Big Technology Podcast - Did Klarna Really Automate 700 Jobs With AI? — With Sebastian Siemiatkowski

Episode Date: July 17, 2024

Sebastian Siemiatkowski is the CEO and co-founder of Klarna. Siemiatkowski joins Big Technology Podcast to discuss the company's aggressive efforts to implement AI across its operation, including in customer service where it says the technology is now doing the work of 700 people. We also discuss how Klarna is using generative AI in marketing, and whether it can do the work of strategists or simply discreet tasks. Stay tuned for the second half where we discuss how Klarna is using ChatGPT enterprise, the company's roller coaster ride amid big market swings, and whether the company plans to continue to grow its workforce in the age of AI. --- Enjoying Big Technology Podcast? Please rate us five stars ⭐⭐⭐⭐⭐ in your podcast app of choice. For weekly updates on the show, sign up for the pod newsletter on LinkedIn: https://www.linkedin.com/newsletters/6901970121829801984/ Want a discount for Big Technology on Substack? Here’s 40% off for the first year: https://tinyurl.com/bigtechnology Questions? Feedback? Write to: [e-mail]

Let's talk with the CEO of Klarna, one of the most enthusiastic adopters of generative AI, about how the technology is being applied in practice and whether it can really do the work of 700 customer service rubs. That's coming up right after this. Welcome to Big Technology Podcast, a show for cool-headed, nuanced conversation of the tech world and beyond. We're joined today by Sebastian Shiemkowski, the CEO and co-founder of Klarna, which is a fintech company that specializes in Buy Now, Pay Later, but also has its own shopping, app. And it's also a company that's been leading the charge in implementing generative AI, one of OpenAI's very early partners. Sebastian, welcome to the show. Great to see you.

Thanks for having me. It's great to be here. So let's just start off with this stat that you guys put out a couple months ago about how you have effectively built AI that's capable of doing the work of 700 customer service reps. If listeners of this show know I have a stance on this, which is that when I see these numbers from company, I generally think that they have haven't actually replaced that many people with AI. Here's like one example, right, like IBM. They had an announcement in August 2023 saying that they're going to replace nearly 8,000 jobs with AI, but also as they're going through a layoff. And look, maybe this is a case where you've actually been able to, and I think your wording is pretty interesting and we'll get into it.

Maybe this is a case where you actually been able to hand off the work of 700 people to AI, but also like you guys did do a layoff of 700 people right beforehand. And so I always wonder, like, is this replacing people with AI or is it something a little bit different once you get back into the details? So you tell me, like, did you hand off the work of the 700 people directly to AI or is it something different that we should know about? Right. So the layoff of like the comparison to the 700 layoff is actually a misquote by a news magazine online. It is not accurate. It was two years ago when we had to, you know, change the amount of investments we were doing. I mean, we had to make layoffs, and it just happened to be that the numbers are fairly similar, but it's just coincidence. So that's a separate thing. But taking that aside to answer your question, once you think about it this way, right?

Like, I think it's almost better to give the like a little bit of the story and the context. When AI came along, we started a lot of initiatives internally to explore the capabilities of it. And we were quite free internally in the sense that we said, look, you know, it doesn't need to be like core business that you go after. So we build a lot of different things, somewhere that are more related to our business and some that we would expect other companies to build. build that aren't directly to say the core to what Klanah does. And one of the teams that happened to be very successful internally was a team that started creating basically a kind of co-pilot for customer service to resolve disputes. And disputes is one of the like more tricky thing in our customer service world because

you collect data from immersion from a consumer. You have to decide based on that data, you know, one claims that they didn't receive the package. The other one claims that they sent the package and you have to decide whether the customer should keep the money or not or whether you should persist in trying to have them pay for something right and so it's kind of a difficult it's almost like a mini court decision that you have to decide on and so this team managed to do this in a very very good way so they managed to create a co-pilot the co-pilot started helping these customer service agents collect this information from both merchants and consumers started applying a more methodical approach to

assessing these errands and then also providing decision support and like what you should we do on this specific occasion? So that was kind of the beginning. And the disputes at us had always been like, there was a backlog of like 30 days. It's a quite complex matter. It's always frustrating to us because consumers want to get answers very quickly and we need to collect a lot of information.

So it takes a little bit too long time, et cetera, et cetera. Once they took the co-pilot live first, that in itself basically meant that in a few weeks, our backlog was down to zero, which was unheard of. And we even had this fun in ton of Slack message where an engineer is like, we're out of errands, send us more errands. we want to run it on a co-pilot, right? So it was at least an indication. And then we tasked that team to say, hey, would you be, would you want to try to actually build a full

service, customer service kind of agent, right, based on what they have done? And then they worked on that for another six months. Now, to your point, when you then think, did we really replace 700 or not, what has happened is that, like, I think if you look at Klanah, you know, if you look at American companies, a lot of them will have. fairly advanced IVRs, right? So you would call in press one for this, press two for that. And then they make... What's an IV? So, like, you know, when you call a customer service support number and it will be like press one, press two, whatever, they will have these like fairly, you know, and we all hate them.

