Klarna's AI Chatbot Saved a Reported $60M. Here's What It Means for Support

Klarna's AI assistant, built with OpenAI, handled two-thirds of the company's customer service chats in its first month and, by Klarna's own account, has saved about $60 million since. In May 2025, Klarna's CEO said cost had weighed too heavily in how support was organized and that quality had suffered, and Klarna began hiring so customers can always reach a person. The lesson for support teams: measure resolution quality alongside speed and savings, and keep a clear path to a person, with AI and human agents working together.

Key points

• Scale: 2.3 million conversations in the first month, two-thirds of all chats, the work of about 700 full-time agents (Klarna, February 2024). By late 2025, Klarna put it at more than 853.
• Savings: a $40 million profit improvement estimated for 2024, and about $60 million in savings credited to the AI by late 2025, both by Klarna's own figures.
• Quality: Klarna reported satisfaction on par with human agents and 25% fewer repeat inquiries. In May 2025, its CEO said cost had been "a too predominant evaluation factor" and the result was lower quality.
• People: customers could always ask for a live agent. In 2025, Klarna started hiring so a person is always available, and tested an "Uber-type" remote setup for its staff.

Klarna's AI customer service timeline

The $60M savings and launch strategy

Klarna built its AI assistant with OpenAI and launched it globally in early 2024. It ran in 23 markets, around the clock, in more than 35 languages, and handled requests from refunds and returns to payment questions and disputes.

A month after launch, Klarna said the assistant was doing the work of about 700 full-time agents. People stayed in the picture: the same announcement said customers could still choose to talk to a live agent, and Bloomberg later reported that Klarna outsourced a few thousand human agents.

Early results: what the data showed

Klarna's February 2024 announcement reported strong first-month numbers:

Metric	Before the AI assistant	First month with AI (Klarna, Feb 2024)
Time to resolve an errand	About 11 minutes	Under 2 minutes
Conversations handled by AI	None	2.3 million, two-thirds of all chats
Work equivalent	None	About 700 full-time agents
Repeat inquiries	Baseline	25% lower
Customer satisfaction	Baseline	On par with human agents
Financial impact	None	$40 million profit improvement estimated for 2024

On paper, it was a clear win: faster answers, fewer repeat contacts, and a large cost benefit. First-month numbers measure speed and volume well. They say less about the hardest cases.

Where Klarna's cost-first approach fell short

In May 2025, Klarna's CEO Sebastian Siemiatkowski told Bloomberg that cost had weighed too heavily when the company organized its support, and that quality had dropped as a result.

Klarna hasn't detailed which requests went wrong. CX Dive reported customer complaints that the AI gave generic answers and struggled with more complicated questions. Those are common failure modes of AI-first support in general: answers that loop or stay generic when the problem is specific, and no easy way to reach a person. In financial services, where customers are often worried about money, gaps like these cost trust quickly.

Klarna's response was to invest more in people, so customers can always reach one, and to keep AI agents that escalate complex queries to human staff.

Speed and cost tell part of the story. Resolution quality, repeat contacts, and customer satisfaction show whether the AI is solving problems.

sbb-itb-97114f1

Klarna's ROI and performance data

Cost reduction and speed improvements

Klarna's own figures grew over time. Read them with their dates and sources, since they measure different things:

Date	Source	Figure	What it is
February 27, 2024	Klarna press release	$40 million	Estimated profit improvement for 2024
February 27, 2024	Klarna press release	About 700	Full-time agents' worth of work done by the AI
November 2025	CEO on the Q3 2025 earnings call, as reported by CX Dive	About $60 million	Savings Klarna credits to the AI assistant
November 2025	CEO on the Q3 2025 earnings call, as reported by CX Dive	More than 853	Full-time agents' worth of work, up from about 700

Resolution time fell from about 11 minutes to under 2, a drop of more than 80%, and in 2024 the AI handled about two-thirds of all chats.

Building an AI assistant like Klarna's has two kinds of cost: a one-time cost to set up and customize the models, and an ongoing cost to run them, for example API usage per conversation. Klarna hasn't broken these out. If you're pricing AI support for your own team, check which of the two your vendor charges and how the running cost scales with volume. Resolution-based pricing, such as CoSupport AI Agent's $0.50 per AI-resolved ticket with unresolved tickets free, ties the running cost to tickets the AI actually resolves.

Customer satisfaction scores

Klarna's 2024 release said customer satisfaction with the AI was on par with human agents, and that repeat inquiries fell 25%. Klarna repeated the "on par" claim in late 2025. Both are Klarna's own measurements.

