Klarna's AI agent carries the workload of 853 employees. The case study covers three decisions that worked, one lesson missed and a 9-12 month adaptation plan.
Link copied
At the start of 2024, the head of customer service at a Stockholm-based fintech was close to resigning. The pressure points were clear: call volumes had exploded across 35 languages, average resolution time had crept up to 11 minutes and NPS was wobbling.
Eighteen months later, the same manager was looking at a very different dashboard. Average resolution time had fallen to 2 minutes, NPS had held steady and the company's leadership was reporting an annual saving of 60 million dollars.
The Klarna AI agent case was consumed by the popular press under the headline "AI beat the humans". Read the file line by line and a different picture emerges. The result was produced by tight operational discipline, a measured rollout plan and a workforce kept in the right place, rather than by any single clever model.
In the figures shared publicly, Klarna's leadership reported that the AI assistant handled 2.3 million conversations in its first month, a volume equal to the monthly workload of 700 full-time employees.
Customer satisfaction stayed at the same level as conversations handled by human agents. Repeat contact rates fell by 25 per cent. Operational savings started at 40 million dollars and reached 60 million once annualised.
A second measurement came in the third quarter of 2025. This time the figure was updated to the workload of 853 employees. Klarna had grown, so the base had grown with it; the AI's capacity expanded faster than the human workforce.
The story carries on past the headline. In the same period, the company's leadership announced that certain categories had been handed back to human agents. The AI had struggled to capture enough nuance in premium customer disputes and rare edge-case combinations. Klarna's leadership treated this as calibration. That is the lesson worth keeping.
A narrow starting point. The Klarna team first deployed the AI agent on returns and refund scenarios. These accounted for 40 per cent of total contacts, success was measurable and the cost of error was low. It was a textbook application of the bowling pin strategy.
Two-way logging. Every AI interaction was recorded alongside a signal that could route the customer to a human agent. When a customer said "I don't want to talk to an AI", the system handed over without resistance. Having watched contact centres wrestle with this for years, I can say most of them set the AI up as a barricade in front of the customer. Klarna set it up as a door that opens onto a human agent. Customers sense the difference within seconds.
A feedback loop that keeps feeding the model. Every week, Klarna's quality team sampled conversations in selected categories and reviewed the AI's answers. Errors flowed back into prompt refinement, training data and fallback rules. Without that structure the model would have degraded within 6 months.
The gap Klarna's leadership acknowledged at the end of 2025 was this: as the team shrank, redeployment moved too slowly. When a workload equal to 700 people shifted to AI, the freed human capacity failed to flow automatically into other areas of value. Part of it left the organisation.
The company's leadership corrected course later. The lesson stands: if an AI agent is rolled out without a parallel redeployment plan, trust inside the brand takes a hit.
McKinsey's 2025 State of AI report points to the same spot. The workforce impact of generative AI is defined as "substitution plus reallocation". Apply substitution on its own and you collect the efficiency while absorbing the cultural loss.
Across retail, fintech and subscription businesses the question is always the same: which contacts should a human take and which should the system take? A Klarna derivative is achievable within 9-12 months.
Months 1-2, data collection and classification. Pull the last 6 months of contact records and tag them by topic, duration, resolution and customer segment. Split them into 6-8 core categories such as returns, delivery status, account access and billing queries, then measure each category's volume, marginal cost and tolerance for error.
Months 3-4, the first narrow pilot. Choose the single category with the highest volume and the lowest risk; it is usually the "where is my order" question. In 80 per cent of these contacts the answer is already known. Build the AI agent, route the remainder to human agents and measure CSAT weekly.
Months 5-7, scaling. Once the first category is stable, open up returns, billing and account access in sequence. Every new category gets a 4-week observation window. If NPS dips, roll it back.
Months 8-10, the premium customer trigger. This is the point Klarna spotted late. Certain customer segments want to speak to a person directly, and the CRM needs to see that. The AI gives the initial answer, then says "I'm connecting you to your dedicated adviser".
Months 11-12, reallocation. Redirect the freed agent capacity towards outbound calls, retention and premium service. Skip this step and you repeat Klarna's mistake.
| Channel | Volume | Error Tolerance | First-Phase Fit |
| Chat (web) | High | Medium | Start here |
| High | Medium-high | Phase 2 | |
| Voice (calls) | High | Low | Phase 3 |
| Medium | High | Quick win | |
| Social media | Low | Low | Keep with humans |
The voice channel goes last. Voice AI remains the least mature layer; accents, dialects and background noise still disrupt it, so text-based channels give a far more controlled rollout.
Over 18 months, Klarna's AI agent absorbed the workload of 853 employees and generated an annual saving of 60 million dollars. The result came from a narrow starting point, continuous quality measurement, a hybrid flow between human and AI, and a reallocation plan run in parallel.
Organisations can apply the same model within 9-12 months using category-based pilots. To bring this case to your leadership team as a keynote, contact Speaker Agency about our artificial intelligence speakers. For concrete selection criteria, the guide on how to choose an AI keynote speaker is a practical starting point.
In Klarna's first announcement, the AI assistant had taken on the workload of 700 full-time employees. A later update raised that figure to 853; as the company grew, the AI's capacity scaled with it. Annual operational savings reached 60 million dollars and NPS held steady.
A realistic window is 9-12 months. The critical point is keeping the first pilot category narrow, starting with a topic that combines high volume with high error tolerance. Work in 30-day measurement cycles and open the second category after 90 days.
With a poor setup, yes; with the right setup, no. Klarna's NPS held steady because the "I want to speak to a person" signal was accepted by the system without resistance. Set the AI up as a barricade and customers get angry; set it up as a door and customers start choosing the AI themselves.
E-commerce, fintech, telecoms and insurance are the best fits. They share the same profile: most contacts are repetitive, the need for nuance is low and regulation sits at a moderate level. In healthcare and law, the same model demands a much stricter human approval layer.
It started with returns and refund scenarios. The category accounted for 40 per cent of total contacts; success was measurable and the cost of error was low. The narrow start gave the model room to be calibrated in the field.
Voice AI still delivers inconsistent results when faced with accents, dialects and background noise. Text-based flows on chat and email run under far tighter control. That is why the voice channel sits in the final phase of the decision matrix.
After Dinner Speakers
Contents
