Klarna

Mixed evidenceDecision rights movedCreator to judgeScope widened on evidence

First-line customer-service errand resolution

That a first-line customer-service errand requires a human agent to read, interpret and resolve it, so resolution capacity can only grow by adding agent headcount and outsourced capacity market by market.

Customer service · Global; 23 markets and more than 35 languages at launch

Collections: Embodied work

A growing queue of routine customer requests contrasts with automated resolution and expert human handling for complex conversations.

Executive brief

The operating-model shift, in one view.

Klarna moved first-line resolution authority to an agent and kept humans for the conversations where judgement and empathy actually change the outcome. The instructive part is not the speed figure, it is the sequencing: Klarna set a narrow errand scope, instrumented two quality proxies it was willing to publish (CSAT parity and repeat-inquiry rate), preserved a standing customer right to reach a human, and only then widened the agent's authority into level-two work. The 2025 rebalancing is the more useful lesson than the 2024 launch - when quality complaints surfaced, Klarna did not retreat from automation, it re-invested in higher-skilled in-house humans for the residual and kept expanding the agent.

AI value · Average customer errand resolution time

Company-reported

Average resolution time: 11 → under 2 minutes

Before

Opens an in-app service chat about a refund, return, payment issue, cancellation, dispute, or invoice. → Picks the chat off the queue, interprets the errand, looks up account and order data, and resolves it or escalates. → Handles complex or sensitive errands the first-line agent cannot resolve. → Adds agent headcount…

After

Opens a customer-service chat in the Klarna app and is met by the AI assistant. → Interprets and resolves the errand end to end - refunds, returns, payment-related issues, cancellations, disputes. → Handles higher-end outsourced conversations and any issue a customer chooses to route to a human. → Extends the…

Human boundary

The AI assistant resolves errands within its scope without a disclosed per-errand human approval. Klarna does not publish an authority matrix; the disclosed control is the customer's standing right to choose a live agent instead. Scope boundaries (which errand types the assistant may close) are set by Klarna product…

Why it matters

Leaders should also read '700 agents' correctly: it is avoided hiring during growth, not a headcount cut, and treating an equivalence claim as a savings claim is how these cases get discredited.

How the work changed

Before

How the work ran before the change.

Step 1 of 4
Customer
Opens an in-app service chat about a refund, return, payment issue, cancellation, dispute, or invoice.
ControlSelf-service help content; queue wait before a human picks up.
Step 2 of 4
First-line human service agent (largely outsourced)
Picks the chat off the queue, interprets the errand, looks up account and order data, and resolves it or escalates.
ControlAgent handbook, quality-assurance sampling, supervisor escalation.
Step 3 of 4
Level-two human agent
Handles complex or sensitive errands the first-line agent cannot resolve.
ControlSupervisor review.
Step 4 of 4
Service operations management
Adds agent headcount and outsourced capacity to absorb volume growth across markets and languages.
ControlHiring plan and outsourcing vendor contracts.

What changed

Routine first-line service does not require a human agent.

Decision rightAI resolves bounded errands; people own complexity and escalation

After

How the same work runs now.

Step 1 of 4
Customer
Opens a customer-service chat in the Klarna app and is met by the AI assistant.
ControlKlarna states customers 'can still choose to interact with live agents if they'd prefer' - customer-initiated escalation is the disclosed safeguard.
Step 2 of 4
AI assistant (OpenAI-powered)
Interprets and resolves the errand end to end - refunds, returns, payment-related issues, cancellations, disputes.
ControlCustomer-satisfaction score tracked against human-agent CSAT; repeat-inquiry rate tracked as an accuracy proxy.
Step 3 of 4
Human service agent (increasingly employed by Klarna)
Handles higher-end outsourced conversations and any issue a customer chooses to route to a human.
ControlKlarna-employed staff used deliberately for quality-sensitive conversations; CEO-stated intent that a customer can always reach a human.
Step 4 of 4
Product and service operations
Extends the assistant's scope as capability improves, including into level-two support.
ControlExecutive review of scope expansion; volume and equivalence metrics restated publicly.

Process model built from the published workflow evidence for Klarna. Every step, actor, and control appears in full below.

Every step, actor, and control

Exception path

Customer-initiated escalation is the primary path: Klarna's launch release states customers can still choose to interact with live agents if they prefer. Klarna additionally routes higher-end conversations to human agents and, from 2025, hires in-house staff specifically for those conversations rather than relying on outsourced capacity.