We're not too happy about them. They're always like a little bit annoying because you want to call to, you want to talk to the human. But the truth is these systems, they do resolve quite a lot of customer service errands in that they start presenting you with facts and you're going to hang up and not talk to the agent because you kind of got what you needed, but it's a little bit annoying because you want to talk to a human. So Clana was not very advanced in that. There are other companies that are more advanced in building out such kind of services, you know, some semi-automated services, which requires you to collect a little bit of information and present some information. And all companies have been doing this in a way to kind of reduce

the number of errands that agents actually deal with on a day-to-day basis, right? Like, it's a pretty standard procedure. So one has to take that into account. when you think about the potential savings that we had because we didn't have as much of advanced such systems, it was a little bit of a lower threshold for us to achieve this accomplishment, right? In addition to that, however, what was clear is that, like, when we started exposed this chat AI agent to customers

and they had the opportunity to interact with it as an alternative to a human agent, the customer satisfaction on that was equal to a human agent in many cases. And so that allowed us to say, well, we should then scale and propose to more customers to use this as an alternative. And to be fair, like all product development, all things that you've been doing to improve your app or whatever, I mean, partially you want customers to be able to self-serve and serve themselves. So any product improvement partially has an implication on, you know, reducing number of variants to customer service, right? If you have a really bad app, people will call you more often. If you have a good app, people will call you less often, right?

So there's always going to be that. Now, the difference was when we got this customer service AI agent to reach a level where it actually served a lot of errands and on a satisfaction level equal to what human agents many times did, and we took it live, the number of errands that are human agents needed to deal with that was removed in a single day was the equivalent of what 700 people used to do manually before, right? so that is actually true and that has led to a saving now in our case we don't hire these people ourselves we use customer service companies and so these agents would go on and do other jobs for other companies in the shorter term because these companies employ hundreds of thousands of people so when clana has less errands somebody else will have more errands and they will go and work on them instead but we still wanted to share this metric because we felt look directionally speaking if this continuous it will obviously have implications on the number of customer service jobs that exist

in the wider economy right yes so we still thought it was a very worthwhile statistic to share so did you then reduce so you outsource your customer service did you then reduce sort of the headcount that you get from these outsource vendors by 700 like what so 700 less people from these companies yes and like so we on average would have like a about 3,000 it depends on because you have Remember also just like Amazon, we have much more transactions around Christmas because we're very online. There will always be variations in these numbers, but like on the average, you would look at like 2 to 3,000 agents. And that was removed by about 700 when this AI chat agent went live, which meant that it's actually a reduction in cost for us as well for paying these customer service companies that we, you know, we basically are suppliers. We reduced the spending with about $40 million on an annual basis.

What did these vendors say to you when you were able to make this happen? Well, they were not very happy because we tweeted about it and a few of them had very severe implications on their market cap because like one of them lost like over a billion dollars in market cap on the stock exchange. Now that was definitely not our intent and we felt a bit embarrassed about it. That was not what we were trying to to accomplish by sharing this statistic. But no, but they, I think that, you know, they are, you know, some obviously were, you know, they were first they were asking like, where are all the. errands. That was the first thing because they were surprised, obviously, to see that dramatic shift in number of errands that we were shipping to them. But then there was also like, I mean, since we share the statistic, you know, some of them are more like, well, what else could we do

for you and how could we grow the relationship? Some of them are more keen to learn and understand how did we do this because they're trying to offer similar services to their customers and so forth. So there's been like a mix of reactions. Right. So give me an example of like what an AI customer service bot would be doing that a human customer service bot would have. have done previously? Well, it could be very specifically. It could be like more simple errands. Like, you know, hey, I want to find out, can I delay the payment on this transaction?

Because I don't want to pay now. I want to pay next week, right? Like, and previously customer service may have instructed that individual on like, where in the app do you go and do that? Or I may even help you prolong it and give you a new due date on this specific payment, right? In this case, the AI would basically show, like basically serve you. directly in the app the button to click and delay that, right? So that would be a good example of something that, you know,

nowadays would be handled by AI as opposed to the human agent. One thing that we also observed in this, which I think is worth adding in this, is that, and I think most of us have had these experiences, when you interact with human agents over chat, right, many times, and, you know, companies will always say that we try to avoid this, but it still happens and it's a fairly general applicable practice, is these agents will have five, six chat conversations going on at the same point of time.

And we always, as customers, we experience that because we write something and then we're like, why are not answering immediately? Like, give me an answer, right? But since I've also sat on the other side as a customer service agent, I also know that that's also normal because maybe you, Alex, are pinging me something

and then you get a phone call or something so you're not writing anything. So I'm not going to sit and wait for Alex. You know, I'm going to go and have a few other conversations simultaneously. It makes sense because otherwise it would be very, very inefficient. However, we experienced that. Now, the difference with AI is that we don't need to do it that way. So what you see is the, I think the biggest difference is that in general, when a human started a conversation with another human to resolve a single task took, on average, 14 minutes, right?

Just because of those delays that happened because people aren't really actively talking to each other all the time. And somebody said, oh, let me go and check. And something is a delay, et cetera. and here the resolution time went for 14 minutes to two minutes and that is because you get instantaneously response from the AI instead right and like the AI is kind of focused on your conversation so to speak and so that is I think one of the biggest differences and also something that drives up to customer satisfaction

because you feel it's more immediate right but it doesn't mean obviously that all errands are you know can be answered by the AI today there's obviously still a lot of things that humans deal with because they're more complex and more difficult. And there's also a huge amount of customers. The first thing they write when we exposed to AI agent to them is agent, right? That's the first thing. Because a lot of people have had so much bad experiences with these AI bots that they just want to talk to human, right? So like that's also another thing that we see a lot of them. Yeah, because I gave it a shot. I was in your app yesterday and I wrote just, I want to refund. I want a refund

