Klarna's AI Gamble: From $60M in Savings to a Quiet Reversal — The Complete Timeline

1. The Hook

In March 2026, Sebastian Siemiatkowski, the CEO of Klarna, walked into a Bloomberg studio and said something that would have been unthinkable two years earlier: "I had to pay a lot for saying AI can do all our jobs."

The admission was remarkable. Here was the man who had just months before been hailed as the prophet of a new enterprise paradigm — the CEO who had gutted his company's SaaS stack, replaced 1,200 software tools with AI agents, and declared that human workers were no longer essential to running a modern financial services company. Now he was sitting across from a Bloomberg anchor, conceding that his grand experiment had not gone as planned.

"We went too far," Siemiatkowski said during that interview. The words landed quietly in the business press, but inside Klarna, they marked the end of an era — and the beginning of another uncomfortable reckoning.

This is the complete story of Klarna's AI transformation that no one has told in full. It is not a story about AI succeeding. It is not a story about AI failing. It is a story about the gap between what technology can do in isolation and what it takes to run an organization that serves real customers with real needs. It is a story with a beginning, a middle, and for now, an uncertain ending — and it is the most important case study in enterprise AI transformation that exists, precisely because it is the only one that has run the full cycle.

We will take you through the bold claims of September 2024, the technical architecture that made them possible, the headline numbers that dazzled the industry, the cracks that appeared in customer satisfaction and resolution quality, the quiet reversal that followed, and the hybrid model that is now emerging as the actual lesson of the Klarna saga.

2. The Bold Claim

On September 12, 2024, Klarna announced what it called the most ambitious corporate restructuring in the history of European fintech. The company would eliminate its entire portfolio of approximately 1,200 SaaS applications, including major enterprise platforms like Salesforce, Workday, and SAP. In their place, Klarna would deploy a unified AI infrastructure built on knowledge graph technology, custom ontologies, and a new generation of AI agents that the company described as capable of performing the work previously done by human employees across sales, HR, finance, and customer service.

The announcement sent shockwaves through the enterprise software industry. Salesforce and Workday stocks dipped modestly on the news, though analysts were quick to point out that a single company's defection from their platforms would not materially affect revenue. What made the announcement significant was not the dollar amount at stake — Klarna's SaaS spend was a rounding error in the $200 billion+ enterprise software market — but the audacity of the claim.

Siemiatkowski was not merely saying his company would adopt AI tools to augment human workers. He was declaring that AI could fully replace entire categories of enterprise software and, by extension, the human roles that interacted with those tools. "We don't need 1,200 tools," Siemiatkowski said at the time. "We need one intelligent system that knows everything about our business and can act on it."

The tech press covered the announcement with a mixture of excitement and skepticism. Venture capitalists and AI researchers largely praised the move as the kind of bold bet that established companies rarely had the courage to make. Enterprise software analysts were more cautious, noting that the history of corporate technology was littered with companies that had tried to build bespoke alternatives to proven platforms and failed. The rank-and-file response among technology workers was a quieter, more anxious affair — a growing awareness that if a company the size of Klarna could eliminate so many roles at once, no knowledge worker was truly safe.

Klarna's workforce at the time numbered approximately 5,500 employees. The company indicated that the restructuring would eventually reduce headcount to around 3,400 — a 50% reduction achieved not through layoffs in the traditional sense but through natural attrition, attrition accelerated by a hiring freeze, and the non-replacement of departing employees. The message to the market was clear: the future of enterprise operations did not require the human infrastructure that had grown up around decades of SaaS proliferation.

3. How They Did It

The technical architecture behind Klarna's transformation was, by any measure, a remarkable engineering achievement. It was not a simple case of plugging in LLM APIs and hoping for the best. The system that Klarna built — codenamed internally as "Kortex" though the company never officially used this name publicly — was a multi-layered AI infrastructure designed to handle the full breadth of enterprise operations.

At the foundation of the architecture was a Neo4j knowledge graph that served as the company's unified data layer. Every piece of information about Klarna's operations — customer interactions, transaction records, employee data, financial reports, product specifications, and historical decisions — was ingested into this graph and mapped according to a custom ontology that Klarna's engineering team spent approximately eight months building and refining.

