marketplace, a retailer, there is an incredible amount of work that goes on behind the scenes. There needs to be a fulfillment team that fulfills the order. And before that happens, a marketing team runs a whole lot of operations to reach the customers. And some inventories gets returned. A lot of people have questions. The access inventory needs to get resold and liquidated. So there's an incredible amount of work that goes into moving trillions of dollars worth of products across the world.

2:36Lightwork Host: And the operational cost of moving those trillions of GMV is in the hundreds of billions every year. And what SEAL has done is we created an AI workforce that enables a lot of these complex digital commerce operations without needing to build a substantial global operation. The way I think about it is, just like people talk about how Vibe Coding enables this one-person billion-dollar software companies, I think we build the infrastructure that enable one-person billion-dollar brand.

3:10Zack Peng: And I think something that came up for me that was a really helpful analogy is almost like if you have Shopify as the storefront for these brands, Stripes is sort of this payment layer and then Seal is just everything operations. Is that the right way to think about it?

3:25Lightwork Host: Absolutely. Yeah. I think opening a store has become incredibly easy, right? Like it used to be that, you know, you have to have an engineering team build a website. Shopify made that super easy and then accepting payments was a huge deal back in the days. And, you know, now that's incredibly easy. Now, if I want to start a brand, if I wake up today and decide to start a new brand, I can get everything set up by lunch. But it is still incredibly hard to operate an e-commerce business, the operations that goes into it. I don't think a lot of founders and founding teams really started the brand because they want to be answering tickets or thinking about, you know, fulfillment details.

4:00Lightwork Host: You know, we want to empower brands to do what they do best, which is building physical products that people want. We'll take care of the boring stuff.

4:08Zack Peng: I think it's almost like in the era where cost of creation is so low, then everything hinges on that execution layer. And you all are just making that execution layer so much more seamless. Absolutely. Yeah. And Zach, on this show, we're used to talking about venture returns, but not so much about e-commerce returns. And your background is quite interesting. You started as an engineer at Orbital Insight analyzing satellite imagery. How did you go from that to e-commerce and returns?

4:42Lightwork Host: Yeah, Open Insight was the first and only company that I really worked at. And one, it was an incredible experience for me because I was a philosophy and physics major in college. And that experience is where I learned how to code. A second profound realization I had was the impact that algorithm can have on financial service businesses. So OpenInside was getting a lot of data and they built these incredible algorithms that can predict a lot of events for hedge funds, insurance companies, etc. And the first wedge that SEAL had was we took that algorithmic approach to e-commerce returns and refunds.

5:21Lightwork Host: We saw there was a whole industry around payment chargeback. And I thought, OK, payment chargeback is less than 1 % of the total GMB. and return was like 15 to 25%. So if there are a lot of investment that went into being able to predict chargebacks, why wouldn't people sort of build algorithms that predict returns and refunds? That was the starting insight for the company. So I think we were one of the first teams in the world to try to use pre-purchase behavior data to predict returns and refunds. And that sort of snowballed into a whole operational capabilities that we had to build in-house and eventually morphed into the AI operations platform that we are today.

6:07Zack Peng: So what I'm hearing is, in short, you all have built this AI that is very good at predicting what gets returned or not. I'm curious, what kind of signals are you taking in? Or is there even a weird signal that you've found to predict a return?

6:23Lightwork Host: Yeah, absolutely. All of this happened before the whole LIM wave started. And it was like deep learning and convolutional neural net for computer vision was sort of the thing back then. And the characteristic of those algorithms was that it was very much a black box. And so we took the approach of maximizing information to the model and start from there. And then we started fine tuning the model. We ended up using about 200 different signals pre-purchase to predict returns and refunds. So some of them, you look at them and you see, okay, that makes sense. Some of them are pretty surprising. My favorite one was for apparel specifically, if there's a human model in the image, that apparently made a huge impact on the returns and refund probability.

7:09Lightwork Host: I would say most of them are probably things you can think of, like how many seconds someone spent looking at a product page or what other products they like to buy, all of those things sort of feed into the probability of return in refunds. So you talked a little bit about how in the early days

7:28Zack Peng: you had deeply negative contribution margins while the model was learning. What kept your conviction and where was that turning point where the underwriting started to pay for itself?

7:41Lightwork Host: Yeah, honestly, it was the scale of the data. It's incredibly, there's this concept of curse of dimensionality that as the number of dimensions of the problem increases, the complexity of the problem increases exponentially. So when you have 200 different dimensions that go into predicting returns and refunds, it consumes a huge amount of data. There needs to be a massive number of orders that we have both pre-purchase insight into and the post-purchase label of whether or not it ended up being a return. So the conviction we had was really scientific. It's that I had deep conviction that return and refunds are fundamentally a predictable problem.

