what klarna really teaches

Everybody filed the Klarna reversal under proof that AI in customer service just doesn’t work. Wrong read. What Klarna really is, honestly, is a perfect example of AI hype, and way too much of it. You know the AI mindset I keep banging on about, the one that’s too small, the owner who never gets past the chat box? This is the mirror image. Here it went too big. Too much trust in AI isn’t any better than too little, it’s the same mistake landing in the wrong spot from the other side. And the numbers say it plainly: by mid-2024 the assistant was already clearing two thirds of every service request, the work of 700 full-time agents (eMarketer, 2025). So the tech held up fine. What broke was everything Klarna built around it. They rolled the AI out as a straight swap for the service team and cut the human safety net clean away. That’s the real lesson sitting under the headline, and it’s an architecture one at heart. You’re picking between full replacement and a hybrid where AI is the front door and a person’s standing right behind it.

And look, there’s a reason you want that person there. Your employees don’t want the all-AI world, that’s one. But your customers don’t want it either, a place where every last thing runs through a bot. And it just isn’t fair yet to hand your complex, messy, emotional questions fully over to AI. You still need the human part. Klarna learned that one the expensive way.

what exactly went wrong at klarna?

Klarna pulled the people out from under its own AI, and service quality sank right along with them. Hiring got mostly frozen in late 2023, and somewhere around 700 jobs quietly went away between 2022 and 2024, handed off to an assistant built on OpenAI tech (MLQ.ai, 2025). On paper it held for a good while. By mid-2024 the bot was resolving two thirds of all service requests, the whole of that departed team’s work (eMarketer, 2025).

Then May 2025 flipped it. CEO Sebastian Siemiatkowski said he’d start hiring people again, because the quality just wasn’t there anymore. Cost, he told Bloomberg, had grown into too big a factor in the company’s decisions, and service was the thing paying for it (eMarketer, 2025). And in October 2025, right after the US IPO, he put it about as plainly as a CEO ever does: “We went too far.” Klarna’s been building a hybrid ever since, flexible service agents working next to the AI (MLQ.ai, 2025).

Now read what that actually says, because the headline gets it backwards. The bot kept handling two thirds of the questions, no drama. The trouble sat in the other third: the complex, the unusual, the emotional, the cases that no longer had a decent route to a human. So what Klarna hired back in 2025 was mostly the safety net it had ripped out itself. That’s a very different story from “Klarna quits AI.”

are you liable when your chatbot makes things up?

Yes. Whatever your bot tells a customer, you’re on the hook for it. Air Canada found that out in February 2024, when a tribunal in British Columbia ruled the airline had to pay for a bereavement discount its chatbot had invented out of thin air. Air Canada’s defense? That the chatbot was a separate legal entity, responsible for itself. The tribunal swept that straight off the table (American Bar Association, 2024). The money was tiny, about 650 Canadian dollars. The precedent’s the whole point: a company can’t go and hide behind its own software.

The second one’s quieter, and honestly I find it more useful. April 2025, the AI support bot at the coding platform Cursor invented a policy on session limits that had simply never existed. The bot was called “Sam” and never once said it was AI, so the annoyed user figured they were talking to an employee. People canceled their subscriptions in frustration before a cofounder could get out in public and say no, there’s no such policy (The Register, 2025). Cursor’s labeled its AI answers as AI ever since.

For European companies this is about to be more than a reputation bruise. From 2 August 2026, a few weeks out now, article 50 of the EU AI Act makes chatbots identify themselves as AI at first contact. And the revised product liability rules, live at the end of 2026, put companies squarely on the hook for misleading or wrong chatbot information (Latham & Watkins, 2026). What the law does and doesn’t ask of you beyond that, I get into in the ai act without panic.

do customers actually want ai customer service?

Yes, more than half are open to it, on one hard condition: it’s got to actually fix their problem. The Dutch National Voice Monitor 2026 puts 52 percent as positive or neutral-positive on autonomous AI customer contact (Consultancy.nl, 2026). And then the same survey shows how far the practice trails that goodwill: only 12 percent of questions put to a chatbot or voice assistant get a satisfactory answer. That gap, right there, is the whole problem in one line.

