AI vs. Customer Service: The Klarna Playbook and the Death of the Call Center

Executive Summary

In February 2024, Klarna announced that its AI assistant — built on OpenAI's GPT-4 — was handling two-thirds of all customer service interactions within its first month of deployment. The numbers were striking: the equivalent of 700 full-time agents, 2.3 million conversations resolved, average resolution time cut from 11 minutes to under 2 minutes, and a 25% reduction in repeat inquiries. By Q1 2026, Klarna has reduced its global workforce from approximately 5,000 to under 3,500, with customer service bearing the largest share of cuts.

Klarna is not an outlier. It is a leading indicator. The global contact center industry employs roughly 17 million people and generates approximately $340 billion in annual revenue. It is one of the largest employers of non-degree workers in the global economy, and it sits squarely in the crosshairs of large language models that can understand intent, retrieve information, follow scripts, and generate natural-sounding responses — the exact skill set that Tier 1 and Tier 2 support agents are hired for.

This report examines the Klarna case in depth, maps the economics of call center operations before and after AI deployment, analyzes the counterintuitive quality dynamics, and projects displacement timelines across support tiers. We also assess the strategic response of major BPO (Business Process Outsourcing) providers and identify which human roles will persist in an AI-dominated customer service landscape.

The Klarna Case Study: Anatomy of a Displacement Event

Before AI: The Cost Structure

Klarna, the Swedish buy-now-pay-later company, operated customer service across 23 markets and 35 languages before its AI deployment. The operation relied on a blended model: approximately 700 in-house agents supplemented by outsourced capacity from BPO providers in the Philippines, India, and Eastern Europe. The fully loaded cost per agent (salary, benefits, training, facilities, technology) varied by geography but averaged roughly $35,000-$45,000 per year for offshore agents and $55,000-$75,000 for onshore agents in Western Europe and the United States.

The dominant cost drivers were predictable: labor accounted for 65-75% of total contact center spend, with technology infrastructure at 10-15%, facilities at 5-8%, and training and quality assurance consuming the remainder. Agent attrition ran at 30-45% annually — an industry-standard figure that meant Klarna was effectively retraining a third of its support workforce every year.

The unit economics told a clear story. Average handling time (AHT) for a Klarna customer interaction was approximately 11 minutes. With agents handling roughly 40-50 interactions per shift, the cost per resolution landed in the $3.50-$7.00 range depending on geography and complexity. Multiply that across tens of millions of annual interactions, and customer service represented a nine-figure cost line for the company.

The AI Deployment

Klarna's AI assistant launched globally in January 2024 as a collaboration with OpenAI. Rather than deploying a narrow chatbot with decision trees, Klarna implemented a conversational AI system with access to customer account data, order history, payment schedules, and refund policies. The system could process refunds, adjust payment plans, provide order tracking, answer policy questions, and escalate to human agents when it detected situations beyond its capability.

The results, disclosed publicly by Klarna CEO Sebastian Siemiatkowski, established a new benchmark for AI customer service deployment:

• 2.3 million conversations handled in the first month
• Two-thirds of all inbound interactions resolved without human involvement
• Average resolution time: under 2 minutes (down from 11 minutes)
• Customer satisfaction scores: on par with human agents (no statistically significant difference)
• Repeat contact rate: down 25%, suggesting first-contact resolution improved
• Available 24/7 across all 23 markets and 35 languages simultaneously

Klarna estimated the AI assistant was doing the equivalent work of 700 full-time agents. At an average fully loaded cost of $40,000-$50,000 per agent, that translates to $28-$35 million in annualized labor cost savings from a single deployment. The cost of running the AI system — API calls, compute, fine-tuning, monitoring — was a fraction of that figure, likely in the $3-$5 million range annually based on publicly available OpenAI enterprise pricing.

The Workforce Impact

Klarna's headcount trajectory tells the story that the press releases sanitize. The company employed approximately 5,000 people at its peak in 2022. By the end of 2024, that number had fallen to around 3,800. As of Q1 2026, Klarna operates with fewer than 3,500 employees globally. The company has implemented a hiring freeze for roles that AI can perform and has publicly stated its intent to let natural attrition — combined with AI capability expansion — continue reducing headcount.