8:25Lightwork Host: And really, it's about scale. And once we got to scale, that's when sort of a lot of the problem resolved themselves. And, you know, we got to fine tune the algorithm as we started seeing more brands and marketplaces cross different verticals. Like, you know, the return probability of someone buying a pair of socks versus buying a refrigerator are obviously very different. I remember in the early days, we optimized specifically to work with marketplace a lot. Marketplaces. Because marketplaces represent a broader inventory and broader category of goods. that really helped accelerate the velocity of product innovation.

9:04Zack Peng: Is there things that you've learned from working with maybe the smaller, you know, smaller shops to some of these larger marketplaces and bigger companies that you now work with today? And maybe talk to us about a few of those brands and who they are.

9:18Lightwork Host: Yeah, no, absolutely. Working closely with like startup brands helped shape the identity of the company in two ways. One is as a product slash engineering founder, I got to work really closely with a large number of brands and really spend a lot of time with the founders. That helped us have the identity of being a very product-focused company. I would say even to this day, I spend 70, 80 % of my time on direct product-related work. There's no sort of right or wrong approach to enterprise eventually. I think we stayed really focused on product as opposed to other aspects of the business. The second way that I think the early SMB days influenced us was focus on simplicity.

10:03Lightwork Host: That became even more important in the, I think, the post-LIM days. When we get into things like building support agencies and compliance and marketing agencies for brand and marketplace partners today, there is this push to build a lot of customized features. And then if you're not careful, they can quickly snowball into a complex piece of software that nobody knows how to use. But working with SMB really taught us the difference between getting 99 % of the way they're in terms of performance versus like 50, 60%, the ability to get users to use the product well. You simply can't achieve that without simplicity in the product.

10:46Zack Peng: I love the way that you phrase that because I think so many of the greatest founders, one of the common themes is that it is on them to figure out how to create a simple product. Like that user delight should come from the product itself, not from them having to navigate all of these frictions and building a team to figure out how to use your product. And that stands out so clearly in the way that, you know, I think Seal has become a very well-loved product in the e-commerce space. I think to just make it a little bit more tangible for folks who are less familiar, I want to get a little bit into the business model as well as why this is such a core product for any sort of merchant.

11:25Zack Peng: One, how do you benefit the merchant? How does SEAL technically work in the payment process? What benefits do merchants get out of this? And then two, how does that translate into the broader product and the suite of tools that you offer them? After they kind of get that, the underwriting of potential return, how does that extend to the other product suite that Seal now offers?

11:50Lightwork Host: At the very beginning, we had a choice of either selling the return and refund signals or selling liability. And we chose the latter because, you know, it's really hard to convince people that their signal is accurate. We took a rather unusual approach looking back. So we would approach brands and marketplaces with a free product. We would not charge brands and marketplaces anything. Instead, we would be a checkout add-on as part of their checkout flow. So consumers that want an upgraded return and post-purchase experience can pay a few extra bucks and get the option. And for the brands, it's a free product.

12:25Lightwork Host: If they have final sale things on their website, this would cover the final sale. if they have, you know, two day turnaround time for email support, we offer all day 24 seven live cost of support, right? So we try to put the post purchase service in a box as an upgrade to whatever default service that is already offered. So the go to market is powered by a free product, the free checkout add on the checkout add on had I think of it as having two sets of features. One is an insurance backed refund policy, you know, insurance is just, you know, the best, the highest attaching checkout add-on of any category.

13:00Lightwork Host: That's why you see them in when booking flights, when booking hotels, when buying an iPhone. Insurance-backed refund policy builds trust with consumers, which drove a very high attach rate at checkout. And we also had a dozen different post-purchase features. 24-7 support the option to resell a product with one click after you've used it for 90 days. Things like the ability to make the shipment carbon neutral, just a dozen different features that bring the shoppers back. And when we bring shoppers back, these are moments of service that can also be moments of sales. We turn these moments and help the brands sell more products.

13:39Lightwork Host: So if someone asks, hey, what's the right camera stand configuration for this camera I just bought? Instead of just giving some numbers, it's better to just share the product link. Effectively, brands benefit from working with us in two ways. One is they save on the operating expense around support and chargeback. The post-purchase service handles that at that level. And then they also enjoy post-purchase traffic that we bring back. So for us, it is a model of go-to-market with a free product. And then once the brands get to see how we interact with their shoppers, we also become an extension of their team.

