Customer Support Automation
AI customer support works. Here is where it stops working.
Klarna's AI assistant became the most-cited example of AI support, first for its results and then for its reversal. Here is what both halves of that story mean for a growing brand.
If you have read anything about AI in customer service, you have probably heard about Klarna. The full story is more useful than the headline.
The headline
In early 2024, Klarna reported that a month after launch, its AI assistant had handled 2.3 million conversations, about two-thirds of its customer service chats.12 The company said that was the equivalent work of 700 full-time agents, that customer satisfaction was on par with human agents, and that average resolution time had dropped to under two minutes.12 It estimated a $40 million profit improvement for 2024.1
The reversal
By May 2025, the tone had changed. Klarna's CEO told Bloomberg that cost had become "too predominant an evaluation factor," which led to "lower quality," and the company began recruiting human customer service workers again.2 According to the company, the AI assistant was still handling about two-thirds of inquiries, with response times 82% faster and repeat issues down 25% since launch.2
Read together, these are not contradictory. The AI kept doing the work it was good at. The mistake was treating every conversation as if it were that kind of work.
What this means for your business
AI is excellent at questions your own data can answer. "Where is my order?", "Can I return this?", "Does this come in a larger size?", "When will this be back in stock?" These are high-volume, repetitive and factual. An AI that can look up the order, the policy and the product catalog will answer them instantly, at 2 a.m., in your brand's voice.
AI is the wrong tool for moments that need judgment or empathy. A damaged wedding gift, a billing dispute, a loyal wholesale account that is unhappy. These conversations decide whether a customer stays, and they deserve a person.
The design question is the handoff. Good support automation is not about how many tickets the AI closes. It is about how cleanly it recognizes what it should not handle and passes it to a person with the full context, so the customer never repeats themselves.
How we build it
When we set up support automation, we follow a few rules:
- Start with the top ten questions. Most inboxes are dominated by a handful of repeat questions. Automating those first delivers most of the value with the least risk.
- Answer only from real data. The assistant reads your live orders, policies and product information. If the answer is not there, it does not guess.
- Escalate early and generously. Refund requests above a threshold, complaints, anything emotional, and anything the AI is unsure about go to a person.
- Measure quality, not just volume. We track accuracy, customer satisfaction and repeat contacts alongside deflection, so cost savings never quietly erode the experience.
The takeaway
The lesson from Klarna is not that AI support fails. It is that AI support should be designed around your customers, not around a headcount target. Done that way, your team spends its time on the conversations that matter.
If your inbox is full of questions your systems already know the answer to, request a consultation. We will show you which ones AI can take on, and which ones it should leave to you.