Critically, Klarna did not execute mass layoffs in a single event. The reduction was achieved through attrition (not replacing departing agents), contract non-renewals with BPO providers, and targeted restructuring. This approach — sometimes called "silent displacement" — is likely to be the dominant pattern across the industry. It avoids the reputational cost of layoff headlines while achieving the same economic outcome over a 12-24 month window.

Call Center Economics: The Before and After

The Traditional Cost Model

The global contact center industry operates on remarkably thin margins. BPO providers typically earn 8-15% operating margins on customer service contracts, with labor arbitrage (offshoring to lower-cost geographies) as the primary margin lever. The economics have been well-understood for decades:

Cost Component	% of Total Spend	Notes
Agent labor (salary + benefits)	65-75%	Largest single cost; varies 3-5x by geography
Technology (telephony, CRM, WFM)	10-15%	Increasingly cloud-based; falling per-seat costs
Facilities	5-8%	Declining with remote/hybrid work post-COVID
Training & quality assurance	4-6%	Recurring due to high attrition
Management & overhead	5-8%	Supervisors, team leads, operations

A typical offshore contact center operation (Philippines or India) runs at $8-$12 per agent-hour fully loaded. Onshore operations in the U.S. or Western Europe run at $25-$45 per agent-hour. The blended global average sits at approximately $15-$20 per agent-hour.

With agents handling 6-8 interactions per hour (depending on channel — voice is slower than chat), the cost per interaction ranges from $1.50 (simple chat, offshore) to $7.00+ (complex voice, onshore). The industry processes an estimated 250-300 billion customer interactions annually worldwide.

The AI-Augmented Cost Model

AI fundamentally reshapes these economics. Based on data from Klarna, early deployments at Salesforce Agentforce customers, and several undisclosed enterprise case studies, the emerging cost structure looks dramatically different:

AI-handled interactions (no human involvement):

• Cost per interaction: $0.05-$0.25 (API compute + infrastructure)
• Resolution time: 1-3 minutes average
• Available 24/7 with no capacity constraints
• Quality consistency: high (no bad days, no Monday morning slumps)

AI-assisted interactions (human agent with AI copilot):

• Cost per interaction: $2.00-$4.00 (reduced handling time + AI summarization)
• Resolution time: 4-7 minutes average (down from 8-12 minutes)
• Agent can handle 30-50% more interactions per hour
• Quality improvement: AI provides real-time policy lookup and response suggestions

Human-only interactions (complex escalations):

• Cost per interaction: $8.00-$25.00 (higher skill agents, longer duration)
• Resolution time: 15-45 minutes
• Requires experienced, well-compensated agents
• These represent 5-15% of total volume but 30-50% of remaining cost

The blended cost per interaction in an optimized AI-first operation drops to $0.50-$1.50 — an 70-85% reduction from the traditional model. For a company processing 50 million interactions annually, this translates to $150-$300 million in potential savings.

The Quality Paradox: When AI Outperforms Humans

One of the most counterintuitive findings in AI customer service deployment is that AI systems frequently score higher on customer satisfaction than human agents. This seems paradoxical — customers generally report preferring to speak with humans, yet rate AI interactions more favorably. The data, however, is increasingly consistent across deployments.

Klarna reported no statistically significant difference in CSAT between AI and human agents. But other deployments have shown AI outperformance:

• A January 2026 study published by the National Bureau of Economic Research (NBER), building on earlier work by Erik Brynjolfsson and colleagues at Stanford, found that customer service agents using AI assistance saw a 14% improvement in customer sentiment scores, with the largest gains among the lowest-performing agents.
• Zendesk's 2025 AI Impact Report found that AI-resolved tickets received CSAT scores 3-8 percentage points higher than human-resolved tickets for Tier 1 issues.
• An internal study at a major U.S. telecommunications company (disclosed at a 2025 industry conference but not publicly named) found that AI chat interactions scored 12% higher on Net Promoter Score than human chat interactions for billing and account inquiries.

The quality paradox has several explanations:

Consistency: Human agents have variance. They have bad days, they get frustrated with difficult customers, they forget policy details, they make errors under time pressure. AI systems deliver the same quality at 3 AM on a holiday weekend as they do at 10 AM on a Tuesday. For routine interactions, consistency is more valuable than occasional brilliance.

Speed: Customers value their time above almost everything else. An AI that resolves an issue in 90 seconds generates more satisfaction than a human who resolves the same issue in 8 minutes, even if the human interaction is warmer and more empathetic. The NBER research found that resolution speed was the single strongest predictor of customer satisfaction, outweighing empathy, personalization, and agent knowledge.