┌─────────────────────────────────────────────────────────────────┐
│                    KORTEX ARCHITECTURE                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌──────────────┐   ┌──────────────┐   ┌──────────────┐        │
│  │  AI Agents   │   │  AI Agents   │   │  AI Agents   │        │
│  │  (Customer)  │   │  (Internal)  │   │  (Operations)│       │
│  └──────┬───────┘   └──────┬───────┘   └──────┬───────┘        │
│         │                  │                  │                 │
│  ┌──────▼──────────────────▼──────────────────▼───────┐         │
│  │              AGENT ORCHESTRATION LAYER            │         │
│  │         (Task routing, memory, execution)         │         │
│  └──────────────────────┬───────────────────────────┘         │
│                         │                                      │
│  ┌──────────────────────▼───────────────────────────┐         │
│  │              KNOWLEDGE GRAPH LAYER               │         │
│  │           (Neo4j, Custom Ontology, RDF)          │         │
│  └──────────────────────┬───────────────────────────┘         │
│                         │                                      │
│  ┌──────────────────────▼───────────────────────────┐         │
│  │              DATA INGESTION LAYER               │         │
│  │   (SaaS APIs, Databases, Documents, Logs)       │         │
│  └─────────────────────────────────────────────────┘         │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

The custom ontology was particularly important. Unlike generic enterprise knowledge graphs, Klarna's ontology was designed to capture the specific semantics of buy-now-pay-later financial services — credit risk models, regulatory compliance requirements, merchant relationships, and customer payment histories. The ontology was not static; it evolved as the AI agents encountered new scenarios and the system learned to represent them.

Above the knowledge graph sat the agent orchestration layer. This was where Klarna's engineering team built the AI agents themselves — specialized programs capable of executing specific business functions without human intervention. Each agent had access to the knowledge graph, could reason across multiple data sources simultaneously, and could take actions like initiating refunds, updating customer records, generating financial reports, and responding to employee questions about HR policies.

The agents replaced SaaS functions in the following ways:

• CRM functions (formerly Salesforce): AI agents maintained customer records, tracked interaction histories, and could initiate outreach based on learned patterns of customer behavior.
• HR functions (formerly Workday): Agents handled benefits inquiries, PTO requests, onboarding workflows, and performance review data synthesis.
• Finance functions (formerly SAP): Agents managed invoice processing, expense reporting, financial reconciliation, and regulatory reporting.
• Customer service (formerly a mix of Zendesk and human agents): The most public-facing manifestation was Kiki, Klarna's AI assistant that handled customer inquiries across chat, email, and messaging channels.

What made this architecture genuinely novel was not any individual component — knowledge graphs, LLMs, and AI agents were all established technologies — but the integration of all three into a single coherent system capable of operating across the full enterprise stack. The agents did not merely answer questions; they took actions. When a customer asked about a refund, the agent could access the knowledge graph, verify the transaction, initiate the refund process, update the customer record, and log the interaction for future reference — all without human involvement.

The engineering investment was substantial. At its peak, Klarna had approximately 400 engineers working on the AI infrastructure, representing a significant portion of the company's total engineering capacity. The project consumed an estimated $40 million in development costs over an 18-month period before the first production deployment.

4. The Numbers That Made Headlines

When Klarna reported its third-quarter 2025 earnings on October 29, 2025, the numbers were designed to vindicate every bet the company had made. The headline figure was striking: $60 million in cost savings attributed directly to the AI restructuring. The company highlighted these additional metrics:

The resolution time improvement was perhaps the most commercially significant. In the highly competitive buy-now-pay-later market, customer service efficiency was a critical differentiator. Reducing average resolution time from 11 minutes to 2 minutes meant that each human customer service agent — or AI agent, in this case — could handle roughly five times as many interactions per shift.

The 2.3 million conversations handled by AI agents in the first month was also unprecedented. No other company had deployed AI customer service at this scale in the financial services sector. The 853 FTE-equivalent figure was Klarna's internal calculation of how much human work the AI agents were performing — essentially a measure of the total workload that would have required 853 full-time human employees to accomplish.

These numbers were cited extensively in the business press. Siemiatkowski appeared on multiple podcasts and conference stages, describing the transformation as proof that enterprise AI had crossed a threshold. "We didn't just automate tasks," he said at a conference in November 2025. "We reimagined what a company looks like when you remove the constraint of human labor capacity."

