The AI Implementation Failure That Looked Like a $60 Million Win
Created on 2026-02-26 12:03
Published on 2026-02-26 12:07
My sales team resigned. Colleagues avoided me.
I had optimized ruthlessly for the metric that got measured and destroyed everything that actually mattered. Relationships. Trust. The willingness of people to follow me into the next quarter.
The year was 2001. I was a GM at Electrolux. Star on the spreadsheet. Cancer in the hallway.
My Chairman, Gunnar Broberg sat me down and said nine words just before he fired me: “You are brilliant, but you are not ready for management.”
I thought about Broberg earlier this year when I read what happened at Klarna. Because Klarna’s AI agent made my exact mistake. Only faster, at greater scale, and across 23 countries simultaneously.
The $60 Million Success Story
In early 2024, Klarna rolled out an AI customer service agent. The numbers were staggering. In its first month, the agent handled 2.3 million conversations across 23 markets in 35 languages. Average resolution time dropped from 11 minutes to two. By January 2025, Klarna reported the agent was doing the work of 853 full-time employees and had saved the company $60 million.
On every dashboard, this was a triumph.
Then customers started pushing back. Generic answers. Robotic tone. No ability to handle anything requiring judgment or empathy. By mid-2025, CEO Sebastian Siemiatkowski told Bloomberg that while cost had been “a predominant evaluation factor,” the result was “lower quality.”
Klarna began frantically rehiring the human agents it had let go.
Most people tell this story as evidence that AI can’t handle nuance. That’s a comfortable reading. It lets you believe the technology isn’t ready and you have time to figure this out.
Here’s the reading that should concern you more.
The AI was extraordinarily good at what it was told to do. It resolved tickets fast. It cut costs dramatically. It scaled with impressive consistency. The agent didn’t fail. The agent succeeded brilliantly at the wrong objective.
Klarna’s organizational purpose wasn’t “resolve tickets fast.” It was “build lasting customer relationships that drive lifetime value in a competitive fintech market.”
Those are profoundly different goals. They require profoundly different decisions at the point of customer interaction. And nobody encoded that difference into the system.
The Acceleration Trap
There is a pattern here that extends far beyond one Swedish fintech company. I’ve watched it destroy value at every scale across 25 years in APAC, the UK, the US, and Australia.
AI doesn’t change your organization. It accelerates whatever your organization already is.
If your customer service operation genuinely values relationships, AI will help you build deeper ones at scale. If your operation secretly values cost reduction above everything else, AI will cut costs so efficiently that the damage compounds before anyone thinks to check whether customers are still there.
Klarna’s AI agent did precisely what the organization’s incentive structure rewarded. The deeper problem was that Klarna’s stated values and its operational values were not the same thing. The AI found the gap between them and drove through it at speed.
MIT’s Project NANDA research, published in July 2025, tracked over 300 AI implementations and found that 95% of organizations are getting zero return from their AI investments. The critical finding: “This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.”
Approach. Not technology. Not budget. Not which model you picked.
BCG’s research corroborates this from a different angle: 74% of organizations fail to achieve or scale AI value, and 70% of the challenges are people and process, not technology.
Klarna had a frontier model. It had budget. It had global scale. It still became a cautionary tale. Because approach includes something most organizations have never had to articulate: what does the organization actually want, as opposed to what can it most easily measure?
The Knowledge That Walked Out the Door
A customer service agent with five years at a company carries something no prompt can replicate.
She knows when to bend a policy. She knows when to spend three extra minutes because the customer’s tone says they’re about to churn. She knows when efficiency is the right move and when generosity is. She knows this because she absorbed the company’s real values over years. Not the values on the website. The values encoded in the decisions managers make every day, in the stories veterans tell new hires, in the unwritten rules about which metrics leadership actually cares about when things get tight.
The 700 human agents Klarna let go carried all of that institutional knowledge in their heads. When they walked out, the knowledge walked with them. It had never been documented. Nobody thought it needed to be.
Humans just knew.
The AI didn’t. It had a prompt. It had access to account histories and interaction data. What it lacked was the accumulated judgment about what matters more when two legitimate values conflict: speed or care, policy or relationship, efficiency or generosity.