No hold times: AI systems eliminate queue wait times entirely. The psychological cost of waiting on hold — which averages 13 minutes globally for voice calls, according to Talkdesk's 2025 benchmark data — creates a negative satisfaction baseline that human agents must overcome before the interaction even begins.

Information accuracy: AI systems with proper data integration provide accurate, policy-compliant answers 95%+ of the time. Human agents, working from memory and training materials that may be outdated, typically achieve 82-88% accuracy on policy-related questions, according to quality assurance industry benchmarks.

This does not mean AI is universally superior. The paradox breaks down in specific scenarios: emotionally charged situations (bereavement-related account changes, fraud victims, service failures causing significant harm), complex multi-issue interactions, and situations requiring creative problem-solving or exception-making authority. These scenarios represent a minority of total volume but are disproportionately important for brand perception and customer retention.

Tier 1/2/3 Support: A Displacement Timeline

Customer service operations typically organize support into tiers based on complexity. The displacement trajectory differs dramatically across these tiers.

Tier 1: Basic Inquiries (60-70% of volume)

Current AI capability: High. Tier 1 interactions — password resets, order tracking, billing inquiries, FAQ responses, simple returns — are well within the capability of current LLM-based systems. These interactions follow predictable patterns, require access to structured data, and have clear resolution criteria.

Displacement timeline: Already underway. By the end of 2026, we estimate that 50-60% of global Tier 1 interactions will be handled by AI without human involvement, up from approximately 25-30% today. By 2028, this figure will reach 75-85%.

Surviving human role: Quality auditing and exception handling. A small team of humans will monitor AI performance, handle the 5-10% of Tier 1 interactions that AI fails to resolve, and continuously improve AI training data.

Tier 2: Moderate Complexity (20-30% of volume)

Current AI capability: Moderate and improving rapidly. Tier 2 interactions — disputed charges, product troubleshooting, service modifications, complaint resolution — require more contextual understanding, access to multiple systems, and occasionally judgment about policy exceptions. Current AI systems handle approximately 30-40% of Tier 2 interactions successfully when given proper system access and policy guidelines.

Displacement timeline: Lagging Tier 1 by 18-24 months. By the end of 2027, we estimate 40-50% of Tier 2 interactions will be AI-handled. By 2029, 60-70%. The constraint is less about language capability and more about system integration and exception authority.

Surviving human role: Exception authority agents who can make judgment calls outside standard policy, particularly when customer retention economics justify concessions. These roles will require higher skill and pay than current Tier 2 agents.

Tier 3: Complex and Specialized (5-10% of volume)

Current AI capability: Low to moderate. Tier 3 interactions — regulatory complaints, legal escalations, technical debugging of complex systems, VIP/enterprise account management — require deep domain expertise, multi-step problem solving, and often cross-functional coordination within the organization.

Displacement timeline: Slow and partial. By 2028, AI will assist Tier 3 agents (summarizing case history, suggesting resolutions, drafting communications) but will not autonomously handle these interactions. By 2030, AI may handle 20-30% of current Tier 3 volume, primarily cases that were previously misclassified and are actually solvable with better data access.

Surviving human role: This tier is the most durable. Complex escalation specialists, enterprise relationship managers, and regulatory compliance agents will remain human-staffed for the foreseeable future. However, AI augmentation will allow fewer humans to handle the same volume — expect 30-40% headcount reduction even in Tier 3 by 2029.

The BPO Industry Reckoning

The Business Process Outsourcing industry built a $350+ billion global enterprise on a simple value proposition: labor arbitrage. Western companies could reduce customer service costs by 40-60% by moving operations to the Philippines, India, Colombia, South Africa, and Eastern Europe, where English-speaking workers earned a fraction of U.S. or European wages. The industry's three largest public players — Teleperformance (€8.3 billion revenue), Concentrix ($9.6 billion revenue), and TaskUs ($900 million revenue) — collectively employ over 600,000 people, overwhelmingly in offshore locations.

AI does not just threaten this model. It renders the core arbitrage obsolete. When an AI system costs $0.10 per interaction regardless of language or geography, the difference between a $4 offshore agent-hour and a $35 onshore agent-hour becomes irrelevant. The entire cost curve that justified the BPO industry's existence flattens to near zero for routine interactions.

