AI Customer Experience Strategy: Lessons From Klarna, Air Canada, and the Metrics That Matter

Two stories from the past few years frame what AI can and cannot do for customer experience.

In February 2024, Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of all its customer service chats. It said the assistant did the work of 700 full-time agents, cut average resolution time from 11 minutes to under 2, and reduced repeat inquiries by 25%. Fifteen months later, CEO Sebastian Siemiatkowski told Bloomberg that the company's focus on cost had led to lower quality, and Klarna began making sure customers could always reach a human. Klarna disputes that this was a reversal; AI still handles most of its chats. But the company changed course on how far to push it.

The same month as Klarna's announcement, a British Columbia tribunal ruled in Moffatt v. Air Canada that the airline had to honor a bereavement discount its website chatbot had wrongly promised. Air Canada had argued the chatbot was responsible for its own words. The tribunal disagreed.

Between them, those cases cover the two ways AI customer service goes wrong: pushing automation past what customers will tolerate, and forgetting that the bot speaks for the company.

What customers say they want

Gartner surveyed 5,728 customers in December 2023 and published the results in July 2024. 64% said they would prefer companies did not use AI for customer service, and 53% said they would consider switching to a competitor over it. The biggest worry was that AI would make it harder to reach a person, followed by job losses and wrong answers.

That does not mean customers reject quick automated answers. Most people are happy to reset a password or check an order status without waiting for an agent. What they reject is AI as a wall. The design principle that follows is simple: use AI to make easy things instant, and make the path to a human obvious and short.

Where AI works in customer experience

Resolving routine requests end to end

Order status, password resets, address changes, refund status, plan changes, and "how do I" questions answered from documentation. These are high volume, low risk, and well suited to an assistant connected to your systems through tools that can actually take the action. An assistant that can only explain how to change an address resolves far less than one that can change it.

Helping agents on everything else

For conversations that reach a person, AI can summarize the history so the customer does not repeat themselves, suggest answers from the knowledge base, draft replies, and fill in after-call notes. This is lower risk than full automation because a person checks the output, and it often pays back faster.

Fixing problems before they become contacts

Many support contacts are predictable. For subscription businesses, failed card payments are a big one: expired or replaced cards cause involuntary churn that customers never intended. Card network updater services (Visa Account Updater, Mastercard Automatic Billing Updater), reminders before a card expires, and retry logic timed to when payments are likely to succeed all reduce it. Usage drops, repeated errors, and stalled onboarding are similar signals worth acting on before the customer writes in. Our SaaS churn reduction guide covers these in detail.

Routing and prioritization

Classifying incoming messages by intent, urgency, and sentiment and sending them to the right queue is a mature, reliable use. It shortens resolution time without the customer ever noticing AI was involved.

The chatbot speaks for you

After Moffatt v. Air Canada, the safe assumption is that anything your chatbot tells a customer is a statement by your company. Practical consequences:

• Ground answers in approved content. Retrieval from your current policies and help articles, with the assistant instructed to say it does not know rather than guess.
• Keep policies in one place. The Air Canada chatbot contradicted a policy page elsewhere on the same website. If the knowledge base is out of date, the bot will be too.
• Restrict what it can promise. Refunds, discounts, exceptions, and anything with legal or financial consequences should come from a system rule or a person, not from generated text.
• Log conversations. You need to know what was said when a dispute arises.
• Disclose it is AI. Required in the EU under Article 50 of the AI Act since August 2026, and good practice elsewhere.

Build or buy

Customer service AI is a crowded, mature product category. Most companies should buy the platform (ticketing, chat, voice, and the assistant layer) and spend their own engineering time on the integrations that let the assistant take actions in their systems and on keeping the knowledge base accurate. Building in-house makes sense mainly when the product itself is unusual enough that no vendor's assistant can handle it, or when support data is sensitive enough that it cannot leave your environment. The SaaS vendor management guide covers evaluating vendors.

Measuring honestly

AI support metrics are easy to flatter. The most common problem is counting a conversation as "contained" because it never reached a person, when the customer actually gave up.

A useful test: sample 100 "contained" conversations each month and have someone read them. It takes an afternoon and catches problems the dashboards miss.

A rollout that avoids the common mistakes

1. Find the volume. Pull three months of contacts and group them by reason. Usually a handful of reasons account for a large share.
2. Pick two or three reasons that are routine and low risk, where the answer or action is clear-cut.
3. Fix the knowledge base first. Remove contradictions and outdated articles; the assistant will repeat whatever it finds.
4. Connect actions, not just answers. Let the assistant look up orders and make permitted changes through controlled tools.
5. Design the handoff. A visible "talk to a person" option, a summary passed to the agent, and a target for how long escalation takes.
6. Measure resolution and repeat contacts from day one, then expand to more contact reasons only when those hold up.

For the wider AI picture, see AI business strategy. For the data protection side of feeding customer conversations into AI tools, see the AI and SaaS data privacy guide. And for automating the processes behind the conversation, workflow automation.

This guide is for informational purposes only and is not legal advice. Requirements for AI disclosure and consumer protection vary by jurisdiction.