Klarna's AI Took 700 Customer-Service Jobs. The Secret Was the Manual

Pat Grady and Sonya Huang introduce Sebastian Siemiatkowski as the poster child for putting AI into a real business. In February, Klarna had said its OpenAI-powered assistant handled two-thirds of customer-service chats in the first month, with better satisfaction scores. Sequoia had been in business with him since 2010, when he was a young man in Stockholm. By 2024 the company was a global payments and commerce player, and every founder the hosts met wanted the Klarna customer-support thing.

Siemiatkowski gives the two-word version of the company. Klarna started as a payment solution for shopping online, often tagged as buy now, pay later. Today it does about $100 billion of volume, with half a million merchants and about 100 million consumers who can pay in full or in installments. It is also a fintech and a neobank, with cards and balances, a fully regulated bank, about 4,000 employees, and customers in more than 20 countries.

The number that made him famous comes later in the conversation and is worth putting on the table now. The assistant is doing the work of 700 full-time agents and adding about $40 million of annual profitability. Klarna does about $2 billion of revenue. Plenty of companies had contemplated something like this. Very few had actually put it into production.

The origin story starts in November 2022. Siemiatkowski is on Twitter, people tell him to try ChatGPT, and he is blown away. The last time anything felt like this was Google, twenty years earlier, and this is more impressive. After the holidays he decides Klarna has to lean in. He also decides Sam Altman will soon be impossible to get a meeting with, so he has to try before everyone else.

Sequoia is a shareholder in both Klarna and OpenAI, which makes a decent opener. He flies to San Francisco pretending he has other business. The original slot is two hours. By the time he arrives, the secretary has down-prioritized him and thirty minutes are left. He sits down and pitches a European bank as a guinea pig: always try the latest thing, help OpenAI find the bugs. They set up a joint Slack channel and start experimenting.

Back home he clears the banker's objections first. Data, GDPR, the structures a regulated firm needs. Then he encourages everyone to play. Some people are curious and passionate. Some are not. That is fine. Success, he says, comes back to the creativity of the team. One of the teams that leaned in picked a hard problem: dispute resolution. A customer says the package never arrived. The merchant says it shipped. Klarna has to be a mini court, gather evidence, and decide who covers the cost. Those tickets are complex, slow, and always backlogged.

They start with a copilot. The team builds what is now called RAG, demos it internally within two months, and speeds up decisions while actually raising quality, because the model takes in the relevant information instead of skipping it. They put it in production as a helper. Two months later a Slack message arrives: they are out of errands, send more tickets. That has never happened. The same team then takes on answering customer-service tickets themselves. One rule is non-negotiable: the customer should always know whether they are speaking to AI or a human. A bug breaks that rule for a few thousand conversations. Reading those transcripts, they realize the AI is doing a heck of a good job. That is when they set the bar. Customer satisfaction has to be equal or greater than a human agent's before they get less nervous about real production.

Huang wants the architecture. Where does OpenAI end and Klarna begin? Siemiatkowski, who is usually "freaking transparent," gets coy. He thinks of this as a strategic advantage. What he will say is that instructions have to be clear. If you sit a random human in customer service and the documentation is subpar, because the real knowledge lives in sessions and tribal assumptions, that human will fail. If the FAQ and manuals are detailed enough that any random individual can slowly answer correctly, it works.

That is how he thinks about the AI today. It is an employee that turns up every day and has forgotten everything about what Klarna is and how it works. You have to tell it again. That may change. For now, that is partially the game. Garbage in, garbage out, the old data-science line. Feed the model bad manuals and you get bad answers. Cleaning those manuals had a side effect. The human agents' tools were also subpar, and nobody had put enough managerial attention on them. Fixing the docs made both the people and the model better.

Is he going to replace the CS department 100 percent? Not anytime soon. Some customers will prefer a human for any reason, and he wants to serve them. The biggest quality jump is speed. Human agents run multiple chats. Customers type, wait, and ping "hello, where are you?" He has worked the queue himself. You cannot sit idle while a slow customer thinks. The cost is an average resolution of about 14 minutes. With AI it is two minutes, because the reply is instant. Customers who try it say they want that experience.