The market has noticed. Teleperformance's stock declined approximately 40% from its 2023 peak through Q1 2026. Concentrix shares followed a similar trajectory. TaskUs, with higher exposure to digital-native clients (who adopt AI fastest), has seen the steepest decline among publicly traded BPO companies.

Strategic Responses

The major BPO players are pursuing three strategies, with varying degrees of conviction:

1. Becoming AI-first service providers: Teleperformance has invested heavily in its "TP Interact" AI platform and acquired several AI startups. The strategy is to transition from selling agent-hours to selling AI-managed customer experience outcomes. The challenge: this business model generates far less revenue per client, even if margins are higher. A $50 million annual BPO contract that gets automated might become a $10 million AI platform contract — better margins but dramatically less topline.

2. Moving up the complexity ladder: Concentrix has repositioned toward higher-value services — consulting, CX design, technology integration — that are harder to automate. This is a sound strategy but limited by the talent pool: the skills required for CX consulting are fundamentally different from those of a contact center agent, and retraining at scale has proven difficult.

3. Offering hybrid human-AI models: TaskUs and others are positioning "AI-augmented agents" as a middle ground — human agents equipped with AI copilots who can handle higher volumes at lower cost. This extends the runway for the traditional model but does not change the long-term trajectory. As AI capabilities improve, the "human in the loop" becomes less necessary for an expanding set of interactions.

The most likely outcome for the BPO industry is significant consolidation. The top 3-5 players will survive by transforming into technology-enabled service companies. Smaller BPO operators, particularly those with limited technology investment, will see contract values decline sharply as clients bring AI-automated interactions in-house or shift to lower-cost AI platform providers.

For investors evaluating this sector, the key metric is revenue per interaction rather than total revenue. BPO companies that maintain or grow revenue per interaction — by moving to higher-value, harder-to-automate work — will survive the transition. Those whose revenue per interaction declines in lockstep with AI cost reductions will not.

Which Roles Survive?

Not all customer service roles face equal displacement risk. Based on our analysis of capability trajectories and organizational needs, we identify five categories of roles that will persist through 2030 and beyond:

1. Complex Escalation Specialists

When a customer's house floods because of a faulty appliance, when a billing error has cascaded across six months of statements, when a product defect has caused physical harm — these situations require human judgment, empathy, and authority to make decisions outside standard policy. AI systems can assist (summarizing the case, calculating financial impact, suggesting resolutions), but the decision-making and emotional labor remain human tasks. These roles will command premium compensation as organizations recognize they are the last line of defense for brand reputation.

2. Enterprise and VIP Relationship Managers

High-value accounts — enterprise clients, wealth management customers, VIP loyalty program members — expect dedicated human contact. The economics support it: a single enterprise client generating $500,000 in annual revenue justifies a $100,000 relationship manager. These roles will evolve to incorporate AI-generated insights and recommendations but will remain fundamentally human-to-human relationships.

3. Emotional and Sensitive Situations

Bereavement, fraud victimization, disability accommodations, crisis situations — these interactions require genuine human empathy and the ability to adapt communication style to the emotional state of the caller. While AI systems are improving at detecting emotional cues, the ethical and reputational risks of handling sensitive situations with AI remain prohibitive. Companies that automate bereavement calls will face severe backlash.

4. AI Trainers and Quality Engineers

A new category of role is emerging: specialists who train, evaluate, and improve AI customer service systems. These roles require a combination of customer service experience (understanding real interaction dynamics), data analysis skills (evaluating AI performance metrics), and prompt engineering or fine-tuning capability. We estimate that for every 50-100 AI-displaced agent roles, 3-5 AI training and quality roles are created.

5. Regulatory and Compliance Agents

In regulated industries — financial services, healthcare, insurance, utilities — certain customer interactions must be handled by licensed or certified humans as a matter of law. Insurance claims adjusters, financial complaint handlers, and healthcare patient advocates will remain human-staffed even when AI could technically perform the work, because regulatory frameworks require human accountability.

Cost Savings and Handling Time: The Metrics That Matter

Across deployments we have analyzed, the following metrics have emerged as representative of AI's impact on customer service operations:

Resolution time reduction: 60-80% for Tier 1 interactions. Klarna's drop from 11 minutes to under 2 minutes (82% reduction) is at the high end but not anomalous. The primary driver is elimination of agent lookup time — AI systems access customer data and policy information instantaneously.

