what klarna really teaches

Klarna’s reversal does not prove that AI in customer service failed. By mid-2024 the AI assistant was handling two thirds of all service requests, the work of 700 full-time agents (eMarketer, 2025). What broke was the construction around it: Klarna deployed AI as a full replacement for its service team and cut away the human safety net. The real lesson for anyone automating customer contact is therefore an architecture lesson: choose between full replacement and a hybrid setup in which AI is the front door and a human is the safety net.

what exactly went wrong at klarna?

Klarna removed the people underneath its AI, and service quality sank with them. The company largely froze hiring in late 2023 and let roughly 700 jobs disappear between 2022 and 2024, replaced by an assistant built on OpenAI technology (MLQ.ai, 2025). On paper it worked for a long time: by mid-2024 the bot was resolving two thirds of all customer service requests, the equivalent of that entire departed team (eMarketer, 2025).

In May 2025 came the turn. CEO Sebastian Siemiatkowski announced he would hire people again because quality was falling short. He told Bloomberg that cost had become too dominant a factor in the company’s decisions, at the expense of service (eMarketer, 2025). In October 2025, just after the US IPO that valued Klarna at 19.65 billion dollars, he got more explicit still: “We went too far.” Klarna has since been building a hybrid model with flexible, partly remote service agents working alongside the AI (MLQ.ai, 2025).

Look closely at what these facts actually say. The bot simply kept handling two thirds of the questions. The problem sat in the remaining third: the complex, unusual and emotional cases that no longer had a decent human route. What Klarna restored in 2025 was mostly a safety net it had cut away itself. That is a fundamentally different story from the “Klarna quits AI” headline that echoed everywhere.

are you liable when your chatbot makes things up?

Yes: as a company you remain responsible for what your bot tells customers. Air Canada learned that in February 2024, when the Civil Resolution Tribunal of British Columbia ruled that the airline had to pay for a bereavement discount its chatbot had invented out of thin air. Air Canada argued that the chatbot was a separate legal entity with a responsibility of its own; the tribunal swept that defense off the table (American Bar Association, 2024). The damages, for the record, were small, about 650 Canadian dollars. The value of the case sits in the precedent: a company cannot hide behind its own software.

The second incident is subtler and therefore more instructive. In April 2025, the AI support bot of the coding platform Cursor invented a policy on session limits that had never existed. The bot was called “Sam” and did not identify itself as AI, so the complaining user assumed they were talking to an employee. Users canceled their subscriptions in frustration before cofounder Michael Truell could publicly set the record straight that no policy of that kind existed (The Register, 2025). Cursor has labeled AI answers explicitly as such ever since.

For European companies this is about to become more than a reputation issue. From 2 August 2026, a matter of weeks from now, article 50 of the EU AI Act requires chatbots to identify themselves as AI at first contact; depending on the violation, fines can run to tens of millions of euros or a percentage of global annual revenue. The revised European product liability directive, which takes effect at the end of 2026, additionally makes companies liable for misleading or incorrect chatbot information (Latham & Watkins, 2026). What the law does and does not ask of you beyond that is covered in the ai act without panic.

do customers actually want ai customer service?

Yes, over half are open to it, on one hard condition: it has to genuinely solve their problem. The Dutch National Voice Monitor 2026, a representative survey of over a thousand Dutch consumers, shows that 52 percent are now positive or neutral-positive about autonomous AI customer contact (Consultancy.nl, 2026). The same survey shows how far practice lags behind: only 12 percent of questions put to a chatbot or voice assistant get answered satisfactorily, and as a preferred channel the chatbot scores 0 percent, against 48 percent for the phone. Poorly performing chatbots sit near the top of the irritation list at 50 percent, just behind long wait times (52 percent).

The same pattern shows up internationally. US research by the software company SurveyMonkey finds that only 8 percent of consumers prefer AI over a human, and that in financial disputes 85 percent choose a human against 5 percent for AI (SurveyMonkey, 2025). US consumer research cited by the telecom company Avaya adds nuance in turn: 56 percent are perfectly happy with automation as long as it solves the problem quickly, and 70 percent find it important that an agent already knows the context of the AI conversation after a handover (Avaya, 2026). Customers, in other words, are more pragmatic than the bot hatred on social media suggests; they judge the result, and the channel only starts to matter to them once things go wrong.

And then there is the perception gap on the company side. In the software company Salesforce’s own State of Service study of over three thousand service professionals, 65 percent believe customers fully trust AI; independent consumer data from Metrigy, cited in that same report, lands at 44 percent (Salesforce, 2026). A gap of 21 percentage points between what companies think and what customers themselves say. That is the Klarna story in miniature, playing out every day.

Meanwhile the Dutch public is getting stricter. Private AI use rose from 47 percent in late 2024 to 65 percent in early 2026 according to the Dutch AI-Barometer, while the share of Dutch people with a negative view of AI grew from 25 to 31 percent (Nederland Digitaal, 2026). People know the technology better and better, and so they forgive a failing bot less and less easily.

where does ai demonstrably work in customer contact?

On narrow, factual, high-volume questions with one verifiable right answer: order status, return procedures, opening hours, account questions, rescheduling appointments. There the question is predictable, the answer lives in your own systems, and a mistake surfaces quickly. AI also solves a real problem there: according to the software vendor Zendesk’s own trend research, 74 percent of consumers now expect customer service to be reachable 24 hours a day because of AI (Zendesk, 2026). For a small business without a night shift, that is unaffordable with people alone.

Adoption is moving fast in the meantime. According to the same Salesforce study, 66 percent of customer service organizations now use agentic AI, up from 39 percent a year earlier, and customer satisfaction is the KPI that improves most often after adoption (Salesforce, 2026). At the same time, stay sober about vendor numbers: the automatic resolution rates chatbot vendors claim vary wildly by source, from modest to near perfect, and none of those figures has been independently verified. That spread is itself the lesson. The distance between the marketing demo and your own practice is exactly the gap Klarna fell into; why so many AI pilots never cross that gap is something I described earlier in the production gap.

The researchers behind the National Voice Monitor accordingly see the market shifting from “human in the loop” to what they call “AI in the loop”: AI that supports the agent instead of replacing them (Consultancy.nl, 2026). For most small businesses that is the sensible order. First AI that makes your team faster, only then AI that talks to customers on its own, and always with a human within reach.

automating customer contact: seven steps for small business

This is how you approach it without repeating the Klarna mistake:

1. Analyze three months of customer questions and sort them by volume and type. Factual, recurring questions are candidates for automation; anything involving emotion, money or custom work stays with people.
2. Pick a narrow task set of five to ten question types and configure the bot for those alone.
3. Design the escalation before the bot. Hard triggers (a complaint, a money issue, an angry tone, a second failed attempt) route straight to a human, conversation history included, so the customer never has to tell their story twice.
4. Have the bot introduce itself as AI in its first message. From 2 August 2026 that is mandatory in the EU; the Cursor incident shows it was wise well before then.
5. Restrict the bot to your documented policy and have it refer onward when in doubt. A bot that says it does not know costs you nothing; a bot that invents policy does.
6. Measure three things separately: resolved by the bot, resolved after escalation, and abandoned. Measure customer satisfaction on all three.
7. Expand only after at least a month of measuring, task by task.

And what you do not need to do: fire anyone, train your own language model, or buy a six-figure enterprise platform. A standard tool with a well-designed safety net gets further than an expensive bot without one.

My position, after a year and a half of reading Klarna postmortems and building customer contact myself: automate the predictable and guarantee the human, and never let a bot improvise policy. “AI or people” is the wrong question; the right question is where your own boundary sits and how strong your safety net is.