In May 2025, Klarna CEO Sebastian Siemiatkowski admitted the Swedish fintech had "gone too far" in replacing most of its customer support with an AI assistant. The line that summed up the reversal was another one: it is critical to make clear to the customer that there will always be a human, if they want one. Klarna was hiring human agents back.
The detail that tends to get lost in this story isn't the number of reopened positions. It's the reason the executive himself gave: cost had weighed too heavily in the decision, and the result was lower quality. The company had not decided, ahead of time, where the AI should stop — and found the limit only after customers had already felt its absence.
Klarna's case isn't isolated. A late-2025 study by research firm Metrigy, surveying consumers in the United States, found that 84.7% prefer interacting with a person over an AI agent — and 80.1% keep that preference even when guaranteed the AI would resolve the issue. For anyone deciding on automated support, the relevant number isn't about technical capability. It's about where the line for "on its own" should have ended.
It is critical to make clear to the customer that there will always be a human, if they want one.
The Cost of Removing the Person
Beyond Klarna, the pattern shows up study after study. CX Mag Today, citing Glance's 2026 customer-service trends report, found that 75% of consumers report frustration with automated AI support — even when the response arrives fast. Nearly 90% say they trust a brand less once the option to talk to a person disappears entirely. One in three (34%) say AI support "made things harder," not easier.
The common thread across these studies and the Klarna case is telling: dissatisfaction rarely comes from AI getting things wrong. It comes from AI being the only option. Metrigy's research makes that clear: even when guaranteed the issue would be resolved, most people still prefer a person — because they trust more that they'll be understood, and because explaining a problem once to a human feels safer than repeating it to a system.
That changes what "fixing AI support" actually means. It isn't training a better model, or writing more polite responses. It's deciding, before any conversation happens, exactly where the AI stops deciding on its own — and making sure that point actually exists, not just in the company's talking points.
Klarna acknowledged its own mistake in public, under the market's eyes. Most companies that automate support won't get that kind of visibility: they'll lose customers quietly, with no press statement explaining why.
The Handoff That Fails on Its Own
The most common response to this frustration is adding a "talk to an agent" button at the end of the AI flow. In practice, that button solves less than it seems to.
A Salesforce study on consumer behavior toward AI agents found that 45% of people are more willing to use an automated assistant when a clear handoff path to a person is defined ahead of time. The reverse also holds: when that path isn't clear — when the button exists, but nobody knows what happens after pressing it — distrust remains, even with the option there.
There's an even simpler problem that no button solves by itself: if the handoff to a person doesn't carry the conversation history along, the customer starts over. Repeating a problem to someone who knows nothing of what was already said is, for the person on the other end, nearly as frustrating as not having the option to talk to a human at all.
The common mistake is treating that handoff as an interface feature — one more link in a menu — when it's actually a process decision: what AI never resolves on its own, who receives the case, with what information, and how fast.
What Has to Be in Place
Closing that gap takes concrete mechanisms, not a generic promise of "humanized support."
A declared scope limit before the conversation starts. The list of what AI never decides on its own — cancellations, discounts, deadline exceptions, any promise with contractual weight — needs to exist in writing before the first customer talks to the assistant, not get discovered case by case.
Human approval before the action, not after the complaint. When a conversation touches one of those sensitive points, the right flow is to pause and request a person's confirmation before promising anything to the customer — inside the same conversation, without restarting support on another channel.
Handoff with the full history, never from scratch. Whoever picks up the conversation gets what was already said, what was already tried, and what the customer already explained. That's exactly the point Salesforce's research ties to trust: a handoff without context is just a second line.
A record of what was promised, dated and sourced. Every relevant response gets tied to the document or policy that backed it. If a customer later says "the AI guaranteed me this," the company can check what actually happened.
This is how Skyller was designed: human approval before a sensitive action and conversation handoff with the full history, as the default — not as something a team has to remember to turn on.
The Payoff of Deciding This Early
The benefit of setting that limit ahead of time doesn't show up only in the customer experience. It also shows up in how the company avoids drastic decisions later.
Klarna had to do two things at once, under the market's watch: admit the mistake publicly and rebuild a support team from scratch, at a delicate moment, shortly before going public. A company that designs the handoff point from the start doesn't face that binary choice between "AI only" and "back to all humans" — it adjusts the limit as it learns, without needing a press statement to do it.
There's also an internal-visibility gain that rarely makes the headlines: once every handoff to a person is logged, the company starts seeing a pattern that used to be invisible — what kind of request most often escapes the automated script, at which step, and how often. That pattern is exactly the material that helps decide where to invest the next round of automation, instead of guessing.
The lesson applies at least as much to companies in Brazil and Latin America, where a growing share of support already runs through automated assistants on channels like WhatsApp. The handoff point to a person isn't an implementation detail — it's a business decision that needs to be made before the first customer talks to the assistant, not after the first complaint.
Questions to Decide the Limit
Before expanding AI support to another channel, it's worth bringing these questions to the next meeting with legal and operations:
- Is there a written list of what AI never decides on its own? Cancellations, discounts, and any contractual promise should be on it — defined ahead of time, not discovered in a live conversation.
- When someone asks to talk to a person, does the conversation history go with them? If the answer is no, the handoff button solves less than it looks like it does.
- Does the company know, today, how many conversations asked for a person and couldn't get one last quarter? If the answer is "we don't know," that's the first gap to close.
- Is there a record of what was promised to the customer, and which document it was based on? Without it, a complaint becomes one person's word against another's.
