Klarna CEO Sebastian Siemiatkowski on Getting AI to Do the Work of 700 Customer Service Reps

0:00I feel it's different with copy and image in copy though, there in the LLMs are much less impressed. And I think the reason for that is that at least how the LLMs work is they work towards the average. So they are trained towards the average and creativity is not the average.

0:36It was 2010 when we first got into business with a young man named Sebastian based in Stockholm. Fast forward to 2024 and Clarner is a global payments and commerce behemoth. Clarner recently made its mark in the world of AI by sharing some of the results of a product that they built for customer facing workflows. Cloud has been one of the more aggressive experimenters in the world of AI, both with external workflows, as well as internal use cases. Sebastian joins us today to say a few words about what they built and where he sees this world headed. BEEP BEEP BEEP BEEP Sebastian, welcome to the show.

1:18Thank you for having me. So you have become a poster child, perhaps the canonical example of putting AI into production inside of your business to make life better for your customers and to make things more efficient internally. And so the thing that everybody is desperate to know about is how did you do it, why did you do it, what lessons have you learned, what are the pros and cons, everything related to the customer support implementation that you guys have done. But before we get into that, unless we put the card before the horse, can Can you just give us two words for people who may not be customers?

1:55What is Clarner? Give us a sense for the size and the scope and the business that Clarner's in. Sure. So, I mean, it basically started as a payment solution for shopping online, often associated with this buy -and -out pay later thing. But actually, today, we do about $100 billion worth of volume across the world. We have a half a million merchants. We have about 100 million consumers. They can all both pay the full amount, what we call debit, and they can pay an installments and use credit and it's also fintech and a neo bank in a sense that we have card sources, balances, the whole thing. We are fully regulated bank.

2:34And this about 4 ,000 employees. Got it. In 100 million plus customers in 20 plus countries, we can see how the customer support implementation got to the scale that it did. I think there are a lot of people who have contemplated doing something like what you guys did. There are very few people who have actually executed against it. And so maybe the first question we'll ask, ask you about this, is how did you do it? Like, how did you guys get into production so quickly with something that seems to be pretty darn effective? Sure. So I think you can start it at the script. I mean, the first thing that happened to me at least was like November 2022 on Twitter.

3:11And then I see people say, you should really try out this thing called ChatchitB. I tried out and I'm blown away. like I was like, wow, this is like amazing. I've never seen anything like this. Or at least maybe when I tried Google 20 years ago, but this is no, this is even more impressive. And so at the up one time, I was like, okay, this is really cool, but then holidays came, Christmas, and then after that, I was like, oh, you know, we have to lean into this and let's see if we can get hold of Sam Oldman and open AI and so forth. And I realized that soon Sam is going to be a person who's going to be impossible to get a meeting with.

3:41So I got to try before everyone else. And so finally, you know, lucky for me, I had Sequoia as a shareholder in both Clana and OpenAI. That was a good opener. I got it and I flew to San Francisco pretending that I was going there for other business, but the meeting Sam was the only reason. Originally my time for meeting was two hours, but the time I got there was only 30 minutes left because the secretary was like down prioritizing my meeting. And I came in and I sat out with Sam and I was like, I got a pitch to him that working with this European bank, FinTech is going to be a great client to test OpenAI products on because what I really wanted to accomplish was to use us as a guinea pig.

4:19I wanted to make sure that we would always try to lay this greatest and that they would find us as a great client to work and develop things with. And we managed to establish that relationship and we had a joint slack channel start experimenting a lot. And then the second thing that we did was important to me was to encourage people internally to really lean in and try it. And originally there was tons of these like concerns, what about data, what about this. So we made sure to very quickly solve these things so we would GDPR comply in and that we could set the right structures around it since we were a bank and all that and really make sure that everyone in the company would experiment with it.

