Klarna: the AI assistant did the work of 700 full-time agents in its first month

• Klarna AI assistant handles two-thirds of customer service chats in its first month — Klarna (пресс-релиз), 2024-02-27
• Klarna's AI chatbot: how revolutionary is it, really? — The Pragmatic Engineer, 2024-03
• Klarna changes its AI tune and again recruits humans for customer service — CX Dive, 2025-05-09

Background

Klarna is a Swedish payments and shopping service, one of the pioneers of the buy-now-pay-later (BNPL) model. Its scale at the time of the AI assistant launch: over 114 million active users, 3.4 million transactions per day, and more than 850,000 retail partners across 23 markets. Customer support at this scale means millions of chats per month in dozens of languages: questions about refunds, payments, invoices, and limits arrive around the clock from jurisdictions with different rules.

The BNPL model has a structural trait that makes support especially loaded: nearly every purchase spawns a payment schedule — and therefore potential questions. Where is my invoice, why this amount, how do I postpone a payment, where is the refund for a cancelled order, why did my spending limit change. In classic e-commerce a customer writes to support mostly when delivery goes wrong; in BNPL there are far more contact points, because the financial relationship continues for weeks after the purchase.

An important detail Klarna disclosed itself: before AI, first-line support was handled by an average of about 3,000 full-time agents — employees of the company's outsourcing partners. Support, in other words, was not an internal team but a large procured function with linear economics: more customers meant more agents and more cost.

When ChatGPT arrived in late 2022, customer support almost immediately became the top candidate for automation with large language models: a conversational format, repetitive scenarios, measurable metrics. Klarna became one of OpenAI's most aggressive early enterprise partners — OpenAI COO Brad Lightcap said in the joint press release that Klarna was 'at the very forefront among our partners in AI adoption'. In February 2024 Klarna became the first major fintech to publish the operational results of moving its first-line support to an AI assistant — with concrete numbers that spread across the industry, made business-press headlines, and became the reference point for the AI-in-support market: 'the equivalent of 700 agents' is still quoted in every other deck about AI in customer service.

Problem

Before the assistant launched, the average errand took 11 minutes to resolve in chat. Routine requests — refunds and cancellations, payment status, invoice and payment-schedule questions — dominated the queue. Each such dialogue required the agent to find the order, check the terms of a specific retailer and the rules of a specific market — routine but not instant work.

Scale made the problem structural. Twenty-three markets mean 23 sets of local refund and payment rules and a need for support in dozens of languages. Covering that with human agents only works extensively: hiring and training through outsourcing partners, peak loads in sale seasons, attrition and constant retraining. An average of about 3,000 agents is one of the company's largest operating cost lines, growing almost linearly with the business.

Repeat inquiries were a pain of their own: customers came back to chat because the issue had not been resolved accurately the first time. Every repeat contact doubles handling cost and compounds the frustration of a customer who has already explained their problem. For a fintech this is not just cost but risk: an inaccurate answer about a payment or refund directly erodes trust in a product people move money through. That is why Klarna later made the reduction of repeat inquiries one of the project's headline public metrics — it is an indicator of accuracy, not just speed.

Finally, previous-generation scripted chatbots did not solve the problem: rigid decision trees could not handle the variety of phrasings across 35+ languages, nor the need to account for the context of a specific order. Anything outside the tree went to an agent — and the economics of support did not change.

Solution

Klarna launched the AI assistant on OpenAI models (via the ChatGPT API) inside its own app — globally, within one month, across all 23 markets at once. The assistant runs 24/7 and communicates in more than 35 languages. Per the press release, it covers the core first-line scenarios: refunds and cancellations, payment issues, disputes, invoice accuracy, current balance, payment-schedule reminders, and spending-limit explanations.

Placing the assistant inside the app is not a technicality but a product decision. The customer is already authenticated in the app, so the assistant immediately sees their orders and payments — no need to burn half the dialogue on 'please provide your order number and the last four digits of your card'. That is what makes answering 'where is my refund' a matter of seconds rather than minutes of clarification.

What we know about the architecture comes from an independent teardown by The Pragmatic Engineer (Gergely Orosz tested the assistant live and analyzed its design). It is a RAG setup over Klarna's own support documentation: the assistant retrieves relevant knowledge-base sections and injects order-specific context — price, merchant, purchase date. On top sit hard guardrails: a whitelist of permitted topics in the system prompt, a verification step comparing the generated answer against allowed topics, and automatic handoff to a human agent as soon as a question goes out of bounds. In the teardown's tests the assistant proved resistant to hallucination within its zone — largely because the zone is deliberately narrow: the model does not 'know' answers, it assembles them from vetted documentation and order data.