for my order. And the bot writes back, I understand you're looking to get a refund to assist you better. This is what you have to do. One, exit this chat. Two, go to the customer service section under settings. Three, it's like the purchase you need help with. And once you've done that, you can proceed with the steps you need for your refund. So effectively, it's not like the AI is necessarily going out and accomplishing it for me, but it's directing me to the place that I need to go to finish this action. And that's a very common thing. So what you'll see is that what, you know, one of the things where we were actually a little bit lucky when we compare, because we talked a lot of companies that are trying to do similar things, right? And so Klan was a little bit lucky in the sense that

like before, already like a few years ago, we had this vision for our customer service service. It had nothing to do with AI at that one of the time. It was just that like if you, for example, ask like you did in that chat for a specific action, rather than just inform you of where to do it in the app, we would actually serve like a small widget that allowed you to do it directly in the chat thread, right? So as I said, like, you know, we don't have that, we don't have those widgets for all actions. So the actions that you happen to ask for, we don't have that widget for, but there are other actions that we had such widgets for. And so what we see now with a lot of other companies trying to catch up with this idea is that they don't have widgets for anything. Like, it's always been,

you know, customer service interacting with Alex and then going and doing that in a separate GUI or, you know, in separate software, right? In our case, we did had already a few, quite a few of such widgets that we could serve into that thread that would allow you to do it. And that is partially what has allowed us to you know to get this going at a bigger area but to a point you will always be able to find things that like you know are yet not working or not yet at that level and so forth right so and that's why I said I also made a comparison with IVR if you had a very advanced such you know press 1 press 2 system then you know the difference between what we did and the outcomes would have been not as great as 700 it would still have been something but it would

not been at the same level yeah I definitely hate those systems so yeah that's a subject for another conversation let me ask you this there have been some problems with these customer service chat bots that they've hallucinated sometimes i mean i think like one of the examples that i've been given was that um an auto sales bot was like someone basically you know uh conversed with it and then convinced it to give them a car for like half price and then the dealer kind of had to owner had to honor that decision have you have you have you had any of these issues with your bots hallucinating or like pulling the wrong data

and if so, how have you navigated that? Not the wrong data, right? So you have to be very, I mean, they obviously have to put very strict standards into what is accessible, not accessible to the A application itself. So there you can control that. But it has definitely hallucinated

and it has answered incorrectly. But to us, the way we think about that is that like you have to make a compare, you have to also recognize the fact that humans don't necessarily hallucinate hopefully, but they also do errors and they will also answer incorrectly. So what we simply do is we read a lot of these transcripts on a continuous basis

and we do continuous quality checks to ensure that the error rate is not higher for the AI chatbot than it is for our human agents. And if we see that they are at least on par, then we think that's an acceptable outcome. But it would be, you know, we could never like promise that it never makes errors just like you can't promise that your human agents won't make errors because our human agents unfortunately also make errors right like so it's just about making sure that there are not you know a substantial bigger amount of errors that the AI is doing than the human agents are doing it's kind of like the logical self-driving car threshold which we'll never see but if like you can kill less people than human drivers you should probably roll out the self-driving cars but we don't

see it but but the difficult thing for self-driving cars obviously is that human lives are stakes. So the acceptance within society for those mistakes will probably be at a very different level. In our case, it's a little bit more fine if it makes a mistake, right? It's still financial services. It's not like, you know, some music app or whatever. So, like, you still have a different level of compliance. And we're a bank. We're fully regulated. So we have a lot of things that we need to live up to. But obviously, at least it's not life's at stake, right? Like, in that sense. So I was speaking with somebody about this move and they basically said, listen, like if Karner was able to figure out a way to automate customer service this way, I think you've had

the less public number is 2.3 million conversations and two-thirds of your customer service chats happen with the bot. Basically, they're like, if this was working as well as Clark claims that they wouldn't be talking about it because they have this like edge over the competition and why give that up in public. So I put that question to you. Why talk about it? For two reasons. One is that like, you know, I've been, I've been part of a controversial industry for quite a while called buy now pay later and people have criticized buy now pay later and i also see i see why but i also see strengths to what buy now pay later offers to the market compared to credit cards and the traditional banks and so my learning from going through that of being first like a hailed amazing

tech company doing awesome things and then starting being criticized for some of the also risks and you know challenges associated with i mean if you provide credit it's still credit right like you can provide better credit or worse credit, but it's still credit, then people will have opinions about credit. And my learning from that is that it's helpful to be proactive, that it actually is better to kind of lean forward and share things on a proactive basis rather than kind of have them, you know, appear later, right? So that's one reason. My learning from that made me believe that I think is better of us to kind of be proactive and share these statistics and so forth. I also feel partially like, you know, that I

I sometimes feel that the, you know, when I see so much noise and discussions about AI, this and that or whatever, I actually feel that we have a more kind of a moral responsibility to share that we are actually seeing real results and that it's actually having implications on society today and hope to encourage people specifically politicians in society to actually treating this as a serious change that's coming and start thinking proactively about what to do that. So that's more like on a human level. And then finally, the third thing, obviously, is also self-promotion, for sure. Like, you know, that's definitely also a play, right? So obviously what we've seen as a consequence of sharing this is that more of the AI startups wants to work with Klauna because we're regarded as, you know, a thought leader or somebody doing something exciting in the space. We see more people want to work here, you know, et cetera.