Work removed

Queue wait before a human agent picks up a first-line chat
Manual interpretation and account look-up for routine refund, return, payment, cancellation, dispute and invoice errands
Repeat contacts caused by inconsistent or inaccurate first-line answers (reported down 25%)

Decision authority

Before

01
Customer
Opens an in-app service chat about a refund, return, payment issue, cancellation, dispute, or invoice.
Control: Self-service help content; queue wait before a human picks up.
02
First-line human service agent (largely outsourced)
Picks the chat off the queue, interprets the errand, looks up account and order data, and resolves it or escalates.
Control: Agent handbook, quality-assurance sampling, supervisor escalation.
03
Level-two human agent
Handles complex or sensitive errands the first-line agent cannot resolve.
Control: Supervisor review.
04
Service operations management
Adds agent headcount and outsourced capacity to absorb volume growth across markets and languages.
Control: Hiring plan and outsourcing vendor contracts.

After

01
Customer
Opens a customer-service chat in the Klarna app and is met by the AI assistant.
Control: Klarna states customers 'can still choose to interact with live agents if they'd prefer' - customer-initiated escalation is the disclosed safeguard.
02
AI assistant (OpenAI-powered)
Interprets and resolves the errand end to end - refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies - 24/7 in more than 35 languages, and provides balance and payment-schedule information.
Control: Customer-satisfaction score tracked against human-agent CSAT; repeat-inquiry rate tracked as an accuracy proxy.
03
Human service agent (increasingly employed by Klarna)
Handles higher-end outsourced conversations and any issue a customer chooses to route to a human.
Control: Klarna-employed staff used deliberately for quality-sensitive conversations; CEO-stated intent that a customer can always reach a human.
04
Product and service operations
Extends the assistant's scope as capability improves, including into level-two support.
Control: Executive review of scope expansion; volume and equivalence metrics restated publicly.

Work that left the path

Queue wait before a human agent picks up a first-line chat
Manual interpretation and account look-up for routine refund, return, payment, cancellation, dispute and invoice errands
Repeat contacts caused by inconsistent or inaccurate first-line answers (reported down 25%)
Language-by-language staffing to cover 35+ languages and 23 markets around the clock
Incremental first-line hiring and outsourced capacity expansion to absorb volume growth

Human role before

First-line human agents, largely outsourced, read and resolved the majority of customer-service chats; service capacity scaled with headcount, and level-two agents absorbed complexity.

Human role after

Humans no longer hold first-line resolution for the majority of chats. Klarna retains and actively hires human agents for higher-end conversations it previously outsourced, and any customer may choose a human at any point. The human role shifts from volume absorption to judgement-heavy and quality-sensitive conversations.

AI roleAutonomous first-line resolution agent. It interprets the customer's errand, retrieves the relevant account and order context, and completes the resolution - including refunds, returns, cancellations and disputes - without a per-errand human review step, across 23 markets and more than 35 languages.

Outcomes

Average customer errand resolution time

Company-reported

11 minutes under human first-line handling, as stated by Klarna→Under 2 minutes

First month of global operation, reported 2024-02-27; restated by Klarna in May 2025 as an average of two minutes versus eleven minutes with human agents · 2.3 million conversations in the first month, equal to two-thirds of Klarna's customer-service chats, across 23 markets and more than 35 languages; approximately 1.3 million errands per month as of May 2025

Company-measured and company-defined. No independent measurement of resolution time was located: Forbes explicitly attributes the figures to a company statement, and CMSWire's 2025 restatement is attributed to Klarna ('Klarna tells me'). Both the 11-minute baseline and the sub-2-minute result are Klarna's own figures, and Klarna does not publish the measurement definition or the errand mix behind either number.

Repeat inquiries after a first-line resolution

Company-reported

Pre-AI repeat-inquiry rate under human first-line handling; the absolute rate is not disclosed→25% drop

First month of global operation, reported 2024-02-27; still stated as 25% fewer repeat inquiries in May 2025 · 2.3 million conversations in the first month, two-thirds of Klarna's customer-service chats

Company-measured. Klarna presents the reduction as evidence the assistant is 'more accurate in errand resolution', which is an inference from repeat contacts rather than a direct accuracy measurement. The absolute baseline repeat rate is not disclosed, so only the relative change can be assessed.