14:17Lightwork Host: And all of the services we provide to shoppers are powered by AI, run by the AI workforce. That model has served us tremendously well in the last three years.

14:25Zack Peng: Yeah, I think it's really smart to be able to start with that moment of very clear signal for the brand. But then very naturally, it does act as such a seamless wedge into providing all of this value with agents, which we know that for a lot of these brands, this is the type of stuff that they don't want to spend. It is just so much execution rigor that being able to have something like Seal out of the box is such a game changer for them. I'm curious with the brands that you've worked with, obviously you have all these chatbots executing on their behalf. You have now a whole swarm of agents that are operating.

15:00Zack Peng: How do you think about who's accountable when things might go wrong and how do you ensure that those instances happen less and less?

15:06Lightwork Host: Yeah, it's a really great question. SEAL is very deliberate about building agencies and not just agents. Take support as an example. If we create a tool as a platform to train AI agents and put the customer, the brands, in a position of having to train those agents, the brands are responsible for prompt engineering, setting up task cases, and calibrating the performance, etc. And one thing we realized is that, you know, in traditional support BPO, the best performing agents do not stick perfectly to the SOP, the scripts. They have the context. The unstructured context is what makes a great support rep great.

15:47Lightwork Host: Those contexts are missing in the sort of the agentic platform workflow because, you know, a operations manager would have Slack, have, you know, Zoom meetings, have all these unstructured conversations, providing context to a BPO team. There isn't really a good way of doing that to agents. You know, agents require, you need to, you know, specify the context that they get, you know, the prompt you inject, and then, you know, what data you feed it, what things you don't give it access to. And that last mile is why a lot of the agentic products stop at, for example, 50 to 60 % resolution rate. And we knew that because we tried virtually every agentic, you know, support products in the market before.

16:30Lightwork Host: You know, we needed to serve. We need to provide service to tens of millions of consumers around the world. So we were a user. When we realized that, okay, the bottleneck is not in the technology, but in the way that information is shared to the agents, we thought we want to build agencies. And if you think about how an agency operates, when an agency takes on a set of KPIs and goals and they don't achieve them, it is not the fault of the client. It is the fault of the agency. But if you sell a product, you could always blame it on the user. It's very common in enterprise P2P software. We cannot do this because that information was not shared with us, so that context was not given to us.

17:11Lightwork Host: Well, as an agency, we don't have that excuse. If we don't hit the goal, we don't hit the goal. So we were very deliberate about building AI agencies that can complete an entire project and not just execute on very tactical tasks. tasks.

17:27Zack Peng: Is there a say, you know, since you all are covering such a horizontal stack of AI operations for all commerce, and I feel like you all have been very intentional in the way that you build out all of these different solutions for these brands, is there a part of the stack that you think SEAL will never try to tackle?

17:47Lightwork Host: Yeah, the world spends, you know, $200 billion every year on just, you know, on the digital commerce operations alone. That does not include warehousing and shipping, right? So that's just the marketing, the support, the vision of the business is to power the global GMB with this digital AI workforce. Physical AI has its place. We don't necessarily have the most expertise there, but from having served tens of millions of shoppers ourselves and from having worked with thousands of brands, I think the digital commerce operation is the part that we'll stay focused on.

18:20Zack Peng: I was going to ask, since you do have such a unique window into this ecosystem, you've been building in it for many years now and have seen evolutions in the space. One thing that kind of is top of mind for a lot of people is this idea that we are now increasingly not only having human shoppers, but agents shopping on our behalf. I'm curious for your take on how this changes the ecosystem and how your business is also reacting to this new era of agentic commerce.

18:46Lightwork Host: I saw this really interesting stats a couple of days ago that apparently the agentic traffic has exceeded human traffic. We talk about it on the show all the time. Yeah. There's no question that that's happening in e-com and the majority of the orders and also post-purchase. I think there's very little question that most of the post-purchase traffic will probably end up being agentic. And how we think about it is shopping by itself is really pleasant. People enjoy it. It's not a chore. People enjoy shopping because of the discovery, et cetera. Nobody likes the legwork of having to click through Term Portal, having to email back and forth on issues.

19:26Lightwork Host: So our view is that agentic commerce, most of the legworks will be done by agents. And humans will get to enjoy the pleasant part of shopping a lot more. So we actually wrote out a consumer app ourselves. We think of it as the AI shopping sidekick. So it has things like, you know, price drop monitoring, one click return as a virtual try on. Basically all the things that take the legwork around guessing sizes, you know, returns and, you know, the unpleasant stuff. Take all of that away and help users focus more on the enjoyable part of e-commerce, which is the discovery and the joy of finding products you truly love.