First contact resolution improvement: 15-30% increase. AI systems do not forget to check all relevant systems or fail to mention applicable policies. The consistency advantage translates directly into fewer repeat contacts.

Cost per interaction reduction: 70-90% for AI-handled interactions relative to human-handled. The cost of an API call plus infrastructure overhead is typically $0.05-$0.25, compared to $3.00-$7.00 for a human-handled interaction.

Agent productivity improvement (for remaining human agents with AI copilot): 25-40%. AI copilots that summarize customer history, suggest responses, auto-fill forms, and handle post-interaction documentation allow human agents to focus on the conversation rather than administrative tasks.

Quality score improvement: 5-15% improvement in CSAT/NPS for AI-handled Tier 1 interactions, driven primarily by speed and consistency. This advantage narrows for Tier 2 and reverses for emotionally complex Tier 3 interactions.

Training cost reduction: 40-60%. AI copilots reduce the knowledge burden on new agents, allowing them to reach competency faster. When AI handles Tier 1 entirely, the training pipeline shrinks proportionally.

These metrics paint a picture of an industry in the early stages of a structural transformation. The question is not whether AI will reshape customer service — that is already happening. The question is how quickly the transformation propagates from early adopters like Klarna to the broader market, and what the transition means for the millions of workers currently employed in contact centers worldwide.

The Broader Context: Adjacent Displacement Dynamics

The customer service displacement we describe here does not exist in isolation. It is part of a broader pattern of AI capability expansion that is reshaping knowledge work across sectors. Several related dynamics deserve attention:

First, the distinction between genuine AI deployment and performative claims is critical. Many companies announce AI customer service initiatives that amount to little more than improved IVR systems or basic chatbots with LLM wrappers. Investors and analysts should apply rigorous scrutiny to distinguish real displacement events from AI washing.

Second, customer service has historically been one of the most offshored functions in the global economy. The interaction between AI displacement and offshoring creates a compounding effect: countries that built economic strategies around BPO employment — particularly the Philippines, where contact centers employ approximately 1.4 million people and generate 7-8% of GDP — face a structural challenge that goes beyond industry restructuring. For a deeper analysis of this dynamic, see our offshoring multiplier report.

Third, customer service displacement is a leading indicator for broader white-collar automation. The skills that make Tier 1 and Tier 2 agents vulnerable — reading comprehension, information retrieval, script following, data entry — are the same skills that underpin large swaths of administrative, clerical, and entry-level professional work. The displacement patterns we observe in contact centers today will propagate to other functions with a 12-24 month lag. Our sector exposure map provides a comprehensive view of which industries and functions face the greatest near-term risk.

Key Takeaways

• Klarna's AI deployment is a template, not an anomaly. The company demonstrated that a well-implemented AI system can handle two-thirds of customer interactions at higher speed, comparable quality, and a fraction of the cost. Every company with a significant customer service operation is now evaluating similar deployments.

• The economics are overwhelming. A 70-90% reduction in cost per interaction for Tier 1 support, combined with quality improvements and 24/7 availability, creates a competitive imperative. Companies that do not deploy AI customer service will be at a structural cost disadvantage within 24 months.

• The quality paradox favors AI. For routine interactions, AI systems already match or exceed human performance on customer satisfaction metrics. The consistency, speed, and accuracy advantages outweigh the empathy and flexibility advantages of human agents for the majority of interaction types.

• Displacement follows the tier structure. Tier 1 (basic inquiries) is being automated now. Tier 2 (moderate complexity) follows 18-24 months behind. Tier 3 (complex/specialized) will be augmented but not fully automated through at least 2030.

• The BPO industry faces existential restructuring. The labor arbitrage model that built a $350 billion industry becomes obsolete when AI costs $0.10 per interaction regardless of geography. The top players will survive by transforming into technology companies; smaller operators will consolidate or fail.

• Five categories of human roles will persist: complex escalation specialists, enterprise relationship managers, emotional/sensitive situation handlers, AI trainers and quality engineers, and regulatory compliance agents. These roles will command premium compensation but will represent a fraction of current industry headcount.

• The global employment impact is significant. With 17 million contact center workers worldwide and an estimated 40-60% of current roles vulnerable to automation by 2029, the customer service industry represents one of the largest single-sector displacement events in the AI era. The transition will be managed through attrition rather than mass layoffs, but the destination is the same.