How Klarna rewired its support operations with disciplined prompt engineering

August 19, 2026 · 5 min · Free, with full transcript

Klarna's AI deployment in customer support became one of the most cited cases of LLMs producing measurable operational results. The prompt discipline behind it offers concrete lessons that transfer well beyond fintech.

Listen to the podcast

4 min

Read the full transcript

Welcome to the Leaders Insights podcast.

Today's episode, how Klarna rewired its support operations with discipline prompt engineering.

Klarna claimed its AI assistant did the work of 700 full-time support agents.

That number gets thrown around like gospel.

Was it real or vendor theater?

Mostly real and worth interrogating.

Klarna said the assistant handled two-thirds of its customer service chats in its first

month back in early 2024, roughly 2.3 million conversations.

The 700-agent figure was Klarna's own estimate of equivalent human workload.

So treat it as a company claim, not an audited fact.

Fine, but every company has a chatbot that annoys people.

What did Klarna actually do differently?

They treated the prompt, the written instruction you give the model, as an engineering artifact

not a suggestion box.

Most companies write a prompt like a wish.

Klarna wrote it like a contract.

Slow down.

Pretend I've never written a prompt in my life.

What is one mechanically?

It's the text you feed the model before it answers.

Imagine hiring a temp who has read the entire internet but knows nothing about your company.

And the prompt is the one-page brief you hand them on their first morning.

Vague brief vague temp.

So the whole game is writing a better brief?

The whole game is writing a brief that survives contact with a real customer.

Klarna's discipline had three habits worth stealing.

First, they scoped the model tightly, told it exactly which tasks it may attempt, and

crucially, which ones to refuse and hand to a human.

That sounds like admitting your expensive AI can't do the job.

It's the opposite.

The refund fraud edge cases, the grieving customer situations.

You route those to humans on purpose.

A model that knows its limits beats one that confidently invents a return policy that doesn't

exist.

That invention has a name, hallucination, when the model states something false with

total conviction.

Second habit?

Grounding.

They didn't let the model answer from memory.

Every response was tied to Klarna's actual policy documents, and the customer's real

account data pulled in at the moment of the question.

So it's not guessing your order status, it's reading it.

Even if the policy changes?

You update the document, not the model.

That's the quiet genius.

When Klarna changed a refund window, a human-edited one source file and every future answer reflected

it.

No retraining, no engineers, no six-week project.

Third?

Relentless evaluation.

They built test sets, collections of real past conversations with known correct outcomes,

and ran every prompt change against them before shipping.

Change one sentence in the brief, rerun a few thousand scenarios, see what broke.

That's the part nobody does, isn't it?

That's the part everybody skips.

Then wonders why the thing goes sideways in production.

Prompting without evaluation is editing a novel with the lights off.

You feel productive.

You have no idea what you've done.

Let me push.

Klarna also quietly rehired some humans later.

Doesn't that puncture the whole story?

It complicates the triumphalism and good.

In 2025, their CEO admitted they'd cut too aggressively and quality suffered on complex

cases.

The lesson isn't AI failed.

It's that the AI was excellent at the repetitive 60% and mediocre at the messy 40, and they'd

blurred the line.

Discipline means knowing which bucket a request falls into before you answer it.

So for an exec who isn't fintech and doesn't have Klarna's engineers, what actually transfers?

The method, not the scale.

Any team drowning in repetitive questions, internal IT tickets, HR queries, first-line

sales can apply the same three moves.

Scope tightly, ground in your real documents, and evaluate every change against real past

cases.

And the thing most people get wrong on day one?

They optimize the clever phrasing and ignore the plumbing.

They spend a week wording the prompt beautifully and zero time deciding what the model isn't

allowed to touch.

Give me the one thing to do Monday morning.

Take your 50 most common customer or employee questions, write down the correct answer to

each yourself, and use that list as your scorecard.

Before any prompt goes live, it has to pass those 50.

If you can't write the answers, your AI can't either.

And now you know that for free instead of finding out from an angry customer.

That's it.

The AI Decision Tools are live at MBA-training.com.

Read and go deeper