Here’s where I’ve felt it myself. Meta. They run all their customer communication by email, all of it through AI, barely a human left anywhere in it. Technically it’s a really good model. And I found it genuinely terrible to be on the other end of. If you haven’t had that experience yet, or you’re very bullish on AI, take it from someone who’s sat in it: it just isn’t there yet, and you don’t want to go there yet.

Customers are pragmatic about the rest of it, mind you, a lot more pragmatic than the bot hatred on social media makes out. They judge the result. The channel only starts to matter to them once things go wrong. US research finds that in a financial dispute 85 percent want a human, against a handful who’d pick AI (SurveyMonkey, 2025). And after a handover, 70 percent care that the agent already knows what they just told the bot (Avaya, 2026). Nobody wants to tell their story twice. Hold onto that number, because it’s the one that tells you how to build the thing.

where does ai demonstrably work in customer contact?

A chatbot’s genuinely handy, and it can do a lot of good: for your customer contact, for convenience, for cost efficiency, for the experience itself. But then you’ve got to hand it a really well-bounded process. So go look at your own data. Which customer questions are directly answerable? Which ones get asked all the time? Which ones are simple? Order status. Return procedures. Opening hours. Rescheduling an appointment. The question’s predictable, the answer already lives in your systems, and a mistake shows up fast. Take those bounded bits, let the bot solve them, and put it somewhere easy to reach. AI earns its keep there, too: because of it, 74 percent of consumers now expect service to be reachable around the clock (Zendesk, 2026), and for a small business with no night shift, staffing that with people alone just isn’t doable.

Here’s how I’ve built these before. Purely on a tight, well-bounded process. The bot can easily answer a hundred-odd questions on its own. But the moment it notices it’s out of its depth, that it hasn’t got the knowledge, it pulls in a service agent fast. And that agent walks in with context already: they get a summary from the bot, so they’re not reading the whole chat back, they can help the customer along with a bit of explanation straight away. Those are the really good use cases. What you don’t do is rip the manual service process out entirely. Do that and you fall flat, and the customer experience nosedives.

One more thing, and it’s the trap. Stay sober about the vendor numbers. The automatic resolution rates chatbot vendors love to quote swing wildly from one source to the next, from modest to near perfect, and not one of them has been independently checked. That spread is the lesson on its own. The distance between the marketing demo and your own reality is exactly the gap Klarna fell into, and it’s why so many AI pilots never make it across, which I got into in the production gap. The researchers behind the Voice Monitor see the market drifting from “human in the loop” to “AI in the loop”, AI that makes the agent faster instead of replacing them. For most small businesses that’s the right order. First the AI that speeds up your team. Then, maybe, the AI that talks to customers on its own. And always with a human within reach.

automating customer contact: seven steps for small business

Here’s how I’d go at it, in the order I actually trust, without walking back into Klarna’s mistake:

1. Take three months of customer questions and sort them by volume and type. The factual, recurring ones are your automation candidates. Anything with emotion, money, or custom work in it stays with a person. Full stop.
2. Pick a narrow task set, five to ten question types, and set the bot up for those alone.
3. Design the escalation before you design 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 tells their story twice.
4. Have the bot introduce itself as AI in its very first message. From 2 August 2026 that’s mandatory in the EU anyway, and the Cursor mess shows it was the smart move well before that.
5. Keep the bot inside your documented policy, and have it refer onward the second it’s in doubt. A bot that admits it doesn’t know costs you nothing. A bot that invents policy? That one bites.
6. Measure three things separately: resolved by the bot, resolved after escalation, abandoned. Then measure customer satisfaction across all three.
7. Only expand after a month of measuring, at least, and do it one task at a time.

And the things you don’t need to do? Fire anyone. Train your own language model. Drop six figures on an enterprise platform. A plain standard tool with a well-built safety net gets you further than a pricey bot without one.

My take, after a year and a half of reading Klarna postmortems and building this stuff myself: automate the predictable, guarantee the human right beside it, and never let a bot improvise policy. “AI or people” was always the wrong question. The real one is where your own boundary sits, and how strong your safety net is underneath it.