So there's like obviously also that aspect, right? But it's a combination of those three. So throughout this conversation, you've hedged a couple times about how powerful this stuff is because saying that if you had like a better phone tree for instance it might not have helped you save 700 you know employees or not employees 700 workers time so and you're not you said there'd be some impact but you're not 100% sure what the impact would be that being said is this stuff actually all that powerful or to just kind of help you paper over like a different problem like how should we because I'm thinking like all right so we should think about the implications

for society but like is this the moment or is this like a little bit less powerful than it seems given the problems that existed beforehand so i think um i want to make sure i understand your question correctly but i look i think it's a very good question in the sense that you know it's always like how how how much is hype and how much is real is that the question basically right well yeah i mean to be more direct you've said that you'd see less of an impact if you had your phone tree built out a little bit more before you turn this over to AI. And I'm like, okay. So like how much of this is actually just like papering over like a pre-existing problem versus. That is a good question. Look, I've, I'm sorry. I wish I could give

you like a great answer to that question, but I haven't worked in another company and I can't really make that comparison. I've tried to talk to under entrepreneurs and I don't want to mention them by name because I'm not sure whether they want to share the details. But I have talked to others who have for example come much further in automation using non-AI so to speak before like IVRs and systems like and also who has been tougher negotiators with their customer service suppliers and hence their cost per errand was lower because they had better prices than we had from a scale perspective and their savings by moving to AI was more limited than ours right so there is definitely an element to that but it's very hard it's very hard to answer that right because it's so

company by company specific my belief though if you ask me is that like I don't know like maybe 70% is AI and 30% is automation or 50 50 yeah but I still think it is that much actually that's still my belief right and then the other thing is like even when I talk to some companies that were really good at like had really low customer service costs really good quality and really high level of automation even before AI I still feel that like I I am super happy that we did this because this is just the first version of this, right? Like, we've learned so much from the implementation, and it's not like we just launched it and now we're letting it be, right?

We're continuously improving on it and making it better and better. So I think that, like, even though maybe the first iteration would only have been a substantial improvement compared to some other companies, if you give it one or two more years, it will definitely be an improvement versus what any other companies doing with general just computers, automation and so forth, right? So I think it's a little, like, that's a more, I mean, maybe it's some boring. It's not the headline-catching answer, but I think it's the more real answer. We want to unpack the nuance of this stuff.

Yeah, yeah, yeah, yeah. But I think it, that's, I think that's the reality of things. That's why I just want to lean in. I just want to learn using the technology as much as possible. But, you know, it's still going to take some time for it to fully mature, right? And this extends beyond your customer service department. It's also in your marketing department. Let's see.

You've tweeted our in-house marketing. team is half the size it was last year, but it is producing more. So, I mean, you said, I think AI, AI is used for 80% of all copywriting within the company. I struggle to believe that could be that good, but I'm curious if you could tell us a little bit more about how AI is working within your marketing division. And is it taking on responsibilities of full workers like it is within customer service, or is it more enhancing the impact that the marketers are having? So Clan is active in over 20 countries, which means that we cover over 20 languages. And that in itself, like you can imagine, like, I don't think everyone on the call as well would like agree.

Like translations have basically been nailed, right? Like if you look at like not only a chat GP and open AI, but also if you look at companies like DeepL, the translation quality is extremely high. So if you think about our copy, you have to remember that a lot of those copy people were actually, you know, also performing copy in different. languages and so forth. And so that in itself is just a massive reduction in number of work, labor hours, just to translate from one to another and so forth. I think in addition to that, I want to ask you as a CEO, are you comfortable allowing an AI translation in marketing copy to go in front of an audience without having a human take a look at it? It depends on what copy it is and where it sits, right? So if you think about like, it's actually quite interesting,

because we're a bank, for example, if we market our credit card, there are very strict requirements on the copy and, you know, how that is expressed, right? But there's also other things in our app. Like we have, for example, descriptions of products that you can buy from our merchants or you have things that are of lower level of sensitivity, right? So the thing is that you have to basically structure your information internally in a way where you start separating and understanding that some things will have higher levels of requirements and some things will have lower and and the point with 80% is obviously that's focusing on the areas where it's less you know critical if the translation isn't perfect but there's a lot of human reviews still

of the translations themselves but it's a little bit like you can have AI write the code but you're still going to review the code many times from a human previously you just had to have both the human to write the code and to review it right so I think that but so that again is a in work effort, but it is still not like a full removal. Then what we've seen is also that AI becomes very powerful when you split it in multiple assistants doing different jobs. So if you, for example, have one AI bought right, the copy and then another one review it according to rules, then you actually make that also even better higher quality and so

for it. So you see a lot of the latest AI that we see at least apply within Klanai is that you actually ask the AI to pretend to be. different roles. One is the reviewer and another one is the writer and then they kind of interact with each other as like a multi-agent team. And that actually gives even better quality outcome. So, but to your point, obviously if it's like credit card promotions, you're still going to have humans review that and look at that. Right. And it doesn't seem like AI can do some of like the core functions of marketing, right? Speak with a group that has something to market. Understand their

objectives. Bring it to the creative agency with a brief. Go back and forth. Find a good midpoint. and then you think so well look i i just there's been some really cool things i've done so like i'll give you one example right we um one of the first a application we actually built internally was and this was again just like an idea that we just did it it's not clana's core business but one of the it's just a concept that we wanted to test it was that when you do these kind classical employee engagement service right like which all companies do like how happy are you working at Clana or how happy you're working at META or, you know, whatever. And you say, like, want on a five. And like, how happy are you with the office? How happy are you with your salary?