4:56And then it just happened to be so that some people were more curious and more passionate and some people were less and that was fine. But some people leaned in and it happened to be so that one of the teams that leaned in started looking at a fairly actually complex challenge in a way, which was what we call dispute resolution. And dispute resolution in the bank is basically a customer calls us and says, hey, I didn't get that package. And the merchant says, but I did chip that package. And then we have to be like a small mini court that basically gathers all the evidence and the sides whether, you know, who has who's done wrong and who's done right.

5:28Right? And what are we going to do on this transaction? Who's going to cover the cost? And these errands are very complex. They require a lot of evidence, a lot of emails back and forth in communication with they're both a consumer and the merchant takes a lot of time. And there's always been a backlog and it's always frustrating because customers weigh a lot before they get the final outcome. And so we started experimenting just to see if Chatchity B and these services could help us make basically like a co -pilot help us take those decisions faster. And I think to some degree what's important here is that a lot of it comes back to the creativity of the team.

6:05We had a lot of teams in Klon, as somewhere more successful as a list. This team happened just to be very, very strong and really good at what they're doing and very creative. And they found and built away what is actually today referred to as Rags. They built already back then. They realized that would be a good solution. And within two months, they were demoing internally to us and others that they had managed to build a copilot that basically helped accelerate the process and also increase the quality of the decision making because it was making sure that we actually really took in all the relevant information and then took a decision on these disputes.

6:43And we said, look, this is amazing. Let's put it in production so far just as a co -pilot. And then the team, you know, which was crazy like two months later, we suddenly get this slack message internally, which says, we're out of errands. Can you send us more tickets? And that's never happened. Like this has been a constant backlog. It was just like, this is really impressive. And then we said, look, it's be crazy to try to see how we could increase the pace of this and actually even answer customer service errands. And that was, and then that team went on that challenge. I think to us, what was most critical was one of the rules that we agreed on was that the customer should always know if they speak to AI or if they speak to human.

7:31That's been important, but when we wanted to start testing this, we had a bug, and the bug was that for a few thousand conversations, it wasn't clear that it was AI. And when we then looked at the, we, which was, you know, not intended, but the conclusion was we could read a few thousand transcripts of conversations where the human was in a way, the customer wasn't aware that it was AI answering. And we realized that the AI was doing heck of a good job. And that made us conclude that it's the most important thing to us was that customer satisfaction would be equal or greater than what it was with a human agent.

8:14And as we saw that we start to reach that point, then it became, we became less nervous about putting this into real production and actually try it, try it. So I think that like, you know, it was really the effort of that amazing team that kind of tested and iterated, and that a few lucky you shots at the wrong way that then allowed us to put it in production. Amazing. There are so many questions that we want to ask to follow upon this. So they will start with you mentioned that rag is part of the architecture. Can you say a few more words about the implementation itself and sort of where does open AI end and Clarenne begin in terms of what you guys have built on top to make this work?

8:54Well, I think that the it's funny because in general, I'm always so freaking transparent, and this is one of the time in my life that I actually feel a bit cagey about telling too much about the secret sauce. Because I actually think about this as a fairly important strategic advantage. But what I can say is that in our case, one of the key elements was that like, it's about making sure that the instructions are clear. If you onboard a human and you ask a random human to sit in our customer service and try to answer a question. And the documentation that's available to that human is subpar because there's an assumption that you can rely on what people have learned in different sessions or assumptions.

9:48You're not going to be successful, but if you've written a documentation that is detailed enough, So you could, even if very slowly put any random individual and they could slowly go through your FAQ and manuals but actually answer a question correctly because it was documented at level detail, then it works. And that's how I think about the AI today. It's basically an employee that turns out at the work every day and has forgotten everything about what Klanai is, how it works and those are. And every time you need to tell it again. And that may change over time, but currently that's partially the game.

10:21And then so that helped us a lot to think about that way, that we just needed to make sure that the documentation and the manuals were clear enough and of quality enough, and then it can actually execute. Because many times, you know, if you, it's, I mean, the truth that has been the truth for data scientists for long period of time, shit in, shit out, right? Like if you feed data models with bad things, you're going to get bad results, right? So you need to make sure what you feed in is good, and then you can get better outcomes. Sebastian, I think you tweeted that your customer support agent is now handling two -thirds of your customer service inquiries.