Share of customer-service chats resolved without a human agent

Company-reported

0% before launch - all customer-service chats were handled by human agents→Two-thirds of Klarna's customer-service chats

First month of global operation (reported 2024-02-27), sustained at two-thirds as of 2025-05-20 · 2.3 million conversations in month one; approximately 1.3 million errands per month in May 2025, which Klarna equates to work previously done by about 800 people

Klarna does not define the denominator. The release states 'two-thirds of Klarna's customer service chats' while separately noting the assistant is 'available in the Klarna app', so it is unclear whether the denominator is all support contacts or app chat volume only. The share is company-reported in both 2024 and 2025 and was not independently audited.

What leaders can reuse

Anti-pattern

Reporting an AI-equivalence figure as a headcount saving. Klarna's '700 full-time agents' is an equivalence calculation Klarna has never published the method for, and Klarna itself told Forbes that AI did not directly explain its headcount decline. Presenting it as 700 jobs removed - or citing the company-wide 40% headcount fall as a customer-service result - overstates the case and invites exactly the credibility backlash Klarna faced in 2025. The second anti-pattern is measuring only deflection: Klarna's own 2025 experience was that quality on complex and emotionally loaded errands is where automation fails first, and an organisation tracking only containment rate will not see that failure until customers do.

Questions

01Which specific errand types are we willing to let an agent close without human review, and who signs off on widening that list?
02What quality proxy would we be willing to publish alongside our deflection rate, and do we measure it today?
03If our agent handles two-thirds of contacts, what is the residual third actually made of - and are we staffing it with better people or with whoever is left?
04Are our capacity gains avoided hiring or headcount reduction, and can we defend the distinction publicly with our own numbers?
05What is our equivalent of the customer's standing right to reach a human, and how much friction have we put in front of it?
06How would we detect quality degradation on complex or emotionally loaded cases before it shows up in press coverage?

Portability conditions

The work is high-volume, transactional and reversible - refunds, returns, cancellations, invoice corrections - so an error is correctable rather than catastrophic
Account, order and payment context is available to the agent through APIs at conversation time
The organisation can publish and defend at least one quality proxy (CSAT parity, repeat-contact rate) rather than reporting only deflection
Customers retain a real, low-friction path to a human, and the organisation is willing to staff it
Volume is growing, so capacity gains can be taken as avoided hiring rather than as redundancies
Multi-language, 24/7 demand exists, which is where the economics are strongest relative to human staffing

Reputation risk

medium - the case is inseparable from a job-displacement narrative, and Klarna's own '700 agents' and 40%-headcount statements have been widely contested. Any use of this case must present the equivalence figure as avoided hiring rather than displacement, must include the 2025 rebalancing toward in-house human agents, and must avoid implying that the reported figures were independently audited. All headline metrics are Klarna-measured.

Evidence and authority

What the public record supports.

Current · updated 2026-08-22

2 independent, 1 primary; publication outcomes are verified and reported.

Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 6dd10530a1615545

Related transformations

Predictive Test Selection for change-based regression testing

That every code change must run every test that transitively depends on modified files in order to catch regressions before trunk. Meta's predictive model estimates each test's likelihood of catching a regression for a specific change, so a much smaller learned subset can be run without loss of regression signal.

Twenty-eight South-East Queensland long-term care facilities · Creator to judge

Sources

Klarna AI assistant handles two-thirds of customer service chats in its first monthKlarna · 2024-02-27 · primary
No AI Retreat: Klarna's Bots Are Taking Over More Customer Service TasksCMSWire (interview by Alex Kantrowitz) · 2025-05-20 · independent
Klarna's AI Assistant Is Doing The Job Of 700 Workers, Company SaysForbes (contributor) · 2024-03-04 · independent

Read the evidence, freshness, caveat, and version policy.

First-line customer-service errand resolution

The operating-model shift, in one view.

How the work changed

Outcomes

Average customer errand resolution time

Repeat inquiries after a first-line resolution

Share of customer-service chats resolved without a human agent

What leaders can reuse

What the public record supports.

Related transformations

Predictive Test Selection for change-based regression testing

PARO robotic-seal engagement sessions

Contact-Center Reskilling

AI-Drafted Customer Email / Arlo Assistant

Sources