How are you, whatever? How are you a team? Do you trust your colleagues, etc. So a lot of companies do these surveys that you collect data, you know, people are supposed to say by on a scale one to five, this and that, you know, whatever. And then you kind of synthesize that information. You try to, you know, analyze it, interpret it, put some kind of report, you know, et cetera, and spread that and so forth. Right. It's a very typical company. doing these things. So we said to ourselves like, wow, you know what? Wouldn't it be happy to, like, wouldn't it be fun to like, it's still like so much of, at least to me, when I look at such employee engagement service, what I really care the most about is the comments where people

have written free text because it's so much more interesting. Like, you know, what does it mean to say one to five, this and that? It could mean so many different things for different people and stuff like that. When I read the comments, at least a little bit more concrete and I get some like input on like, what do people think or whatever, right? So what we decided to do is we created, we built our own internal interviewer. So we let the AI interview our employees in a session, like ask these questions and synthesize and have a conversation with employees. It's not to replace human managers because in the end, your manager is the one that's responsible for you being happy at work and all the stuff. But it's an additional tool on top of that to provide some

additional insights and understanding of what's going on in the organization. And the funny thing is that it worked extremely well. Our employees liked it. It was very thoughtful. And we felt compared to a standard survey, it gave us much more interesting outcomes and understanding of what was going on in the organization. That, again, doesn't replace the management that is still the primary objective, but like it still is much of much higher quality.

So to your point, can you use the same for customers? Can you survey your customers and collect and synthesize information? Yes, we've seen that work as well. It actually becomes much more interesting. so again like it doesn't replace all of these things but what I feel with AI with people forget is if you go to AI and you ask a general question right like whatever give me a great way to work with marketing you get a very general answer it's like reading bullshit corporate management literature like it's a you know it's a bit like good to great hire the best thank you but how the hell do

I find them what interview questions to ask like I get it it's important to hire the best people what does that mean in practice right the interesting thing is though if you take AI and you ask something much more concrete, which is like, imagine the front of a house, what would I expect to see at the front of a house? It will answer very correctly, we'll say, probably a few windows and a door, right? And so a lot of what people, I think, miss is that if you take any work task, whatever work task we as humans do, and you derive it down in very, very small concrete steps that are very specific, then actually AI is very, quite effective at performing those and most of our more complex works tasks are really consisting

of very small specific work tasks that are combined together and when you start thinking about it that way you can actually make it do a lot of things like but you have to think about it in that kind step and so yeah so I actually think it could do more than you than one may think right but these are all tasks and it's yet been able to effectively orchestrate and maybe that's coming but it's not been able to orchestrate like you can't tell me that you would sort of trust like the core of what your marketing department does to a machine yet because it's involved i mean even if they have great you know great surveys that they can send to like different divisions about like what they need to do like the the the ultimate end where you're positioning and figuring out what the message is

and what the benefit is to your target like that is something that human still need to do the synthesizing of information i agree with you at some level you know human still obviously outperform AI a lot. The other thing that I believe humans really outperform AI on is creativity, in my opinion. And it's a little bit, in a way, a little bit different with, it depends if you look at copy or text or if you look at images, because images are more fun in a way. Like, you have seen, you know, AI dream up some pretty cool images, right? That, like, you would like, who? Fun to see an artist that would have dreamt up that crazy image, right, it was pretty cool.

So I think images a bit different, but if you look at copy, one thing that I really see the challenge in how the LLMs work is they work towards the average because that's how they're constructed. So that means that they're constantly, all the answers are pushing through the general answer. And like, I think everyone will know that like if you invest in the stock market on the average, you're going to see average returns. Right.

But if you invest with a very strict contrarian owner. idea you may lose a lot but you may also win a lot and i think that like the same applies to your point to copy is that like if i want some amazing copy that you know articulates what clana does that's different i would not rely on chatyp to do that i would rely on a human because it a human has a bigger is much better at kind of being far-fetched and do something crazy and out there and different and so forth why the lLM is just pushing it to the average all the time and the average doesn't sell right so you've also been out front talking about how like you're using general alternative imagery versus images from, let's say, you know, I don't know, stock image companies.

But it's also like I'm looking at your app right now. I'm holding this up for listeners. It says like shop at Amazon with like a bunch of boxes and now it's about to scroll to top pet deals with like actual specific information about the deals you have with Chewy and then talking about the right season with Yeti. this isn't stuff that machines are able to handle yet maybe there's an image or so tell me the truth tell me the truth yeah yeah yeah i don't know if the images that you showed where i i can't tell you that sorry but i can't they look like general stock stock images that you would get versus like the AI gen okay but i'm sorry go ahead but i can't

so i can't tell you for those specific ones that you show in the show but what i've seen internally us do right is that you currently the problem is if you just go to like one of those tools you know any of those image generating tools and you just write like give me a to your point like give me a picture of a guy with box images or whatever they're not going to be on brand tonality they're not going to look the way clana wants them to look they will be a little bit odd there will be tons of issues and to your point you won't be able to apply them but what we have seen that we can do is we can basically again put this through a mechanism of a number of things so we can first text prompt an image we can then

take the same image as an input and move the image so it looks more on brand we can then make assessment and we basically then move them like through a factory basically of that and the outcome picture you have at the end of that are actually usable directly in the app and are being applied now again would i do that for my you know super bowl campaign no uh would i do that for category pictures in the app that are just there to say, hey, we have shoes, we have this. Yes, that we're already doing and applying. So you can see that you're, that is definitely feasible, but you need to set up this production, you know, environment where you basically are almost like a workflow, you take it through multiple models and multiple things to get to the quality levels that you're

looking for. And then finally, on the product experience, so people are going to be conversing with Klarna, I guess, through your customer service agent, but you have a vision to make that a much deeper experience where you're like looking through different sites and trying to shop and there's effectively a shopping assistant right there with you yeah i think look it's actually not only shopping in the sense like actually this started when you know as as you may know i've been uh you know running clon for 20 years next year's we've been long period of time and we have pivoted kind of direction of the company but in 15 so about 10 years ago um we were pivoting away from being a b to b to more of a b to c company and At that point of time, we asked ourselves, like, okay, where is financial industries going?