The remaining third is not mostly a hunger for human contact. Chatbots have been around for a decade and they have been horrible. Everyone has fought an airline bot about a ticket. Of the 30 percent who skip Klarna's AI, the most common first message is "Agent." They do not trust it to be good. Changing the experience is one job. Changing the perception takes longer. He still expects the AI share to rise. Complex queries remain, and there is tons of work to do. He published the 700-person and $40 million figures knowing people would react, because politicians were too slow and he wanted the world to see this was not just fun demos on Twitter. Klarna uses contracting firms that employ hundreds of thousands of people. As far as he knows, nobody has lost a job over this yet. In the longer term, these jobs will feel it.

Grady hands him a policy magic wand for an era of disruption. The first thing Siemiatkowski wants is surprising: a globally applied electronic identification for humans, a passport that is electronic. Pat and Sonya have no way, right now, to verify that he is not a bot or a fake video. He has seen videos of himself talking to customers that look and sound identical to him, avatars Klarna is about to send to more than a thousand of its top merchants. Being able to authenticate the real Pat, or the real human, is critical if fraud is not going to explode.

The left, he hears, wants to stop the progress. He has a hard time with that, especially because less democratic countries will keep pushing. The right says do not worry, there are always new jobs. He calls that a simplification. In Brussels, as he remembers it, about 10,000 translators turn European legislation into local languages. DeepL and ChatGPT make those roles almost redundant, or at least dramatically smaller. "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." Society can keep progress moving and still support people who cannot simply retrain. That is the vicinity of empathy he wants from politicians.

Which jobs go first? When he met Altman he offered marketing advice: replace the three professions people hate most, CEOs, bankers, and lawyers, and nobody will make a fuss. He holds two of the three. Klarna later tweeted that marketing needed fewer photographers and copyists for day-to-day work, still needing them for the very creative stuff, and the internet erupted. Tweets about AI lawyers get a shrug. He feels sad for the lawyers. On a long horizon he now thinks knowledge work may be easier to replace than driving a truck or building proper robots, the reverse of the self-driving-car fever. Some people would enjoy robots serving while they play football. He is the type who would get depressed. He really enjoys the work.

Internally, marketing is the other sandbox. Everyone has tried an image or a video, been blown away, then asked for brand consistency and gotten stuck. Critics say the tools are unfair to creators. Siemiatkowski partly understands. He also has a colleague who scripted video, sound, and avatars into automatic merchant presentations starring a digital Sebastian. It looks on brand. That person is a creator. When synthesizers arrived, people said sitting at a computer was not playing guitar. Nobody says that anymore. People who may not be photographers or Photoshop experts can now bring what they see in their minds to life.

The production chain shrinks. A German credit-card campaign used to take a month or two, bouncing across teams, with the extra complexity that credit cards are regulated and the copy has to be lawful. Now a few individuals can go from idea to a live campaign in a week. Quality is higher. Compliance is better. Fewer people are involved, which loses some of the old collision of ideas and also removes the bottleneck of an idea that nobody on the team could photograph.

Huang asks what can be outsourced to the model and what still needs the gorgeous in-house photo shoots. He splits copy from image. Imagery can be crazy and imaginary, and he finds that fun. Copy leaves him much less impressed. LLMs work toward the average. They are trained toward the average, and creativity is not the average. Creativity is the extreme of recognizing a new way to combine things. Product descriptions on Klarna's comparison site, millions of listings for clothing and iPhones, are a great LLM job. The quirky line people screenshot and talk about is still much worse. You can generate a lot of versions. They still push toward the middle. Humans remain better at thinking outside the box, because the models are supposed to think in the box.

Alongside the AI wave, Klarna got obsessed with collaborating on information. The company started using Neo4j and graphs, looked at how Wikipedia builds knowledge, and began pulling siloed facts into one place. The side effect is a pile of deprecated enterprise software. Salesforce as a system of record is gone. They remain big Slack customers, which means they are still paying Salesforce, just for a different product. Too many systems had left what the company does inconsistent and dispersed. On top of the unified graph sits Kiki, an internal chatbot. Employees ask how many people work in a unit, what a team does, or what steps you take to launch a system. Kiki shows up in Slack and in a Wikipedia-like view of the graph. Adoption has been tremendous. Productivity impact is obvious to him.