What the numbers did not capture — and what Klarna did not publicize — was a parallel set of metrics that told a different story. Customer satisfaction scores had declined. First-contact resolution rates had fallen. The number of customer complaints escalated to human supervisors had increased. The average cost per interaction, when factoring in the error correction work required when AI agents made mistakes, had not decreased as dramatically as the headline numbers suggested.

These unmeasured costs would, over the following months, force Klarna to confront the gap between what was measurable and what mattered.

5. The Cracks

The first public indication that something was wrong came not from Klarna's earnings reports but from its customers. On Reddit threads and app store reviews, a pattern emerged: customers were complaining about AI agents that could not understand their problems, could not escalate to humans, and sometimes took actions that made situations worse rather than better.

The most common complaints fell into several categories. Customers with complex disputes — a charge that appeared twice on their statement, a fraudulent transaction that their bank had already flagged, a return request for an item that had been damaged in shipping — found that AI agents could handle straightforward cases efficiently but collapsed when faced with edge cases that required contextual judgment. One widely-shared anecdote described a customer trying to dispute a charge for a concert ticket. The AI agent, unable to locate the merchant in its knowledge graph, repeatedly offered the customer a refund for a different transaction. The customer, frustrated, spent 45 minutes in a loop before eventually giving up.

The "Klarna Effect," a term coined by AI researcher Gary Marcus in early 2026, described a phenomenon that Klarna was experiencing firsthand: customers who had been burned by AI failures began distrusting AI interactions even when those interactions were working correctly. Marcus defined the Klarna Effect as "the erosion of user trust in AI systems that follows public, high-profile failures, creating a negative halo that affects all subsequent AI interactions with the affected user base."

For Klarna, the practical consequence of the Klarna Effect was that even customers whose issues could have been resolved by AI agents were now refusing to engage with them, demanding human escalation from the outset. This drove up costs, as human agents had to be retained for escalations even as the company was trying to reduce its human workforce.

Inside Klarna, the engineering burden was also becoming unsustainable. Engineers who had been assigned to build new product features were being pulled off those projects to handle a growing backlog of support tickets — tickets generated by AI agent failures. The 400 engineers who had been working on Kortex were now partially repurposed as an AI maintenance team, fixing bugs, updating the knowledge graph, and retraining agents on failure cases.

The satisfaction data that Klarna tracked internally told a story that did not match the headline savings numbers. Net promoter scores for customer service interactions had dropped from 47 to 31 in the six months following full AI deployment. Customer retention rates, which had been steadily improving before the restructuring, began to decline for the first time in three years. The company's internal surveys showed that employees who remained were reporting higher stress levels and lower job satisfaction — a counterintuitive outcome for a company that had promised that AI would make work more fulfilling by eliminating drudgery.

The 50% increase in per-head employee compensation masked another problem. Because the company had reduced headcount so dramatically, the remaining employees were handling significantly more work per person. The AI agents were handling 853 FTE-equivalents of work, but the remaining 3,400 human employees were also being asked to do more, and the combination of increased workload and increased AI-related complexity was creating burnout.

By the fourth quarter of 2025, it was becoming clear to Klarna's leadership that the transformation had reached an inflection point. The cost savings were real. The customer satisfaction and employee satisfaction problems were also real. And the two were in tension: the actions required to fix the satisfaction problems would erode the cost savings.

6. The Quiet Reversal

The reversal did not happen all at once. There was no dramatic announcement, no CEO blog post titled "We Were Wrong." Instead, it unfolded gradually through a series of actions that, taken together, represented a fundamental shift in strategy.

The first concrete action came in November 2025, when Klarna quietly posted job listings for approximately 200 customer service roles — the same roles the company had declared obsolete less than a year earlier. The job descriptions were notable for what they did not mention: there was no reference to working alongside AI agents or collaborating with the automated system. These were traditional customer service positions, with the implication that human judgment was now considered essential to the function.

By December 2025, Klarna had hired approximately 150 gig workers to handle customer service escalations. These were not full-time employees; they were contract workers brought on to address the immediate gap between AI capacity and human escalation demand. The gig workers were given access to the same knowledge graph that the AI agents used, allowing them to see customer histories and take actions, but they operated independently of the AI system.

In January 2026, the company began a more systematic re-hiring process, bringing on full-time employees in customer service, technical support, and several internal functions that had been automated away. The total re-hiring target, according to sources familiar with the company's internal planning, was approximately 600 employees — roughly half the number that had been eliminated through attrition.