And partially this was because already in 11, we acquired a company in Germany called Support. And they had built, not AI-based, but they had built basically an application that did a very simple thing. In Germany, to transfer money between bank accounts was a horrible experience. The banking apps were horrible, right? It was just like horrible. A lot of clicks, a lot of data. It was just like a bad experience. pretty bad. But sorry, go ahead. It's still pretty bad. So what they had done, they had said, look, they had basically build a plaid-like experience where they said, Alex, give me your banking credentials, your password and your login. And there was basically a macro. It was basically like the old Apple macros that we used that on the Apple computers back in the days. It was basically a macro. They logged in on your banking account, on your behalf, and they scripted to send that payment on your behalf. So you didn't have to go through the ugly gooey of the bank.

And when I saw that to me, it was very inspirational because I felt, I was like, wow, you know what, like, amazing. Can you imagine if you had, like, this digital assistant that kind of did this thing on your behalf? So you didn't have to interact with these horrible goodies and experiences of other companies and it would just do this on my behalf. Wouldn't that be great? And you have to remember, like, that's not a small business. We process, we do about $100 billion worth of volume. We do $30 billion volume on debit through that solution, through basically scripting. This doesn't work that way today. It's more API. base, but the point is like for many years, we basically did millions of payments transactions through scripting on banks' guis, right? And so it worked. You can put it to production. Now, at that point of time, it's not AI. So if one of the banks changed the GUI, there would be a phone call down in Gieson in Frankfurt where the team sits was like, hey, it's broken. You need to fix the script. And somebody would wake up in the night and fix it because the bank had changed the GUI, right? It's like robotic process automation. Yeah, exactly, right? And you will see UIPath and some other companies have come, you know, to some degree on that, on that. So, some of that has already

been done, but it inspired us to believe that, like, okay, what's the future of financial services? Well, the future financial services, you wake up in the morning, your computer tells you, hey, I analyze your mortgage. I realize I could save you $10 by switching from Bank A to Bank A to Bank B. The only thing you need to do is to say yes to execute on that change. And so, like, that to us became the direction of where financial industry is going. It also means a reduction in excess profits because a lot of the banking profits are generated due to the fact that the switching costs are so high and we're not willing to switch. And so the competitive pressure is actually lower than it's perceived to be. And so that became a conviction of ours already in 15. And ever since we've been trying

to build services in that direction. Now, you know, we've come some part of it. We haven't, you know, nailed it but that's the direction that we had at now when we then saw chaty p we felt like oh my god this is probably going to have a little bit faster than we had previously envisioned but it's a bit like self-driving cars i personally believe the self-driving cars at some point of time will happen but i don't know when my bet is my daughter is now 10 and i have always said i don't think she's going to get a driver's license but that's eight years out right so there's still like there's still some time and maybe i'll be wrong but that's been my bet on it and i think it's like a little bit of here right like so um and that's that's what we're trying to do whether it's a shopping assistant or a

digital financial assistant but it's somebody who's helping you save time save money and make you in more control of your finances okay so your company is also one of the most enthusiastic adopters of chatypti enterprise um i want to talk about that on the other side of this break about how so many of your employees are in these tools every day and whether that's helping and whether that's sustainable. All right. We'll be back right after this. Hey, everyone. Let me tell you about The Hustle Daily Show, a podcast filled with business, tech news, and original stories to keep you in the loop on what's trending. More than two million professionals read The Hustle's daily email for its irreverent and informative

takes on business and tech news. Now, they have a daily podcast called The Hustle Daily show, where their team of writers break down the biggest business headlines in 15 minutes or less and explain why you should care about them. So, search for the Hustle daily show and your favorite podcast app like the one you're using right now and we're back here on big technology podcast with sebastian shimian kowski he's the CEO of clorna and the co-founder of the company we've been talking about AI in the workplace and sebastian your company is also using chaty pt enterprise with i'm maybe the most enthusiasm of any company in the world so these are just some stats you put out 90% of your employees are using generative AI tools powered by open ai daily

communications, marketing, and legal are using adoption, are using it with adoption rates of 93, 88, and 86 percent, respectively. You're also seeing a wide variety of additional use cases from building software to streamlining customer services, service, which we've talked about. Okay, first question for you, you're 90% daily active use on chat GPT Enterprise. Has that gone down since you released those numbers, or does it remain consistent? I don't know. I haven't checked, but I would believe that it's consistent. What are the people using it for? You know, it's tons of things on day-to-day stuff like, you know, help me draft this, help me review this, have me check at that text. We have also built a lot of what we internally refer to as CCTV. So this could basically be like different assessment tools. Like if you want to check, you know, is this text good enough for this purpose? As an example, actually, if you want to text, you want to test some copy to see if it's a, you know,

Is it correct from a legal point of perspective? Is it following the policies and routines that we have? Stuff like that. So there's tons of different things that people are using it for. With that said, though, you know, when Bitcoin came along and these technologies, like, I try that as well. I personally didn't find the technology to solve a real problem, in my opinion. I didn't see how it was going to help my mom prefer using Clana over something else.

