Why you still need Pexly’s live agents, who supervise AI

ai-customer-service-pexly

Written by:

Michel Mertens
Managing Director

Klarna became a headline case study for “AI replacing customer service.” In a widely shared LinkedIn Pulse, Jean-Marc Strauven describes how Klarna cut hundreds of customer support roles and pushed a large share of customer conversations to an AI assistant – at first getting the boardroom metrics everyone loves: faster replies and lower cost. Then the customer reality arrived: responses felt generic, issues looped, nuance disappeared, and customers had to repeat themselves after the bot failed. Eventually, leadership publicly admitted they “went too far” and moved back toward a hybrid model where AI handles routine questions and humans step in when the situation exceeds the script.

That story is bigger than Klarna. It’s a pattern we see across industries: companies optimize for volume metrics (tickets handled, response time, cost per interaction) and forget the trust metrics (resolution quality, customer effort, churn, reputation). Reworked summarized the same dynamic: AI can look brilliant on paper, but when customers hit edge cases – or simply need to feel heard – the “efficiency” turns into damage control, rehiring costs and brand repair.

Here’s the uncomfortable truth: AI is excellent at answering. Customers are not asking to be answered – they’re asking to be helped. Help requires judgment. It requires reading between the lines. It requires accountability when a policy exception is needed, when money is involved, when emotions are high, or when the “correct” answer depends on context the bot can’t reliably infer. And when AI gets it wrong, it often does so with confidence – at scale.

it-data-support

The real lesson: don’t replace humans – build a system where AI amplifies them

At Pexly, we like AI. We use it for what it’s great at: speed, routing, summaries, knowledge retrieval and handling repetitive questions at any hour. But we build support operations on a principle that too many companies learn the hard way:

Responsible AI in customer support doesn’t mean “a better prompt.” It means an operating model:

  • Human supervision by design (not as an afterthought)
  • Clear escalation rules (the “red button” to a person)
  • QA and calibration (measuring quality, not only speed)
  • Ownership (a real human accountable for outcomes)

That’s what customers feel. Not “automation.” Not “innovation.” They feel whether your brand takes responsibility when things go wrong.

What “responsible hybrid support” looks like in practice:

A smart hybrid model is simple in concept: AI clears the routine, humans own the moments that decide trust.

  • AI handles repetitive, high-volume tasks instantly: FAQs, order status, appointment reminders, password resets, basic troubleshooting, tagging, routing, and summarizing.
  • Humans take over when stakes or nuance appear: refunds/disputes, escalations, exceptions, complex troubleshooting, vulnerable customers, policy interpretation, and any moment where empathy matters.
  • Together, they reduce queues and protect your brand: faster responses without sacrificing resolution quality.

Klarna’s experience (as described in the sources) is a reminder that the “last mile” of customer support is not a rounding error – it’s the part customers remember.