Klarna Fired 700 People and Had to Hire Them Back. Here's the Lesson for Your Business.

The Numbers Were Perfect. Then Everything Fell Apart.

In early 2024, Klarna deployed an AI agent to handle customer service. The results were immediate and extraordinary. In its first month, the agent handled 2.3 million conversations across 23 markets in 35 languages. Resolution time dropped from 11 minutes to 2 minutes. The CEO projected $40 million in annual savings and called it a landmark moment.

By mid-2025, Klarna was frantically rehiring the human agents it had just let go.

The CEO told Bloomberg that while cost was a "predominant evaluation factor," the result was lower quality. Customers were complaining about generic answers, robotic tone, and an agent that couldn't handle anything requiring judgment. The company that had declared victory on AI had to walk it back, publicly and expensively.

Here's the part that almost nobody talks about when they tell this story: the agent wasn't broken. It was extraordinarily good at exactly what it was designed to do. That's the problem.

The Agent Did Its Job. The Job Was Wrong.

Klarna's AI agent was given a clear goal: resolve customer tickets as fast as possible. Resolution time dropped from 11 minutes to 2. By that measure, it was a spectacular success.

But Klarna's actual goal — the one that makes a fintech company viable in a competitive market — wasn't "resolve tickets fast." It was "build lasting customer relationships that drive lifetime value." Those are profoundly different objectives, and they require profoundly different judgment at the moment of interaction.

A human agent who's worked at a company for five years knows things that never appear in a prompt. She knows when to bend a policy for a customer who's been loyal for three years. She knows when to spend an extra three minutes because something in the customer's tone says they're about to churn. She knows when efficiency is the right call and when generosity is. She absorbed this over years of watching how experienced managers handled situations, what leadership actually rewarded, which unwritten rules mattered.

The AI agent had none of this. It had a prompt. It resolved tickets fast. Mission accomplished.

This is what you might call the intent gap — the distance between the task you gave the agent and the actual goal your business needed to accomplish. The task was measurable. The goal was not. And in the absence of clearly encoded intent, the agent optimized for what it could measure.

84% of companies haven't redesigned their workflows around AI capabilities, according to Deloitte. Only 21% have a mature model for agent governance. And 74% report seeing no tangible value from AI at all. These aren't technology failures. They're intent failures. The models work. What's missing is the organizational infrastructure that connects AI capability to what the business actually needs to accomplish.

Why This Matters More for Small Businesses

Klarna had $40 million in projected savings to absorb the cost of getting this wrong. You don't.

When a small or mid-size business deploys an AI agent on customer-facing work without thinking through intent, the damage shows up faster and hurts more. A long-time customer who gets a robotic response to a complicated situation doesn't write a blog post about it. They just stop calling.

The good news is that small businesses have a structural advantage here. You can define your actual goals — not just the task, but the real intent behind it — because you know your customers. You know the difference between a new inquiry and a longtime client who needs to feel heard. You know which situations need human hands and which ones can run on autopilot.

The question is whether you encode that knowledge before you deploy an agent, or whether you find out the hard way what happened when you didn't. This is closely connected to the idea explored in You Already Have the Most Valuable Thing in AI — your domain knowledge is the ingredient that makes an agent work correctly, but only if you put it in.

Three Questions Before You Deploy Any AI on Customer-Facing Work

Before you point an agent at anything that touches your customers, answer these three questions. Not as a formality — actually think them through.

What does this agent actually need to achieve?

Not the surface task. The real goal. "Handle incoming inquiries" and "make customers feel confident enough to book" are completely different objectives that require completely different behavior. An agent that handles inquiries can close every conversation quickly. An agent that builds booking confidence knows when to slow down, offer more detail, or suggest a phone call. Write down your real goal before you write a single instruction to the agent.

What trade-offs is the agent allowed to make?

Every agent will face situations where two good things conflict. Speed versus thoroughness. Efficiency versus a longer conversation that builds trust. Policy compliance versus making an exception for a specific situation. These trade-offs don't resolve themselves. If you don't decide how you want the agent to handle them, it will decide on its own — and it will optimize for whatever you made easiest to measure.

Klarna made resolution speed easy to measure. Their agent resolved tickets fast. Decide which trade-offs reflect your values and make them explicit.

When does it stop and hand off to a human?

This is the question most businesses skip, and it's the most important one. There are situations your agent should never try to handle alone. A long-time customer who's clearly frustrated. A request that doesn't fit any of your standard scenarios. Two policies in conflict with no clear resolution. A situation where the customer's emotional state matters more than the factual answer.

Write down your escalation triggers before you launch. Not "use your judgment" — specific situations that send the conversation to a human immediately. The agent shouldn't decide in the moment whether it can handle something sensitive. You should decide in advance, and the agent should follow that structure.