An overall average can look steady while specific request types struggle. Track CSAT and repeat contacts by request type, so the hard cases show up on their own.

Why Klarna is investing in human support again

What Klarna's CEO said about cost and quality

In May 2025, Siemiatkowski told Bloomberg:

"As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." - Sebastian Siemiatkowski, CEO, Klarna

He added that investing in human support was the way forward for Klarna. The AI stayed in place: a Klarna spokesperson told CX Dive the same month that it still handled about two-thirds of customer inquiries.

The lesson for other teams is about the scorecard. If cost and speed are the main measures, the setup will lean toward them. Track repeat contacts and resolution quality alongside resolution time, so you can see whether problems get solved.

How Klarna's AI-plus-human model works

In May 2025, Klarna said it was hiring so customers always have the choice of a human, and that it was testing an "Uber-type of setup" where staff log in and work remotely. Reports said it planned to recruit students, people in rural areas, and loyal Klarna users for these roles.

Klarna hasn't published how it splits requests between AI and people. A common split in AI-plus-human support looks like this:

Area	AI agent	Human agent
Task type	Routine, high-volume requests	Complex, sensitive, or unusual cases
Examples	Payment status, standard refunds, return policies	Disputes, custom payment plans, fraud, upset customers
What to measure	Resolution rate, repeat contacts	Quality, customer satisfaction, time to resolution
Handoff	Passes the case on when it can't resolve it or the customer asks for a person	Picks up the case with the full conversation history

In Klarna's case, automation still carries most of the volume: about two-thirds of chats in 2024 and, by its late-2025 count, the work of more than 853 agents.

How to mix AI and human support effectively

Which tasks to give AI

Start with tasks that take up your team's time but don't need a human touch. AI does well on routine, high-volume questions with clear answers: payment status, standard refunds, and return policies.

Keep people for the moments that matter: disputes, custom payment plans, fraud, and frustrated customers. If a case involves a policy exception or strong emotion, a person should handle it. That keeps costs in check and quality high where it counts.

Watch your repeat contact rate alongside resolution speed. If customers keep coming back with the same problem, the AI may be closing tickets without solving them.

Key takeaway: give AI routine, standard questions and leave high-stakes or emotional conversations to people. Then teach the AI when to hand off.

Training AI to escalate properly

Getting AI to know when to hand off matters as much as getting it to answer. Set escalation triggers: repeated frustration, keywords like "fraud" or "dispute", or the AI giving the same answer twice without resolving anything.

Build your business goals into the AI's rules alongside speed and ticket closures. If "keep the customer's trust" is a goal, the AI should escalate before it closes a ticket it hasn't really solved.

Always keep a clear path to a person. Customers should be able to type "I need a representative" or click a visible button. Forcing frustrated customers through endless bot replies erodes trust and drives repeat contacts.

Check what the AI tells customers as well as how fast it answers. In 2024, a Canadian tribunal held Air Canada liable after its chatbot gave a customer wrong information about bereavement fares. The airline was responsible for what its bot said.

Adapting AI to your support workflow

Fit the AI to your business. Generic chatbots fail because they don't know your policies, products, or customers. Limit the AI to verified sources to reduce errors.

Give your team an assistant too. CoSupport AI Assistant drafts replies, summarizes long conversations, and suggests resolutions, and an agent reviews every reply before it reaches the customer.

Test the AI on real past tickets before launch, and keep testing after, so you catch the cases where it fails.

Present human support as the premium option for complex situations. Customers then understand why they start with the AI and still feel valued when a person steps in.

What support leaders can learn from Klarna

Klarna's AI brought real savings and much faster responses, by its own figures. In May 2025, its CEO said that putting cost first had lowered quality, and the company began hiring so customers can always reach a person.

Klarna kept the AI. A spokesperson said that month it still handled about two-thirds of inquiries. What changed was the weight Klarna puts on quality and on human support.

For support leaders, the lesson is to measure what matters. Track repeat contacts, resolution quality, and CSAT by request type alongside speed. Cut staff only as fast as your AI actually resolves your customers' problems, and keep an easy path to a person.

What to measure	Why it matters	What Klarna's story shows
Resolution time	Shows speed	Fell from about 11 minutes to under 2 in Klarna's first month
Repeat contacts	Shows whether problems were solved	Klarna reported 25% fewer in 2024
CSAT by request type	Shows quality on hard cases	Klarna reported CSAT on par with humans; its CEO still said quality fell
Path to a person	Shows whether customers can reach someone	Klarna said in 2025 it was hiring so customers can always reach a person

Key takeaway: balance automation with people. Watch resolution quality as closely as speed, keep escalation paths clear, and invest in both AI and skilled agents.