There were publicly documented potholes too: despite the guardrails, users quickly demonstrated prompt injection — for instance, getting the support bot to generate program code with cleverly crafted requests. No direct harm resulted (the bot has no access to operations requiring such privileges), but it illustrated an industry-wide fact: LLM non-determinism cannot be fully 'locked down' by a prompt, and the security perimeter must be built around the model — at the level of privileges and access — not only inside the prompt.

The price of the trade-off Klarna chose is worth noting: a narrow topic whitelist means a visible share of non-standard questions goes straight to humans. That is a deliberate exchange — the company sacrificed the bot's 'universality' for predictability and accuracy in financial scenarios, where the cost of an error outweighs the savings on escalation.

The key product decision: AI covers the first line rather than replacing the channel — the customer can reach a human agent at any moment. The rollout speed was unprecedented for a fintech: a global launch across all markets simultaneously, with no months-long per-country pilots — Klarna deliberately chose speed and learning on live traffic. The Pragmatic Engineer's author cooled the hype with a precise observation: essentially Klarna automated L1 support — a tier companies were automating with scripted bots before the LLM era. The difference is that an LLM does it in 35+ languages, with order-context awareness, and at a quality customers rated on par with humans — something scripted bots never achieved.

Result

In its first month the assistant handled 2.3 million conversations — two-thirds of all Klarna support chats — doing the equivalent work of 700 full-time agents. Average resolution time dropped from 11 minutes to under 2 minutes. Repeat inquiries fell by 25% — Klarna attributes this to more accurate answers. Customer satisfaction (CSAT) stayed on par with human agents. The company estimated the impact at $40 million in profit improvement for 2024 and said so publicly, in a press release dated February 27, 2024. Keep the frame in mind: all figures are the company's own reporting for the first month of operation, not externally audited.

The story had a sequel that made the case even more instructive. In parallel with the AI rollout, Klarna froze hiring and cut staff in 2024, publicly tying this to its AI strategy. Then in May 2025, CEO Sebastian Siemiatkowski admitted in a Bloomberg interview that cost-driven automation had produced 'lower quality' and announced the return of human agents — in a flexible 'Uber-type' setup with remote hires. 'It's so critical that you are clear to your customer that there will be always a human if you want,' was how he framed the new position. The AI did not go away: as of mid-2025 the assistant still handled about two-thirds of inquiries, and response times were 82% faster than in the pre-AI era.

In our view, the Klarna case should be read whole, together with the 2025 reversal — that is where it is most useful. Part one proved that an LLM assistant on a narrow first line can genuinely hold two-thirds of the volume with human-level CSAT — validated at a scale of millions of dialogues and transferable to any business with a large flow of routine inquiries. Part two showed the price of radicalism: the metrics Klarna published measured speed and volume but not brand trust or the quality of complex cases — and that is exactly where the debt accumulated that humans had to be brought back to repay.

The market frame matters too: Klarna was preparing for an IPO, and loud AI claims served the investment narrative — as did the subsequent course correction toward 'quality'. For practitioners the takeaway, in our view, is this: the target support architecture is a hybrid where AI removes the routine and humans remain the quality guarantee and the escalation path; projects designed as hybrids from day one never have to go through a painful reversal.

Lessons learned

1. Measure impact in operational metrics support already reports: resolution time, share of volume, repeat inquiries. Klarna reported exactly these — and the numbers are hard to dispute.
2. CSAT parity with humans is the scaling threshold: two-thirds of chats went to AI only while satisfaction held at human-agent level.
3. Multilinguality is the main ROI lever for a global service: one assistant in 35+ languages replaces dozens of local support lines.
4. Reliability comes from narrowing the zone: a topic whitelist, answer verification, and automatic human handoff made the assistant hallucination-resistant — because it does not try to answer everything.
5. Prompt injection is not hypothetical: users got Klarna's bot to generate code within weeks. The security perimeter is built around the model (privileges, access), not only in the prompt.
6. A public P&L line ($40M) turns an AI project from an 'experiment' into an investor-defensible item — but a company's own unaudited figures need framing when cited.
7. The 2025 reversal is the case's biggest lesson: speed and volume automate fast, but complex-case quality and brand trust do not. Design the hybrid upfront so you never bring humans back under pressure.

Frequently asked questions