Like that may be wrong, you know, maybe it turns out that, you know, blockchain is the amazing technology that will disrupt the whole world. But I wasn't convinced. When I tried this chat to be for the first time, I got very personally convinced that like this is a technology that will have real time applications. And since then, we've been encouraging everyone in company to learn and apply it and learn how to use it in a productive way. With that said, though, like, you know, I have days when I don't use it at all, right? Like, I have days when I use it much more. And sometimes I have more successful use case and sometimes I have less. But what I want to do is encourage the company to learn because a lot of things have actually

are day-to-day use case. They're actually very helpful. But it's also like by applying in learning that we explore and get to know what are the limitations. Why is it not working? What do we need to improve? And to us, we realize that it can either. be not working because the technology is flawed and not good enough, or it can be not working

because we are not set up to use it in a way that makes it productive. And many times we realize it's the latter. So we're changing a lot of our fundamental processes of how to work and so forth to make sure that they also helps us and can make us more productive. And some of these use cases are quite interesting. So this is from another article about what you're doing. You said you use generative AI, namely open AI's chat chip. Or the communications team uses it to evaluate whether press articles written about the company are positive or negative. I see it. Do you really need an AI to tell you that?

Can't you just kind of tell it by reading it? Well, you could, but I'm actually really proud of our communications team. They're so efficient. You have to remember, like, Clona, as a company today, you would probably read, you know, you would have about 40 articles written every day. And it's not only about the sentiment analysis of those. It's actually even a little bit more detail than that. It's very common. I would say about, I think we've analyzed it.

About 20% of the articles that are written about us contain factual errors. They're incorrect. And this has meant that our communication teams calls and tries to correct those errors, right? It could be anything from where the headquarters of the company is to statistics that are wrong or whatever. And so, and we call or email the newspapers to try to correct this to make sure that it's correct. right that's a lot of manual work right that team would much rather be out there and like you know pitching a new story or building relationship with journalists or doing something else than doing that manual work and so like having it assess those articles and identify those errors and then even sometimes draw off an email to the journalist to ask them to correct it it's like a nice it's a nice thing to avoid that manual work and spend time on something else wait the AI can actually pick out the errors in stories or is it just the sentiment do you just check the negative stories very relevant point Alex. That's actually exactly a very interesting thing because one of the key learnings we had the last 12 months is that there is an old rule in data scientists, which is

shit in, shit out. And that still applies in an AI world as well. If you feed the AI with a lot of noise and incorrect information, you cannot expect it to be able to answer such things. So in order for it to assess whether the information in the articles is, accurate. You also need to have a very strict data set that says what is accurate information about the company and you need to have that in a good solid place. And one of

the things that we've seen is that Klauna has been historically using a ton of enterprise systems, different ERP systems, which actually silos and puts more noise to the internal information. So we have our org charts in Workday. We have our clients

in Salesforce. We have you know, our suppliers in M-Files, we have tons of different pieces of information spread out on preparatory data systems with different structures and so forth. And that's hurting the ability for us to use that information in a standardized way to make AI work better. And so part of what we've been doing in the parallel with this is we realize that we need to standardize and centralize our information about the company of what we're doing, because then both humans actually can make more productive work as well as as AI on top of that both AI and like if if there's too many silos of information in the company if things are not transparent not open it makes it

harder and that's true both for AI as well as that right I think that like that's why I wasn't too convinced when people were like oh look I have a PDF reader and AI can read a PDF and answer questions about like yeah but you know the problem is in many companies you have too much information and too much inconsistent information and too much duplicative information so you have to think about both how do you improve the information that you have as well as like how do you then use that to do things like this which is you know check whether the data is correct or not in the article yeah there was a wall street journal article talk about how companies were like really struggling to get this to work when they try to get AI to pull data it would pull data from

like the wrong year and I think the key takeaway from that is that really that it just has to be the data has to be clean for these things to work and structure to us to a some extent you're also your lawyers are also using uh chaty pt to write first drafts of contracts and that's cutting the hours it's taking to draft a contract i mean that to me seems like a low-hanging fruit type of area where like you you have your lawyers like effectively draft the first draft and then they can do like some of the stuff that you're like actually paying for them for versus having them write up boilerplate contracts well i think that's true again coming back to like low-hanging fruit and you know more difficult things

there will definitely be tasks that people could have like you know instead of drafting this maybe you should have just shared a draft internally so you didn't rewrite it every time right did you really need to you know start from scratch every time and now just because AI is there people are applying it and like oh I'm saving that time but you might as well just share the drafts like obviously there are use cases like that obviously there are things like that as well happening where it's just like this could also have been done but just a little bit more standardization and like simplification or sharing information internally um so So it's a combination, obviously. But it helps provoke the idea, right? It helps accelerate. And as long as I see the business implications and the business results, it doesn't matter me too much, you know, how those are accomplished in that sense. Right. Okay.

I want to end with a question about the state of buy now, pay later. Like we've talked a lot about AI, but I would be remiss if I had the CEO of Klarna on the show and didn't ask a little bit about buy now, pay later. So buy now pay later was obviously like a darling of the fintech industry and of tech. Apple tried it. Your valuation was in the $45.6 billion in 2021, but then went down 85% to $6.7 billion a year later. And this was the CNBC headline. Klarna valuation plunges 85% to $6.7 billion as buy now, pay later, hype fades.

So I'm just curious to hear two things. First of all, from you, how has it been navigating this sort of like up and down of the industry and then what do you and then secondly what is the future by now pay later given that yes apples out and it seems like this thing that used to be in the spotlight is now moving out of it so what should we think about when we think about this service yeah so i mean first on navigating up and downs like i you know i think to some degree i benefit from the fact every newness for 20 years and as much as like this up and down was maybe the most media publicized and I've never been in that like spot eye, you know, or like as visible as this.

