Klarna: Automate the common case, keep a human path.
Klarna put an AI assistant in front of its customer service and reported it handled two-thirds of chats in its first month. Explore what to automate, what to measure and why the way back to a person matters.
See the idea in one picture ↓- of chats in month one
- 2/3
- to resolve, down from 11
- < 2 min
- repeat inquiries
- −25%
- languages
- 35+
Figures as reported by the company.
Independent educational analysis. Scenarios, people, numbers and illustrations on this page are original PMcademy practice material, not company research results or an endorsement.
Most questions are common. Some are not.
Every question waits in a queue
Routine answers in minutes, people for the rest
Read the analysis
Imagine your support team answers thousands of messages a day. Most are about the same few things: a refund, a payment date, a return. A few are complicated, emotional or new. An assistant can answer the common ones instantly, and it can also frustrate someone whose problem does not fit. The product question is not only how many conversations you can automate. It is what happens to the people whose case is not common.
Fast is good. Stuck is not.
I just want to know where my money is. Today.
What would help?
A fast, correct answer. And a person if it gets complicated.
What gets in the way?
Waiting in a queue. Looping with a bot.
Fictional personRead the analysis
Our fictional learner, Elsa, wants to know why a refund has not arrived. She does not care who answers, as long as the answer is right and quick. Her neighbour, Johan, has a disputed charge and wants a person who can take responsibility. Both use the same chat window. Ask which of your users want speed, which want a person, and how each of them can tell which one they are getting.
Where each question goes
- 1One queue
Without an assistant, every message waits in line for a person.
- 2Assistant first
Routine questions are answered in minutes, and the rest go to a person.
- 3Keep the way back
A person stays one step away, and repeat contacts show whether problems were really solved.
Animated illustration. Names, numbers and people are fictional teaching material.
Answer the routine, route the rest.
Illustrative data and simplified behaviour. This model does not represent a live company system or measured company results.
Read the analysis
Use the model below to compare three set-ups. In the first, every message waits for a person. In the second, an assistant answers first and a person is one step away. In the third, there is no person at all. Watch the wait time and the number of people who come back with the same problem. The numbers in the model are illustrative.
Cheaper and faster, until it is wrong.
- Answers any time, in many languages
- Faster resolution
- Lower cost per contact
- Confident wrong answers
- Emotion missed
- Trust lost slowly, savings seen fast
Read the analysis
An assistant can be available all day in many languages and answer common questions faster than a queue. It can also be confidently wrong, miss emotion and leave people looping without a way out. Savings show up at once, while lost trust shows up later. Decide which questions the assistant must hand over, how it says it is unsure, and what a person needs to see when they pick up a conversation.
Count the repeats, not only the replies.
Klarna reports 2.3M conversations, satisfaction on par with people and 25% fewer repeat inquiries in month one.
How definitions were set, or what happened after the first month.
- Problems solved first time
- Hand-overs to a person
- Repeat contacts and complaints
Read the analysis
Klarna reported faster resolution, satisfaction on par with people and a 25% drop in repeat inquiries in the first month. These are the company’s own figures, without published definitions or a long-term follow-up. In your project, measure the problem solved rather than the conversation closed: repeat contacts, hand-overs, complaints and whether people come back to buy again.
Where does the person come in?
Your fictional team runs support for an insurance app. Claims are stressful, and most questions are about claim status. You can let an assistant answer everything, let it answer status questions and route claims to people, or keep people for everything and use the assistant only to draft replies. Choose a set-up, name the moments that must reach a person and the metric that would make you change course.
Choose a next move
Leave with a hand-over map.
Your artifact
Write a hand-over map: your ten most common contact reasons, who answers each, the trigger for a hand-over, and the paired metrics you will track.
Critique my project artifact using only the evidence I provide. Write a hand-over map: your ten most common contact reasons, who answers each, the trigger for a hand-over, and the paired metrics you will track. Separate observed facts, assumptions, proposed tests and missing evidence. Identify a failure mode and a cheap check. Do not invent company results or user research. Ask for my artifact before starting.
Read the analysis
List the ten most common reasons people contact you. For each, decide whether the assistant answers, drafts for a person, or hands over at once, and write the signal that triggers a hand-over. Add how the person picking up sees the conversation so far, and the two numbers you will track together: problems solved first time and repeat contacts.