What Good Intent Engineering Looks Like for a Real Business

You don't need a team of engineers to do this. You need to be honest about what your business is actually trying to accomplish.

A roofing contractor thinking through his estimate follow-up agent might write: "The goal isn't to confirm we sent the estimate. The goal is to help the homeowner feel confident enough to say yes or give us a fair shot to address their hesitation." That single sentence changes how the agent is configured, what it says, and when it escalates.

A physical therapy practice thinking through her new patient intake agent might write: "New patients are anxious. The goal isn't to collect their paperwork. The goal is to make them feel like they made the right choice before they've ever walked in the door." Same technology. Completely different outcome.

Seeing the Difference Side by Side

Here's what this looks like in practice. A home services company deploys an agent to follow up on every job with a review request. Version A (no intent engineering): the agent sends the same message to every client 24 hours after job completion. Version B (with intent engineering): the agent checks the job notes — if there was a complaint, delay, or unresolved issue flagged during the job, it routes to a human for a personal call before any review request goes out. Clients who had a clean experience get the automated message. Clients who had a rough experience get a call from the owner.

Version A generates some reviews and some angry responses. Version B generates more positive reviews and catches problems before they become public. The difference isn't the AI model or the platform. It's the 20 minutes the owner spent thinking through when the agent should stop and hand off.

This is the practical version of moving your AI from Level 2 to Level 3 — not just giving agents access to your systems, but giving them clear enough goals that they can make good decisions without constant supervision.

The Pattern Across Every Industry

Klarna's story isn't unique. It's a precise, well-documented version of the same mistake that shows up across every industry where AI agents have been deployed without careful intent engineering.

A healthcare system deploys a scheduling agent with the goal of filling empty appointment slots. It succeeds — but the slots get filled with patients who aren't a good fit for the available providers, creating no-shows and downstream problems. The intent was efficiency. The actual goal should have been appropriate matches, not filled calendars.

A law firm automates initial client intake to screen leads faster. The agent closes conversations quickly with prospects who are actually good clients, because the agent's goal was to qualify efficiently, not to make prospects feel heard at a moment when they're often anxious. Client conversion drops. The volume goes up. The close rate falls.

The pattern is identical every time: the agent does its measurable job well, and the unmeasured thing — what the business actually needed — suffers. The fix is always the same: define the real goal before you build the agent, not after you see the results.

Frequently Asked Questions

Does this mean AI agents shouldn't be used for customer-facing work? No — it means they need to be configured carefully for customer-facing work. The agents that work well in customer contexts are the ones built around relationship goals, not just efficiency metrics. An agent that handles frequently asked questions, confirms appointments, or sends status updates is a great fit. An agent that needs to read emotional context and make nuanced judgment calls is not a fit — that's human work, and the agent's job is to escalate to a human, not to attempt it alone.

How specific do escalation triggers need to be? Specific enough that you could train a new employee with them. "Escalate if the customer seems frustrated" is not specific enough — a new employee wouldn't know what that means in practice. "Escalate if the customer uses words like 'unacceptable,' 'lawyer,' or 'refund,' or if they've contacted us more than twice about the same issue in the last 7 days" is specific. Write the triggers the way you'd write a training guide, and the agent can follow them reliably.

What if I get the intent wrong — can I fix it after the fact? Yes, but the cost depends on what went wrong and for how long. A bad follow-up sequence that went out for two weeks is recoverable. An agent that handled 500 customer complaints poorly over three months creates real relationship damage that takes time to undo. Get the intent right before launch. Then monitor outcomes weekly for the first month and adjust based on what you see.

How do I know if my current agent is suffering from an intent gap? Look at the outcomes it's producing versus the outcomes your business actually needs. If the agent's metrics look good but you're seeing downstream problems — lower close rates, more complaints, unexpected churn — you have an intent gap. The agent is succeeding at the task you gave it and failing at the goal you actually needed. The fix starts with going back and writing down the real goal.

Is there an easy way to test intent engineering before deploying? Run the agent in "shadow mode" for one to two weeks before it talks to real customers. It processes real inputs and generates responses, but a human reviews every response before it sends. This lets you see where the agent's judgment diverges from yours, catch the edge cases you didn't think to specify, and refine the intent before any customer experiences the unrefined version.

Build the Intent Layer First

The agent is only as good as the intent behind it. Klarna proved that a brilliant agent optimizing for the wrong goal is worse than no agent at all — because it's fast, scalable, and wrong at every interaction.

Define what your business actually needs. Then build the agent around that.

Associates AI helps businesses work through the intent layer before deployment — so the agent serves your actual goals, not just the metrics that are easy to measure. If you're building toward customer-facing AI, book a call before you launch.