Klarna, the Swedish "buy now, pay later" fintech, runs customer support across 23 markets in 35+ languages, 24/7. Routine errands — refunds, returns, payment-schedule and dispute questions — dominated agent time and drove long resolution windows (around 11 minutes) and a steady stream of repeat contacts when first answers were imprecise.
The challenge
What was deployed
In February 2024 Klarna launched an in-app AI assistant powered by OpenAI, handling the most common support categories end-to-end with one-tap escalation to human agents on request. The assistant covered refunds, returns, cancellations, disputes, invoice errors, and balance/spending-limit queries across all markets simultaneously.
Architecturally this is a textbook agentic customer-service pattern: a language model fronts the conversation, retrieves account context, executes scoped account actions, and hands off to a person when confidence or policy requires it.
The results
Klarna reported that in its first month the assistant had 2.3 million conversations — two-thirds of its customer-service chats — doing the equivalent work of 700 full-time agents. Customer-satisfaction scores were "on par with human agents," resolution time fell from 11 minutes to under 2, and more accurate resolution drove a 25% drop in repeat inquiries. Klarna estimated the assistant would drive a $40M profit improvement in 2024.
Twarx analysis
Original interpretationKlarna's real lesson isn't "700 agents replaced" — it's that AI cleared the high-volume, low-complexity tier brilliantly, then exposed the limits of optimising customer service for cost alone. The 2025 re-hire is the more instructive data point than the 2024 launch.
The 2.3M-conversations number is real and impressive, but the story most people miss is the shape of the win: AI excelled at the routine tier — refunds, tracking, payment questions — where speed and consistency matter most, cutting resolution from 11 minutes to under two. That tier is where almost every support org should start.
What makes Klarna unusually instructive is the 2025 correction. By their own CEO's account, leaning on AI as a pure cost lever produced "lower quality" on cases that need judgment, and they re-invested in human support. The durable model isn't AI-vs-humans; it's AI for volume, humans for the quality-sensitive tail. Teams quoting the $40M figure should remember it was a projection — the more defensible claims are the resolution-time and repeat-contact improvements, which are operational rather than financial.
Illustrative Twarx model
Illustrative human agent-hours deflected per month, by ticket volume
100k tickets/mo
6,700 hrs/mo
500k tickets/mo
33,500 hrs/mo
2M tickets/mo
134,000 hrs/mo
Method & assumptions: Assumes a two-thirds AI deflection rate (Klarna's reported share) and ~6 minutes of saved human handling per deflected contact. A Twarx model for scale intuition only — not Klarna's actual figures — and it ignores quality/escalation effects and the cost of the AI itself.
Read the numbers honestly
This is not an unqualified success story, and presenting it as one would be dishonest.
- The $40M is a company projection, explicitly worded as an estimate — not audited realised savings.
- 2025 course-correction: CEO Sebastian Siemiatkowski said a cost-first approach produced "lower quality" service and that Klarna would re-invest in human support. The lesson the industry took away was about balance, not reversal — AI for volume, humans for quality-sensitive cases.
- The "700 agents" figure describes work the AI handled, and is frequently (and wrongly) conflated with a separate 2022 workforce reduction. Klarna's release does not claim 700 people were let go.
- "On par" CSAT and resolution-time gains were self-reported by Klarna.
Timeline
Feb 2024
Global launch; 2.3M conversations in month one
2024
$40M profit-improvement projection for the year
May 2025
Klarna re-invests in human support to lift quality
Frequently asked
Did Klarna's AI assistant replace 700 employees?
No. The 700 full-time agents figure describes the volume of work the AI handled in its first month, not 700 layoffs. It is often conflated with a separate 2022 workforce reduction; Klarna's own release does not claim 700 people were let go.
Was the $40M a real, audited saving?
It was a company projection for 2024 profit improvement, explicitly worded as an estimate. Treat it as directional, not as audited realised savings.
Why did Klarna later bring back human agents?
In 2025 Klarna said a cost-first approach had produced lower quality service and that it would re-invest in human support — a recalibration toward AI for volume and humans for quality-sensitive cases, rather than a full reversal.
AI AgentsCustomer ServiceFintechOpenAIROI
TR
Analysis by
Twarx Research Team · Applied AI Research
Twarx researches and deploys enterprise AI agents with a measurement-first method: every metric traced to a primary source, projections labelled as estimates, and limitations stated up front.
AI agentsAgentic AICustomer service automationEnterprise deploymentSLMs