20:07Lightwork Host: Is that like a next generation of like honey and some of those types of platforms?

20:12Zack Peng: or how are you seeing that platform emerging?

20:16Lightwork Host: I think for us, it is more about reducing the friction between brands and consumers. There are a lot of companies building, you know, the Chrome extension and specific features.

20:27Zack Peng: There's lots of virtual try-on already, but I think what is it that makes it unique?

20:33Lightwork Host: There are two things. One is we think of the SEAL AI Sidekick consumer app as the aggregation of long tail demands, right? Every shopper's priority is different. There are some shoppers that want the best value for their money. Other shoppers want to be the first to try something when a new product drops. It's hard to get distribution for these individual features. But the idea is that maybe in aggregate, that becomes a compelling offer. So in one place, you get to find the best price. You get to be the first to try a new product. The aggregation of all these features that remove the friction in e-commerce, that's the primary thing we're betting on.

21:10Zack Peng: What is becoming so clear is like between both the consumer shopper relationship and also the brand itself, you all are building all of these levers that allow it to kind of go true to what we are there for, which is on the brand side or what these brands and consumers start the transaction for in the first place, which is on the brand side, building a great product, delivering that product. But they don't want to do all of that operational heavy lifting. And then for consumers, same thing where it's, you know, you enjoy the discovery part, but there's so much, again, in that shopping process that is very logistics, very manual, lots of link clicking, filling out forms, talking to customers.

21:50Zack Peng: So much headache that goes into like filing a return or having to deal with something final sale that doesn't fit. Yes. It does feel like you guys are solving a lot of problems, both on the merchant side and the consumer side. Yeah, it's really exciting. And I can already see, you know, it's subtle because it works so seamlessly that it often runs in the background. But that's almost, I think, maybe a sign of a great product is that you don't even think about how many problems it's solving for you on the back end. I guess to maybe round out, I'm curious if you have any one story of a customer moment that sort of captures what SEAL is becoming for the ecosystem.

22:29Lightwork Host: We're just starting to see this, but I think this is a really exciting trend. We also are starting to see a lot of starting Shopify brands that are like with one or two people doing the volume and work that typically requires a 10 to 20 people team. Pretty sizable portion of the GMB is cross-border. It would not be possible just five years ago for someone to run a brand that sells to 20 different countries with, you know, fulfillment and marketing and compliance and support that goes into, you know, service customers and all these different time zones and market. We just started rolling out this subscription for starting Shopify brands, you know, essentially give them access to AI workforce around all these different areas.

23:13Lightwork Host: And we start to see them doing volume that typically require a 10 to 20 people team. What this says about the future is in the past, building a brand means building a global, complex business operation. And now very much like, you know, how a software engineer or product manager can build products that touch millions of people with cursor and clock code. We are starting to see folks, you know, build incredible brand businesses that touch millions of consumers. across the world. And I think the future of, you know, a one person billion dollar brand isn't as far as we used to think it is. If someone heard that, you know, in 2026, one person can build products for millions, that would be ridiculous, you know, just five years ago.

23:57Lightwork Host: And I think in five years, one person billion dollar brand would just be as normal as a Tuesday morning with the right infrastructure.

24:05Zack Peng: And I can't even imagine what that looks like in five years, 10 years. And I think to what you were mentioning, just a common theme that comes up for me is that of all these tools and offerings like a seal is that you are consistently lowering the floor, but also at the same time, raising the ceiling of what's possible. And you see that in the GMV that you're unlocking for a lot of the merchants that you work with. And in the same way that I think you've seen that in other verticals and coding and product, all of that. But cool to just see that in very tangible ways, even for commerce, which is, I think a little bit further away from the heart of AI and all things coding.

24:41Zack Peng: And yeah, I'm excited to see that billion dollar brand start by one person. You're getting me excited. I'm like, I can't wait till Josh and I have our own brand. Right. We can start our product. Let's launch something when you're powered by Seal in a Shopify shop. Let's do it. Thank you so much for joining us. This has been so interesting. I think, again, just being able to peel back everything that happens on the back end when you have a single transaction and Seal just being such a core part of it. But it's so interesting to hear about your vision for the future of commerce and how people transact, especially as we have agents now operating in it and just seal powering all of that.

25:18Zack Peng: So thank you for joining us and hopefully many more conversations to come as this all evolves. No. Yeah. Thank you, guys. Zach, appreciate you coming on Lightwork. All right. Thank you.

25:37Thank you.