How Klarna turned customer service triage into a durable AI workflow

0:00Welcome to the Leaders Insights podcast.

0:07Today's episode, how Klarna turned customer service triage into a durable AI workflow.

0:13Everyone keeps telling me Klarna fired 700 people because of an AI agent.

0:18That's the story.

0:19Is it true?

0:20It's the story people want to be true, which is exactly why you should distrust it.

0:25The layoffs made the headlines in early 2024.

0:28The durable part came later, and it wasn't about headcount at all.

0:32The interesting decision Klarna made was refusing to bolt AI onto their existing customer service

0:38process.

0:39What does bolt-on actually mean, though?

0:41That sounds like a slogan.

0:43Fair.

0:44Bolting on means you keep your old workflow, not one, the ticket comes in, gets categorized,

0:50routed to a queue, then a human answers, and you just drop a chatbot in front of it.

0:56Klarna's first instinct was that, and it failed.

0:59The bot became a very expensive speed bump before you reached a real person.

1:04So what did they change?

1:05They rebuilt the triage.

1:07Triage being the sorting step, deciding what a request even is before anyone touches it.

1:12Around the AI agent, instead of around the human org chart.

1:16The old categories existed because humans needed them.

1:19Refund, dispute, where's my stuff?

1:23The model didn't need those buckets, so they threw them out and let it reason from the

1:27raw message.

1:28That sounds like a fancy way of saying they remove structure.

1:31Isn't structure the thing that keeps these systems from hallucinating?

1:35Inventing answers that sound right and aren't?

1:38That's the contrarian bit, and it's where most teams get it backwards.

1:42They think guardrails mean more rules, more decision trees, more if the customer

1:46says X do Y.

1:48Klarna found the opposite.

1:50The more rigid the script, the more the model would twist a genuine request to fit a box.

1:55Give it fewer categories and one clear escalation path to a human.

2:00And the hallucination rate dropped.

2:02Structure in the wrong place is what breaks these things.

2:05Give me a number or I don't believe you.

2:07Klarna's own figures, and this is the vendor talking about itself, so hold it loosely,

2:12claim the agent handled roughly two-thirds of chats and cut average resolution time from

2:1711 minutes to under two.

2:20The number I actually trust more is the repeat contact rate.

2:26How often the same customer comes back because the first answer was garbage.

2:30That fell, and that one's hard to fake because it costs you real money when it's high.

2:35What did they get wrong first?

2:36You keep hinting.

2:38The big early mistake was measuring the agent on deflection, the share of tickets that

2:42never reach a human.

2:44Deflection is the worst metric in this entire field.

2:47You can hit 90% deflection by making it impossible to reach a person, and your customers will

2:52quietly hate you and leave.

2:55Klarna optimized for it.

2:57Watch satisfaction sag and switch to measuring resolved and didn't come back.

3:02Different incentive, completely different behavior.

3:05So the prompt engineering, the way you write instructions for the model, that's the secret

3:09sauce?

3:10No, and I'll die on this hill.

3:13Prompt engineering is the plumbing, not the architecture.

3:15Klarna's win came from redesigning what the job was, then writing prompts to serve that.

3:21Teams obsess over clever prompt tricks, the way people rearrange furniture in a house with

3:26no foundation.

3:27A beautifully worded prompt pointed at a broken process gives you a very polite failure.

3:33That's a little cocky.

3:34What's the evidence a normal team can copy any of this?

3:37Because the transferable part is boring and free.

3:40One, kill your legacy categories and let the model read the raw request.

3:45Two, pick a metric that punishes bad answers, not one that rewards dodging humans.

3:51Three, make the escalation path to a person one click, not a maze.

3:57None of that requires Klarna's budget.

3:59The expensive part was learning it.

4:01You get it for the price of this podcast.

4:03And the funny thing nobody admits?

4:06Half these AI agents work brilliantly in the demo and fall apart in production for the

4:10same reason as Souffle does, the moment you open the door and let reality in.

4:15So the one thing I do Monday morning?

4:17Find the metric your support team is judged on.

4:20If it rewards avoiding customers rather than resolving them, fix that before you write

4:25a single prompt.

4:27Everything else is downstream of that number.

4:29That's your five.

4:30Full AI track, structured, no fluff, is at MBA-training.com.