10:54Go to question from your audience on Twitter this week. Are you planning to replace your CS department 100 % with AI? I guess from a technological perspective, do you think that's possible on what time span and what are your plans? Well, I mean, it's very hard obviously to predict how far will I go and what can we do in the future and so forth. But I think that it's definitely not going to happen anytime soon. And I do think that there will be customers that prefer a human for, you know, could be for any reason, could be because they have such a belief or a prediction or preference or whatever.

11:29And obviously you want to serve those customers as well. So there's no chance that, you know, the human agents are going away anytime soon. With that said though, I think actually the biggest quality improvement that we see is that

11:49generally speaking, and obviously we as every other companies to some degree want to avoid this, but it's not uncommon that our human agents have multiple chats going on. And we as customers all know that because you go and chat and you're like, you write a question and you don't get the answer immediately, and you're a little bit like, come on. And they forgot about you. And they're like, hello, John, where are you? Like, why are you not answering? And they're like, oh, I'm, you know, and so forth. And I would ask that because I, you know, when I have tried to work in customer service myself, I did the same thing.

12:17Like, you know, it's like, it doesn't make sense because we also sometimes take, customer is slow when not answering and so forth. So you start doing something else. You can just sit in idle and wait for that. You want to, you know, resolve more things. So you get it. But the consequence of that is that the average time of resolution of a customer service chat is about 14 minutes. And when we move to AI, it's two minutes. And the reason of that is because you get instant response as a customer, as opposed to that delay that happens due to that parallel handling of errands. And this is actually the biggest advantage.

12:54And so as a customer, a lot of customers that tried at say, wow, I want this experience. But at the same prototype, we have something else which is funny, which is that AI chatbots have been around for 10 years or something, and they all been of horrible quality. And so each one of us have gone to some airline and tried to converse about some tickets and been like, my god, this is the dumbest thing I ever talked to. And so the funny thing is that of those 30 % currently that do not use our AI chatbot, But the most common reason is that when we start the conversation with them, the first thing they write is agent, right, agent, which basically means they want to speak to a human.

13:34And that's not necessarily because they so deeply want to speak to a human as some of them are so are, but a lot of others is just because they had these horrible experiences and they want to avoid it. They just don't trust it to be good. And so actually what I'm seeing is that what's happening right now is time. Well, it will take some time to educate customers on the fact that like, you know what, but this experience is actually many times better. And a lot of the people that tried it, they want to use it more because they find it more, you know, faster. And I think that takes a little bit longer time, right?

14:05There's the actual experience, but then there's the perception expectation of what the experience is going to look like. And changing that takes a bit longer than changing the experience itself. So I suspect we're going to see even higher proportion of things dealt with AI, but there's obviously a lot of complex queries that it doesn't resolve well today. and that still needs to be proved on and there's still tons of work to be done. What are the trade -offs? Are there ways in which it is consistently worse than what you had before?

14:39I actually know, but it's not entirely, as I said, that is not entirely due to the fact that of AI. that is partly due to the fact that some of the instructions that were manuals that were written to help our human agents were subpar. And the experience already before suffered from that, but not enough managerial attention and focus was put to improving that and helping our agents become better at work. So actually our agents have better tools today to be successful in helping the customers as does the AI So the consequences both experiences are improving as the consequences of that But you know, so I think that like partially it's true that like Things would be worse like yeah, no, I think both sides get better by doing this actually Because just we realized the importance of these things and I think sometimes to some degree previously You just you know there wasn't enough focus on the topic like, yeah, that's great.

