How Klarna's AI Assistant Cut Resolution Time from 11 Min to 2 Min

Klarna's AI customer service assistant, launched February 2024 in partnership with OpenAI, became the most-cited reference deployment in fintech support. The numbers are striking: 2.3M chats handled in month one, 2/3 of all customer conversations automated, resolution time down from 11 minutes to under 2 minutes.

But the Klarna AI story has a second chapter that doesn't get discussed as often — by 2025, Klarna publicly walked back some of its AI-only claims and reintroduced human agents for complex cases. That second chapter is the one worth learning from.

This guide covers exactly what Klarna built, how it performed, what broke, and how you can apply the lessons to your own support operation.

TL;DR: Klarna's AI is one of the most ambitious customer support deployments ever. It works, but not as a full replacement for humans. The real lesson is around guardrails, escalation quality, and knowing what AI shouldn't do autonomously.

What Klarna's AI Assistant Actually Is

Launched: February 2024 Partner: OpenAI (GPT-4-class models) Channels: In-app chat, web chat, help center Languages: 35+ Volume: 2.3M chats in month one

Klarna's assistant reads incoming customer messages and handles the full conversation — answering questions, looking up account details, processing refunds, and scheduling payments — without a human agent in the loop for the majority of cases.

Under the hood, it's a tightly integrated system combining:

• A GPT-4-class language model
• Direct integration with Klarna's account and transaction APIs
• A knowledge base of Klarna policies and FAQs
• Escalation logic that routes complex cases to human agents

The Results Klarna Published

From their Feb 2024 announcement and 2024 earnings reports:

Metric	Before AI	After AI
Chats handled by AI	0%	67%
Average resolution time	11 minutes	under 2 minutes
CSAT	Baseline	On par with human agents
Multilingual support	Limited	35+ languages
Cost savings (estimated)	N/A	$40M/year

The $40M number comes from Klarna's estimate that the AI assistant is doing "the work of 700 full-time agents." That framing made headlines — and also became the piece Klarna later had to walk back.

What Went Wrong

By early 2025, Klarna quietly started re-expanding human support capacity. The reasons, based on public statements and industry reporting:

1. Complex cases degraded CSAT. For simple queries (order status, payment schedules), AI matched human performance. For complex disputes, fraud claims, and hardship cases, AI resolution quality dropped noticeably.

2. Hallucinations on edge cases. The model occasionally gave confident-but-wrong answers about policy, fees, or payment terms. In fintech, wrong answers about money are a compliance problem, not just a CSAT problem.

3. "We replaced 700 agents" was misleading. Klarna wasn't replacing 700 agents with AI — they were avoiding hiring 700 new agents during a growth phase. The framing implied layoffs that didn't actually happen.

4. Customer preference for humans on sensitive issues. Even when AI gave correct answers, some customers wanted a human for sensitive financial matters and rated the AI-only experience lower.

Klarna course-corrected by:

• Re-introducing human agents for complex case types
• Adding stricter confidence thresholds before AI auto-responds
• Refining escalation logic to hand off to humans faster on ambiguous queries

Lessons You Can Apply

Here's what matters for any team deploying AI customer support, based on Klarna's experience:

1. Segment queries by risk, not just by volume

Klarna initially pointed AI at all chat volume. The better approach: classify queries by risk band (informational / account-specific / high-stakes regulated) and only automate the lower-risk bands fully.

2. Set strict confidence thresholds

AI should refuse to answer when confidence is low. The instinct is to tune for high automation rate; the better instinct is to tune for high quality of the automated subset. A 50% automation rate with 98% accuracy beats a 75% automation rate with 90% accuracy in regulated contexts.

3. Invest in escalation quality

When AI escalates to a human, the handoff needs to include classified intent, retrieved context, attempted response, and confidence score. Klarna reportedly had weak escalation handoffs initially — customers had to re-explain themselves, which tanked CSAT on escalated tickets.

4. Don't frame it as "AI replaces agents"

Internally and externally, frame AI as "agents work higher up the value chain." Your best human agents should be solving the hardest 20% of cases, not the easy 80%. That's worth celebrating, not hiding.

5. Audit hallucinations weekly

Pull a sample of 100 AI-handled conversations every week and review. Look for:

• Policy claims that don't match your actual policies
• Dollar amounts, dates, or account numbers the AI invented
• Tone misses (dismissive, robotic, overly casual)

Fix content gaps that caused the hallucinations; don't just prompt-engineer around them.

Can You Replicate Klarna's Approach?

Yes — but with important qualifications.

What Klarna had that you probably don't:

• Direct OpenAI engineering partnership
• Full control over their APIs and data
• Willingness to launch publicly and iterate on feedback
• 150M+ users to justify custom engineering

What you can use instead: Platforms like Twig productize the same pattern Klarna built custom:

• GPT-4-class language model handling incoming conversations
• RAG pipeline that grounds responses in your knowledge base
• Confidence-based routing (AI answers when sure, escalates when not)
• Integration with your existing helpdesk (Zendesk, Intercom, Freshdesk)
• Audit trails for every AI decision

The difference: you don't build the platform, you configure it. Deployment takes weeks, not quarters.

See how Twig replicates Klarna's playbook →

Practical Replication Steps

If you wanted to build a Klarna-style AI support assistant for your company:

Step 1 — Audit your content. The AI is only as good as your help docs and past ticket data. Clean these up first.

Step 2 — Classify query risk. Tag every ticket type as low / medium / high risk based on business impact if the AI gets it wrong.

Step 3 — Pick a platform. Custom build if you have a large engineering team and specific needs. Platform (Twig, Intercom Fin, Decagon) otherwise.

Step 4 — Deploy in human-review mode first. Every AI response gets reviewed by a human before sending for the first 30 days. Track override rate — target below 5% before going fully autonomous.

Step 5 — Deploy autonomously on low-risk queries only. Gradually expand the autonomous scope as confidence grows.

Step 6 — Monitor hallucinations weekly. Sample 100 conversations, audit, fix content gaps, iterate.

Klarna's Specific Numbers — Context

When Klarna says AI is doing the work of "700 full-time agents," here's the math:

• 2.3M chats per month
• Assume avg handle time for a human: 11 min
• 2.3M × 11 min = 25.3M min/month = 421K hours/month
• 421K hours / 160 hours per agent per month = ~2,630 agent-hours worth of work

At 67% automation, AI handles ~1,760 agent-equivalents of work. The "700 FTE" figure reflects the ongoing hiring they avoided during a growth phase, not 700 human agents displaced.

Either way, it's the largest public AI customer support deployment we've seen.

FAQ

What is Klarna's AI assistant? A GPT-4-class AI-powered customer service assistant launched February 2024 in partnership with OpenAI. It handles 67% of Klarna's customer chats across 35+ languages.

How much did Klarna save with AI customer support? Klarna estimated $40M in annual cost avoidance by avoiding hiring ~700 additional agents during growth, not by layoffs.

Did Klarna's AI replace human agents? Not in the literal sense. Klarna avoided hiring new agents but retained existing staff. By 2025, they partially walked back the "AI replaced 700 agents" framing and re-expanded human capacity for complex cases.

Can I build an AI assistant like Klarna's? Yes, using platforms like Twig that productize the same pattern Klarna built custom with OpenAI. Expect weeks of setup, not quarters, plus significant time invested in content cleanup.

What went wrong with Klarna's AI customer support? Complex cases degraded CSAT, hallucinations appeared on edge cases, and the "replaced 700 agents" framing was misleading. Klarna course-corrected by reinstating humans for high-complexity cases and tightening AI confidence thresholds.