This is the part that should worry leaders across APAC in particular.
In markets where business is fundamentally relationship-first, where trust precedes transactions, where a customer’s decision to stay is personal as much as rational, the cost of deploying AI that prioritizes speed over relationship is amplified. And the feedback mechanism is different.
Klarna discovered its problem through public backlash. Customers complained to media. The CEO was forced onto the record. There was a visible correction.
In many Southeast Asian markets, that correction doesn’t happen visibly. Customers don’t go to Bloomberg. They leave quietly. No public backlash. No CEO admission. Just a slow, silent erosion of the relationships your business depends on. By the time you see it in the numbers, the damage has been compounding for quarters.
Why This Is About to Get Worse
Klarna at least discovered the problem within months. The agent was handling real-time customer conversations, and dissatisfaction surfaced quickly.
Now consider what’s coming. AI agents that run autonomously for weeks. Soon, months. Making thousands of decisions without direct human oversight. Where the drift from organizational purpose is gradual enough that no single interaction triggers an alarm, but the cumulative effect reshapes your customer relationships, your market position, your brand.
Deloitte’s 2026 State of AI in the Enterprise report surveyed over 3,000 leaders across 24 countries. The findings should give every executive pause: 84% of companies have not redesigned jobs around AI capabilities, and only 21% have what Deloitte considers a mature model for agent governance.
Four out of five organizations deploying AI agents right now have no systematic way to ensure those agents are pursuing the objectives that actually matter to the business. They’re measuring output. They’re hoping it aligns with purpose.
Klarna proved what happens when it doesn’t.
Three Questions Before You Deploy
If you’re running a mid-market company in APAC and you’re considering AI agents, or you’ve already deployed them, three questions matter more than which model you’re running.
1. What does this agent need to want?
Not what does it need to do. What does it need to want. There’s a critical difference. Klarna’s agent needed to want customer retention, not ticket resolution. In a requirements document, those look similar. In practice, they produce vastly different behaviors.
For every AI agent you deploy, articulate the organizational purpose it serves. Not the task. The purpose. Then ask yourself: if this agent performs its task perfectly, will that advance my organizational purpose or undermine it?
If you can’t answer that clearly, you’re not ready to deploy.
2. What does this agent need to know that “humans just know”?
Somewhere in your organization, experienced people carry judgment that has never been written down. The unwritten rules. The contextual awareness. The instinct for when the standard process is wrong for this particular situation.
Identify it. Document it. Not as a 200-page operations manual, but as decision principles with clear boundaries. When a customer request conflicts with policy, here is the resolution logic. When signals suggest frustration, here is what matters more than closing the ticket quickly.
This is difficult work. It forces senior people to make explicit what they’ve never had to articulate. But if you skip it, your AI agent will do what Klarna’s did: pursue the measurable thing and ignore the thing that matters.
3. How will you know when the agent drifts?
Klarna’s feedback mechanism was customer complaints reported by international media. That is the most expensive way to discover misalignment.
Build feedback loops that detect drift before your customers do. Define what “aligned” looks like in terms that go beyond output volume. Customer sentiment trends. Escalation patterns. The ratio of standard resolutions to exceptions that warranted human judgment. The signals that tell you whether the agent is serving your organizational purpose or merely completing tasks.
The Lesson Broberg Taught Me
When Gunnar Broberg fired me from Electrolux, he wasn’t punishing my performance. My numbers were extraordinary. He was teaching me that performance disconnected from purpose is destruction.
Klarna learned the same lesson 25 years later. At AI speed. At AI scale. The $60 million in savings didn’t come close to offsetting the cost of becoming a public cautionary tale, the expense of frantically rehiring, and the customer trust that is far harder to rebuild than it was to lose.
AI readiness is not a technology question. It’s a leadership question about whether your organization’s stated values and its operational values are the same thing. Because AI will find the gap between them. And it will not be diplomatic about it.
If you want to find that gap before an AI agent does, the AIR APAC Readiness Scorecard takes 15 minutes. Six dimensions. A score that tells you what you already suspect but haven’t been able to quantify.