Academy

Klarna went all-AI, then walked it back

The case for full AI support automation is seductive — until the first wrong action fires at scale.

The Felesh team 6 min

The announcement forced an immediate recalculation of customer operations. In early 2024, Klarna publicly declared that its new AI assistant was doing the work of 700 full-time human agents. According to the press and community reports that immediately followed, the system was handling two-thirds of all customer service chats, driving down resolution times, and projecting annualised profit improvements.

For executives watching customer support costs scale linearly with revenue, it was a seductive premise. The narrative suggested you could finally decouple headcount from ticket volume and replace an entire support floor with a single deployment.

Then came the quiet walk-back. Months later, industry observers noted the company had resumed hiring human customer service agents.

To understand why the narrative shifted, you have to look at the menu of options available to operators today—and, more importantly, where each architecture breaks down under load. Customer support is not a single problem; it is a spectrum of friction.

First, there is the legacy menu or IVR bot. It routes traffic through rigid, predetermined decision trees. It loses the moment a customer’s problem falls outside the hardcoded script. Customers tolerate it, but they resent the friction of being forced into a narrow operational funnel.

Second, there is the pure generative chatbot. Powered by large language models, it can read your documentation, parse a knowledge base, and answer questions fluently. It loses when the customer actually needs something done. Concisely explaining a refund policy is not the same as executing a refund. The conversation reaches a dead end, and a human still has to intervene to push the buttons.

Third, there is the human call centre. It remains the baseline standard for empathy and complex problem-solving. It loses on cost and latency at scale. Scaling human headcount to match peak seasonal ticket volume reliably destroys unit economics, and customers still wait in queues.

Finally, there is the AI-only autonomous front line. This was the Klarna path. It involves full agentic execution, where the system has the programmatic keys to read, write, and modify state in the database without human review. It loses when a wrong action fires at scale.

One message, four architectures, four ways to loseA customer writes about order 83-0192: the jacket arrived torn, refund me. The menu or IVR bot says press 1 for orders, 2 for billing, and loses because a torn jacket is not on the menu. The generative chatbot explains the refund policy fluently and loses because nothing gets refunded. The human call centre refunds it with empathy and loses on cost and the queue at peak season. The AI-only front line issues the 640 kr refund with no review and loses when a wrong action fires at scale.customer, order 83-0192: “jacket arrived torn, refund me”menu / IVRpress 1 for orders, 2 for billing…loses: “torn jacket” is not on the menuchatbotexplains the refund policy, fluentlyloses: nothing gets refundedcall centrean agent refunds it, with empathyloses: cost and queues at peak seasonAI onlyissues the 640 kr refund, no reviewloses: a wrong action fires at scale
The same torn-jacket message loses in four different ways, and only the last one does damage at scale.

That last failure mode is where early adopters of autonomous support stumbled. There is a fundamental architectural difference between an AI system that answers and an AI system that acts.

When a pure conversational model hallucinates, it might tell a customer the wrong baggage allowance or misquote a shipping timeframe. That is an isolated customer experience failure. It is frustrating, but its blast radius is contained to one user.

When an agentic system with write-access hallucinates or misclassifies an intent, the mechanics of failure change entirely. An autonomous agent that misreads a prompt might process a refund for a non-refundable item, cancel an active enterprise subscription, or incorrectly modify an account state. The cost of a bad action in production is multiplicative, not additive. You are no longer dealing with a frustrated user; you are dealing with cascading operational liability.

Blast radius of an answer and of an actionTwo panels. On the left, a chatbot hallucinates and tells one customer that 30 kg of baggage is allowed when the policy says 20 kg: one customer, one bad answer. On the right, an agent with write access refunds 640 kr on a non-refundable item, and the same wrong action repeats across every similar ticket. One answer reaches one person; one action reaches every ticket.answersactschatbot, hallucinating“30 kg allowed”policy says 20 kgone customer, one bad answeragent, write accessrefund 640 kr, issueditem is non-refundableevery similar ticketone answer reaches one person; one action reaches every ticket
A wrong answer reaches one customer; a wrong action repeats on every ticket that looks like it.

The quiet rehiring of human agents was not proof that AI cannot do support. It was the natural consequence of deploying the wrong architecture at the wrong moment.

As builders of agentic systems, we have to be honest about where high-autonomy architectures lose. For read-only information retrieval—explaining FAQs, checking public policy, or returning order status—a full agentic system is the wrong tool. For those workloads, a well-tuned chatbot, or a simple chatbot-human blend, will consistently beat an autonomous agent on cost, latency, and risk. Giving a system more autonomy does not automatically make it better; it just gives it a larger blast radius. An agentic platform optimised for complex, state-changing workflows carries overhead that you simply do not need if you are just answering questions.

The lesson from early industry attempts at full automation is not to avoid AI entirely. It is to know exactly which rung of the automation ladder you are standing on.

The automation ladder, rung by rungFour rungs climbing to the right. First, answers: where is order 83-0192, it reads status. Second, suggests: it drafts a 640 kr refund and a person sends it. Third, acts behind guards: a rule, under 700 kr, otherwise a person. Fourth, acts alone with no review, where a wrong action fires at scale. A rust arrow from the first rung straight to the fourth is the Klarna path. Secure the baseline first, then climb one rung at a time.answers“where is order 83-0192?”, reads statussuggests, a person approvesdrafts a 640 kr refund; a person sendsacts behind deterministic guardsrule: under 700 kr, else a personacts alone, without reviewno review: a wrong action fires at scalethe Klarna path:autonomy firstsecure the baseline first, then climb one rung at a time
Climb the ladder one rung at a time; the Klarna path started at the top.

The most reliable way to build AI into your operations is sequentially. The right first rung for most businesses is a conversational interface that answers reliably and retrieves information accurately. You secure the baseline first. Agentic execution—where the system is actually pressing buttons in your CRM or billing platform—should be reserved for a later phase. When you do introduce agents to take actions, they should be deployed strictly behind deterministic guards, and ideally operate as peers to your human agents through a suggest-mode review process before any irreversible action is taken.

Suggest mode: guards first, then a personThe agent drafts a refund of 640 kr for order 83-0192, torn jacket. The guard checks that the item is refundable, the amount is under 700 kr and the order is the customer’s; all three tick. A person reads the draft and the checks and approves, and the refund is sent and logged. A second draft, a refund on a gift card, fails the guard and is held there: no person ever sees it.agent draftsguard checksperson decidesdraftrefund 640 krorder 83-0192torn jacketrefundableunder 700 krtheir orderreads the draftand the checksapprovesrefund sent, loggeddraft 2: gift-card refundheld at the guard: no person sees it
In suggest mode a draft passes the guards and then a person; a draft that fails a guard never reaches the person.

Automation works when it is calibrated to the risk of the task. Before you hand an AI the keys to your database, ensure it can reliably read the map.

Start here →