15:39And you mentioned the 14 minutes down to two minutes. Are there other statistics you can share that helped to kind of illustrate the impact of this? Well, I think the one that we were most famously, quoted on was obviously that 700 full time in place. But I think that one, and we were very, it's a difficult number to share. And I understand like, we understood that people would react to it. But at the same part of time, I also feel to some degree that like politicians are too slow on Considering and thinking proactively what this is going to how this is going to impact society and we felt that there's some level of importance of sharing such statistics to kind of a little bit say look This isn't just fund demos on Twitter.

16:23This is actually having real -life business and real -life implications Now with that said in our case we are have been using customer service on contracting firms, those firms employ hundreds of thousands of people. And if we historically have improved our products somehow that may also have led to less customer service errands because we fix some issue of flow in the product. But obviously I've never seen an improvement to our product that added push of a button had this dramatic impact on number of customer service agents that we need. And now fortunately for those agents, there are tons of other customers out there.

17:02So nobody has lost their job as of today as far as I know at least as a consequence of this. But obviously in the longer term it will have implications on these kinds of jobs. But that was the statistic that a lot of people obviously reacted to. And the fact that it's about $40 million of improved profitability for the company on an annual basis. Right? So it's fairly significant. I mean, we do about $2 billion a revenue. you so it gives you kind of a sense when it's size. Awesome. And since you've been thoughtful about kind of the broader societal or economic impacts of this technology, I'm curious if you had a magic wand and you could craft a policy or a procedure that would help get us through what's likely to be an era of disruption.

17:50Do you have any thoughts of what you would do with that magic wand? Yeah, for sure. all is easy and programming put in place. The first thing that I think is super critical is actually may sound a bit surprising but it's an electronic identification methodology for humans. Currently there's no globally applied such methodology, just like a passport but an electronic one. And why that is so critical is because the amount of fraud and scams you're going to see increase is due to the fact that I have no there's no ability for you patents on right now to ask me for my electronic identification to verify that I am not a a bot or I'm not some you know what's the called fake you know video created or pretending to be an AI you know not an AI talking to you it's actually the human right and I think being able to verify that you are talking to the real human is critical if we can supply that on a global level but preferably even on a country level that will at least reduce the risk of fraud and so forth because you will be able to authenticate.

18:54Am I talking to Pat? The real Pat? Where am I talking to your boss? I want to be able to know that. That is very critical. I think that needs to be resolved fairly quickly because otherwise we're just going to see an explosion as these. I have seen videos of myself talking to customers that we have produced that look identically to me. Sound like me. well, you know, which have been, we are about to send out to over 1 ,000 of our top merchants. And so like, it is crazy to see those avatars, you know, and be able to impersonate them. So I think that's one thing. The second thing though is if you're left leaning, I hear people on the left side of the political spectrum, they are saying like stop the progress.

19:33I have a hard time, especially considering that there are less democratic countries in the world that may, you know, push this agenda as well. And so I think that's not necessarily the best outcome. But if you're on the right wing of the political spectrum, a lot of people say, oh, don't worry, there's going to be new jobs. There's always new jobs. This happens all the time. There's always new jobs. And I think that's a little bit of simplification as well. When I was in Brussels, there are, I think if I remember correctly, 10 ,000 translators that are employed in Brussels to translate all the European legislation into the local languages of Europe.

20:07Those 10 ,000 translators are basically almost redundant today with the technology of deep hell and chat to the pee and so forth to some degree, right? I mean, you could at least reduce it dramatically. And I don't think it's easy to say to a 55 -year -old translator, don't worry, you're going to become a YouTube influencer. So I think that what you can do from a society political perspective is you can think about, But okay, maybe I don't want to stop progress, but maybe I can offer something to people being affected. Right? Like maybe I can offer something to them. Maybe society can have the luxury of at least support individuals that are affected by these changes, because not everyone will be able to just retrain into something different.

20:46And I think it's in that vicinity. And I hope if there would be such measures or at least plans or ideas among politicians, then maybe you can take society through this change with a little bit or more of empathy and care about the people who are affected, what at the same time not saying you have to stop progress, right? Yeah. Thank you for being so thoughtful about it. It's encouraging to see people in leadership positions like yours being so thoughtful. Thank you. Sebastian, what types of jobs do you think are going to be most affected? What type of jobs do you think will be? What skills are you teaching your kids to learn so that their future livelihoods are AI align and set to speak.