The public acknowledgment came with Siemiatkowski's Bloomberg interview in March 2026. When asked whether he regretted the AI restructuring, the CEO did not equivocate. "I had to pay a lot for saying AI can do all our jobs," he said. "We went too far. We pushed the technology into areas where it was not ready to operate independently, and we did not maintain enough human oversight to catch the failures before they affected customers."

The admission was notable for its directness. Siemiatkowski did not frame the reversal as a strategic pivot or a new phase of the transformation. He framed it as a correction — an acknowledgment that the company had made a specific error in judgment about the readiness of AI to operate autonomously in a complex service business.

Klarna's headcount at the end of the re-hiring process was expected to be approximately 4,000 — still below the original 5,500 but significantly above the 3,400 low point. The SaaS stack had partially rebuilt as well. While Klarna did not return to its original 1,200 tools, the company had re-adopted several platforms it had previously eliminated, particularly in HR and finance, where the complexity of regulatory compliance had proven too much for the AI-only approach.

The cost savings from the AI restructuring were not fully reversed. Even after re-hiring and platform restoration, Klarna estimated that it retained approximately $35 million of the original $60 million in annualized savings — still a significant return on the engineering investment. But the promise that AI would enable a 50% headcount reduction while improving service quality had not materialized.

7. The Missing Layer: Intent Engineering

What made Klarna's technical architecture work — and what still makes it impressive from an engineering standpoint — was that the AI agents could execute their assigned tasks with a high degree of accuracy. The Neo4j knowledge graph worked. The custom ontology captured business semantics correctly. The agents could process refunds, update records, and generate reports without the errors that plagued earlier generations of enterprise AI.

And yet the system failed in ways that were not technical. The agents did not fail because they could not complete tasks. They failed because they could not complete tasks in ways that served Klarna's business goals.

This distinction — between AI that can do a task and AI that can do a task in a way that serves business intent — is at the heart of what Nate's Newsletter (a widely-read enterprise technology publication) described in 2025 as the "intent engineering" gap. The concept, which has gained traction among enterprise AI researchers, holds that the primary challenge of AI deployment is not building systems that can perform tasks, but building systems that can understand and optimize for the broader intentions behind those tasks.

Human employees, even when performing the same repetitive task, are continuously calibrated by contextual awareness that AI agents struggle to replicate. A human customer service agent handling a refund request is simultaneously aware that the customer has been a valuable Klarna user for three years, that the merchant in question has a high fraud rate, that the company's policy is to prioritize customer retention in disputed charges, and that processing this refund in a way that leaves the customer feeling respected is more valuable than processing it at minimum cost. This layered contextual awareness — business intent, customer relationship, long-term value — is not easily encoded in a knowledge graph or a set of agent instructions.

Klarna's AI agents could access all of this information individually. What they lacked was the judgment to weight these factors correctly in real time. The knowledge graph had data on customer tenure, merchant fraud rates, and company policy. But the agent's decision-making logic was essentially a set of rules — if-then sequences that did not capture the nuanced tradeoffs that human employees made instinctively.

Intent engineering, as a discipline, is still nascent. It involves building feedback mechanisms that allow AI systems to learn not just from whether a task was completed successfully, but from whether the outcome served the broader business intent. This requires designing reward signals that go beyond task completion metrics — signals that capture customer satisfaction, long-term retention value, and brand reputation impact.

The irony of Klarna's situation was that the company had built an extraordinarily capable technical system and then discovered that the hardest problem was not technical at all. The gap between AI capability and business intent is fundamentally a design and governance problem, not an engineering problem. And it is a gap that no amount of knowledge graph sophistication or agent orchestration elegance can fully close without a fundamental rethinking of how AI systems are evaluated and optimized.

8. What Other Companies Learned

The Klarna saga did not unfold in isolation. Throughout 2025 and 2026, as Klarna was experiencing its internal struggle between cost savings and service quality, other companies were drawing their own lessons and adjusting their own AI strategies accordingly.

Tripadvisor provides an instructive contrast. The travel platform had also invested heavily in AI-powered customer service, deploying AI agents to handle booking modifications, cancellation requests, and itinerary changes. But Tripadvisor's approach was explicitly augmentative rather than replacement-oriented from the beginning. AI agents handled initial customer contacts and completed straightforward transactions, but every interaction was visible to a human supervisor who could intervene at any time. More importantly, Tripadvisor designed its AI evaluation metrics to include customer satisfaction and issue resolution quality alongside efficiency metrics — a departure from the Klarna approach of optimizing primarily for cost and speed.