I've gone through a lot of up and downs with the company, both valuation wise as well as anything else, right? So I think that like it was obviously very tough and I was sad and I was, you know, very, you know, stressed by that. The public traded companies that we are often compared to like a PayPal or, you know, a square block or whatever, they had the same 85% drop in. the stock market during the same time, right? So we weren't singled out in that sense,

but that was a general fintech and tech kind of reduction in stock price. But still, because we're a private company, you know, it became such a bigger, you know, news. And then obviously also at the same point of time, as investor sentiment changed and we were at that point of time, unprofitable, we had to make, you know, very tough decisions that are very, you know, you don't like making, which was a reduction in staff and stuff and stuff like that. which is very challenging to go through.

But at the same point of time, I feel like, you know, we have to do what we think is right for the company and for the employees that are still here and our shareholders and our customers and so forth. So I think we did the right thing. And now we are a profitable company again, which we actually, you know, people don't know this, but Clona was profitable from 2005 to 2018.

So we had a history of kind of running this a little bit differently. The most tech companies are just like burn money. We have been profitable. But then when we came to the U.S., we invested heavily, and that meant that we were lost making for some years. Now, on the other topic, on Bina P. Later, to answer that, you have to first define what is Bina Pay Later, because at the core, what it is to me is we have had a credit card industry, which basically works in a way where you swipe your card, you get your monthly statement with all of your

transactions, and then you're encouraged to revolve. And if you do so, you start building depth, and that debt earns the bank a lot of money, right? Buy now pay later, the difference with it, the way I define it, is that it's interest-free and it's installment-based. So you take a single transaction, you don't pay interest on it, and you pay it down in installments. You can offer that on a card as well, right? Like, you can rethink your credit card to make it work that way.

You don't have to offer revolving. You can do installments on a credit card, right? So, but the concept of interest-free in installments is to me a healthier credit concept than revolving. In addition to that, the difference with Banapelator also in Klanas cases, we're our own network, right? So we are the equivalent of American Express in the sense that we're a third-party network. We have Stripe and Adion and other PSPs and acquirers that offer Klanas,

the payment method side-by-side with Visa MasterCard, but we have a direct consumer relationship just like Amex. So the network means that the fee that the merchant pays is not through visas, they do on visa, but directly to us. And this means that there are less middlemen in between the merchant fee and the actual income to us as an issuer of giving you as a customer this offer allows us to offer an interest-free product, which merchants still are paying for on equal parts with a credit card, which actually also charges on the consumer side.

So if you look at the total cost of payments in the US, and you look at both what the bank is earning by interest on you revolving, plus the merchant fees, it's actually a crazy $5.5 on $100 spend. And we earn less. We do about $2.5 dollars per $100 spend. So we do much less, but we also have much less cost, right? We're coming from a different cost profile. We're a lean fintech. We're not a big bank, et cetera. So you're accepting a lower revenue per customer with the benefit of providing a slightly more, a better product for consumers from a credit perspective. Now, it's still credit. I always say that. I'm fighting I'm fighting fire with fire. So I'm not saying like credit still has this issue and so forth,

but in a future state, if people want to use an installment product with zero interest over a credit card, I think that's a better outcome. You should use debit and then you occasionally use a buy now, pay later, as opposed to using a credit card for all of your spending. I think that's a better thing. And I always make the comparison back in the days, your card used to have press one for debit and press two for credit. But the banks removed that because you wouldn't build as big of a balance on a monthly basis if some of your transactions went on debit. So by having all of your transactions on credit, you were more likely to revolve and then the banks make more money. The average outstanding credit card balance is $5,000. The average outstanding balance on Clon

is $150. So it's a huge difference, right? So I think it's like it's a, but there was a McKinsey study that said that in the US, there are about 20% of the consumers are what they called self-aware avoiders. These are customers who are like really looking for this kind of product as opposed to a credit card with massive bonus points or, you know, other services that are more interesting for them or heavy revolvers and stuff like that. So this is not the full population necessarily, but I think there's a good amount of people that prefer this kind of products and see value to it versus other options. Yeah. Okay. This is really the last one and it's quick. As you start to put more AI into your business, do you expect to increase headcount or

decrease headcount from here? Well, you know, you didn't ask about that, but actually another announcement that we've made is that we haven't been hiring since September. And since we, as many tech companies, have a natural attrition rate where about, you know, 20% leave on an annual basis. So people stay about five years, which is kind of typical for tech companies. This means that we are shrinking. So we used to be about 4,500.

We're now 3,700. And we are basically shrinking on average about 70%. people per month and and we're doing that because we don't want to do layoffs and so we are actively managing down because we see we can do more with less people at the same point of time in order to create you know some benefit for for employees in this we have said that like and be very clear about it which is that our total employee cost will go down but our cost per employee will go up right so that is basically a commitment to our employees that, like, they will benefit from this in seeing higher salaries and more equity shared with them, which is what we're done. And we just did

us a few weeks ago. Now we distributed a lot of equity to our employees. So we're saying that like for the people that want to stay and participate in this, there's an upside to them as well in the fact that we're doing this. But we are shrinking and we will continue to shrink as a company. We'll be much less, but doing much more. Fascinating stuff. Sebastian, thanks so much for coming on. Great speaking with you. Thank you, Alex. Thanks for having me. All right, everybody. Thanks so much for listening, and we'll see you next time on Big Technology Podcast.