Workflows that used to live on Salesforce went into a mix of Slack workflows, which he admits you could laugh at as one proprietary system replacing another, and Klarna's own stack. The point is the count. Every new place to look for information creates a silo and kills cross-company collaboration. They are moving HR out of Workday into Deel, using Deel for payroll, and deprecating Workday within a few weeks. Org charts, teams, and reporting lines have to live in-house or Kiki cannot understand the company. Paying people on time can stay with a specialist.

Buy versus build is the question every founder asks, with Sierra and Glean already selling adjacent products. Siemiatkowski's example is an employee-engagement bot they built because one-to-five pulse surveys felt too crude. ChatGPT interviewed each willing employee for thirty minutes, then summarized Klarna's strengths and weaknesses. Startups now sell that. Today he might buy it. He is still glad they built it, because the people who did are applying those learnings elsewhere. If something is intuitively better than anything you can build, buy it. There is also massive value in letting people learn a new technology by shipping. They stay a little cautious about buying too much, even while wanting to support startups.

When they debated keeping Workday, he called the CEO and asked to be convinced. Then he tried a funny test. He asked ChatGPT for Workday's API calls. At the time, the documentation sat behind a login, so the model had not been trained on it. Slack's APIs are public, so ChatGPT knows them. Open source is even more familiar. Public docs and public APIs suddenly have a second job: they teach the models how to use your product. Lock that behind a door and it will not get used to the same degree.

On the customer-facing product he gets even more coy. A beta goes live in a few weeks: the customer-service assistant on steroids, better at tickets and starting to advise people on Klarna's own services. Looking at internal projects, he thinks they will launch truly disruptive services within six to twelve months. The direction is older than ChatGPT. In 2015 Klarna was trying to compete with Stripe and Adyen as a payment service provider. When Adyen signed Spotify, a neighbor, they had to look in the mirror. They pivoted, and the management team, most of whom have been with him almost ten years, already described the end state. You wake up and a digital financial assistant says it analyzed your mortgage, it can save you ten bucks by switching banks, and you only have to say yes.

They could not predict the date, the way nobody could for self-driving cars. They could see that bank profits sit on customer immobility and friction. An AI that lets you move by saying yes evaporates a lot of that excess. He sees it in the coming two or three years. Klarna does not want to be one of those banks. It wants to be the assistant that helps you save time, save money, and feel in control. ChatGPT made the timeline feel sooner than they had assumed.

Huang wants an AI stylist in the shopping app. He splits commerce into curation, brand, and infrastructure. Nike will still be Nike. Retailers are more exposed because curation already leaked to TikTok, Instagram, influencers, and the Best Buy associate. Recommendation will rise, and it is harder than it looks. He tried AI for a family road trip in the U.S. this summer and it was pretty bad, because the model does not know who he is. Contact-lens refill reminders are the easy version. An internal test blew his mind: category images in the Klarna app generated from his own purchase history. A Nike-ish shoe he immediately wanted to buy, and a beautifully designed lawn mower matched to his gardening orders. He is not a big shopper. He felt he had seen the next step. Images of things that do not exist yet, generated from a profile, produced only after you click. He notes that Shein-style small-batch testing also means less waste than retailers who buy stock that never sells.

Lightning round: who does he admire most in AI? Sam, an easy question. Does a patron of the arts collect AI work? Not yet. He could. An image should create an emotional reaction, and he does not mind if AI made it, which does not stop him buying human-made art. Advice for founders: smaller companies are already leaning in. Big companies should stop discarding this as Bitcoin, a temporary trend. Be cautious, try it, test it, explore it, learn it. Dinner-table talk about the world ending is fascinating. A meteor might hit him in the head tomorrow. You cannot predict that. The only move is to lean in and understand it better.