21:29Yeah, so I think that like it's funny you say that because when I met Sam back then and I got that meeting I said to him, look Sam, one thing that's going to happen is people will this is going to have impact on jobs. So I think if you want to make this a very popular technology, you should identify like what are the job categories that people hate the most. And you know, I happen to have two of the three ones because I'm both a CEO and I'm both a banker and those are two of the ones and then you have the only their lawyers, right? So those are the three ones. So I said to Sam, like, what you should focus on try to build AI that replaces CEOs, bankers, lawyers, and nobody will make a big fuss about it.

22:07Unfortunately, you know, and I saw that very clearly because when, you know, when we did a tweet later on about the marketing things we're doing about AI, where we have less need for photographers and such copious things, less needs. We still need them, but we need them predominantly for the very creative stuff and less for the kind of day -to -day stuff. That had a violent reaction online and I can understand why because people really feel emotionally resonate a lot to that. While when you see online tweets about AI lawyers, nobody seems to react much. I feel sad for the lawyers in the world.

22:45I hope you, you know, people remember you as well, Lashley. But anyways, you know, it's scary, right? It's scary because I don't know. I find it is a very difficult question to answer. I just, to some degree, definitely physical jobs. It just looks currently as on a very long term perspective, it's going to be easier to replace knowledge jobs, then it's going to be to replace driving a truck or even though we were so convinced about self -driving cars and all that. Or, you know, proper robots seems a little bit further out than, you know, AI. So, it's difficult. But that also assumes that everyone wants to work.

23:35I'm not sure. Like some people would like work. Some people would enjoy a society and we'd draw what's servos and we'd just, you know, hang your out and play football. So, like, it's a little bit, you know, it depends like it's hard to predict where all of this is going to go, right? I preferably love work, so I will be one of the depressed people when AI takes my job and I'm going to sit in like, and be like, okay, that was the end of the fun because I really enjoyed a lot, but like, you know, people are different, everyone's different, right? Let's talk about some more of the stuff that you guys have built internally.

24:04So you see mentioned a little bit what you guys have done in marketing. Can you say more about that? Yeah, I mean, it's been very interesting as well. I mean, because again, there's so many demos, right? And it's like, it's I mean, I think everyone's tried and you've gone, you know, tried to create an image, a creative video and You are blown away first with what you can do But then you're like, yeah, but I wanted to look like exactly like this and I wanted to be consistent with my brand feeling and I wanted to you know Etc. And then you start being more challenged with it, right? Like and I think that's why I also sometimes I feel a little bit like For example, you know people say like, oh It's unfair to the creators that these tools are being created in marketing.

24:45So for the night, I partially understand why people say that, but at the same time, we have this guy who has totally immersed himself in this video creation, sound creation, marketing, creating things, and created this amazing, basically just scripted together, a lot of these different technologies to create like these automatic marketing videos with me making presentation to merchants. As an example, me as an avatar. And it's really nice. It looks really great. It's on -brand and so forth. He is a creator. He is extremely creative. And it's a little bit like I can assume that when the music industry evolved and synthesizers came along and computers to make music, some people were like complaining, that's not playing a guitar, that's not creative.

25:37You're sitting by a freaking computer. Nobody would say that anymore, but I'm sure there was a lot of that criticism. To some degree, I feel that people that are adopting these technologies today, they are just, they're very creative, but they're using new tools to be able to create what they see in their minds. And so, you know, I think that's what I'm seeing. So we are basically having people who may not themselves be photographers, who may not themselves be great at Photoshop, who may not themselves be great at all these things, But now with text and communication with the computers they can create what's in their minds and they can explore ideas and concepts of marketing campaigns of marketing material of doing things in a way That was unprecedented before right like and that's really exciting and they obviously there's yeah, it's true.