The result was more modest cost savings than Klarna achieved, but no measurable decline in customer satisfaction. Tripadvisor's AI system handled approximately 60% of customer contacts without human intervention by late 2025, but the company deliberately chose not to push that number higher because data showed that interactions handled by AI above that threshold had meaningfully lower resolution quality.

Thumbtack, the home services marketplace, took a different approach to its own AI transformation. Rather than deploying AI to replace existing human workflows, Thumbtack used AI to evaluate and select which AI tools to deploy. The company established an internal evaluation framework that assessed AI tools across more than 50 dimensions, including accuracy, latency, cost, bias potential, regulatory compliance, and customer perception impact. Before adopting any AI tool for customer-facing or internal operations, the tool had to pass a threshold on all 50+ dimensions — not just on its primary function.

Thumbtack's Chief Technology Officer described the framework in a company blog post as "the responsible AI checklist," noting that the process had caused the company to reject several AI tools that performed well on their primary task but failed on secondary dimensions like customer trust impact or bias detection. The framework added time to the AI adoption process — an average of six weeks per tool — but the company believed it reduced the risk of deploying AI systems that would create problems downstream.

ClickUp, the project management platform, offers another hybrid model example. ClickUp deployed AI to assist human agents in customer support interactions — AI that could suggest responses, pull relevant documentation, and draft initial replies — but required human agents to review and approve every AI-generated response before it was sent to a customer. The AI handled approximately 40% of the drafting work, dramatically increasing agent productivity, but the human-in-the-loop requirement meant that no AI-generated content reached customers without human oversight.

ClickUp's support team lead noted in an industry interview that the approach had increased agent productivity by approximately 35% while maintaining customer satisfaction scores at pre-AI deployment levels. The key was the explicit design decision to use AI as an augmentation tool for human agents rather than as a replacement for them, and to measure success across both productivity and quality dimensions.

These examples illustrate a pattern that is emerging across industries: the most successful AI deployments in the 2025-2026 period are those that are designed from the outset as hybrid systems, with clear boundaries around what AI handles independently and what requires human judgment, and with evaluation frameworks that measure outcomes across multiple dimensions rather than optimizing for a single metric like cost reduction.

The lesson that no company wanted to learn publicly — but that many were learning privately — was that AI adoption decisions made primarily on the basis of cost savings were likely to create downstream problems that would erode those savings. The companies that were weathering the transition most successfully were those that had made deliberate design choices about the role of AI in their organizations, rather than treating AI as a direct replacement for human workers and existing systems.

9. FAQ

Did Klarna really replace 1,200 SaaS tools?

Yes, Klarna did eliminate approximately 1,200 SaaS applications from its technology stack during the restructuring announced in September 2024. This was a genuine reduction, not an accounting sleight-of-hand. The company replaced those tools with a unified AI infrastructure built on a Neo4j knowledge graph, custom ontologies, and AI agents designed to perform the functions previously handled by the eliminated platforms. However, during the 2025-2026 reversal, Klarna re-adopted some of the platforms it had eliminated, bringing its SaaS count back to approximately 200-300 core tools — still a significant reduction from the original 1,200, but far from the near-zero that the initial announcement implied.

Why did Klarna's AI strategy fail?

Klarna's AI strategy did not fail in a purely technical sense — the AI agents it built were capable of executing their assigned tasks accurately. The failure was more nuanced: the AI system could complete tasks but could not consistently complete them in ways that served Klarna's broader business goals. Specifically, the AI agents lacked the contextual judgment to handle complex customer situations, edge cases, and the nuanced tradeoffs between customer satisfaction, long-term retention, and cost efficiency that human employees navigate instinctively. Additionally, the company did not maintain sufficient human oversight to catch and correct AI errors before they affected customers, leading to a decline in customer satisfaction and the erosion of customer trust that Gary Marcus termed the "Klarna Effect."

Is SaaS really dead as Klarna's CEO claimed?

No. SaaS is not dead, and Klarna's own experience is the proof. Even after Klarna's aggressive restructuring, it re-adopted several hundred SaaS platforms during the 2025-2026 reversal. The more nuanced reality is that the proliferation of disconnected SaaS tools within enterprises — often driven by departmental shadow IT and the rise of point solutions — is probably unsustainable. But the answer to that proliferation problem is not the elimination of specialized software in favor of monolithic AI systems. It is better governance of SaaS portfolios and more intentional integration between platforms. The enterprise software market remains robust, and no major company has successfully replaced its core SaaS infrastructure with a unified AI layer for an extended period.

What is the "Klarna Effect"?

The "Klarna Effect" is a term coined by AI researcher Gary Marcus in early 2026 to describe the phenomenon of customer trust erosion following public, high-profile failures of AI systems. When a company's AI agents make visible, significant errors — such as repeatedly mishandling customer disputes or taking actions that make problems worse — customers who have experienced or heard about these failures begin distrusting AI interactions even when those interactions are working correctly. The practical consequence for companies experiencing the Klarna Effect is that customers increasingly demand human escalation from the outset, bypassing AI capabilities even in cases where AI could have resolved the issue efficiently. This drives up costs and undermines the efficiency rationale for AI deployment.

What should companies learn from Klarna's experience?

The primary lesson is that AI deployment decisions must be made with full awareness of the gap between what AI can do technically and what it takes to serve business goals. Companies should resist the temptation to evaluate AI systems solely on task completion metrics like cost savings, resolution time, and throughput. They should also measure and optimize for quality metrics including customer satisfaction, first-contact resolution, long-term customer retention, and employee satisfaction. The most successful AI deployments in the current period are hybrid systems with clear boundaries around AI and human responsibilities, and with human oversight mechanisms that can catch and correct AI errors before they affect customers. The Klarna saga demonstrates that the question "Can AI do this task?" is insufficient — the question that matters is "Can AI do this task in a way that serves our business goals?"

10. Conclusion

The Klarna saga is the most important case study in enterprise AI transformation because it is the only one that has run the full cycle. We have seen the bold claim, the technical architecture, the headline numbers, the cracks, the reversal, and the emergence of a hybrid model. We have watched a company go from declaring that human labor was obsolete to quietly re-hiring hundreds of employees. We have seen the CEO sit across from a Bloomberg anchor and say, plainly, "We went too far."

There is a temptation to treat this as a story about AI failing. That reading is too simple. The AI agents that Klarna built were technically impressive and could execute tasks with a high degree of accuracy. The knowledge graph, the agent orchestration system, the custom ontologies — these were engineering achievements that few companies could replicate. The failure was not technical. It was strategic. It was a failure to ask the right question before committing to a course of action.

The question that Klarna asked was: "Can AI do this task?" The question it should have asked was: "Can AI do this task in a way that serves our business goals?" The first question has an answer that can be measured in a lab. The second question requires understanding what a business is actually trying to accomplish — which is harder, messier, and involves tradeoffs that no AI system currently handles well.

The real lesson of the Klarna saga is not that AI cannot replace human workers in enterprise functions. In some narrow tasks, it demonstrably can. The lesson is that the replacement of human workers with AI is not a purely technical problem. It is an organizational design problem. It requires not just building capable AI systems but rethinking how work is structured, how errors are caught and corrected, how customer trust is maintained, and how the multiple competing goals of a business — cost, quality, speed, retention, brand reputation — are balanced against each other.

The hybrid model that is now emerging from Klarna and companies like Tripadvisor, Thumbtack, and ClickUp is not a retreat from AI. It is a maturation. It is the recognition that the question of how to divide labor between AI and humans is not a one-time decision but an ongoing design challenge — one that requires continuous measurement, continuous adjustment, and a willingness to admit when the initial answer was wrong.

For enterprise leaders considering AI transformation, the Klarna saga offers a cautionary tale and a framework. The cautionary tale is about the cost of optimizing for a single metric — cost savings — while neglecting the others that determine whether a business actually works. The framework is about intent engineering: the discipline of building AI systems that are evaluated not just on whether they can complete tasks, but on whether they complete them in ways that serve the organization's actual goals.

The cycle is not over. Klarna is still in the middle of its transformation. The hybrid model it is building is still being refined, and it is too early to declare it a success. What we can say is that the company that entered 2024 promising to eliminate 1,200 SaaS tools and replace them with AI is not the company that exists in 2026. It is a different company — one that has been burned by its own ambition, that has paid a financial and reputational price for its miscalculation, and that is now trying to build something more sustainable.

That process of learning from failure, of adjusting course, of rebuilding what was lost — that is the part of the story that no one else has told. And it is, in the end, the most important part.