In February 2024, Klarna deployed an OpenAI-powered customer service assistant and, within a month, it had handled 2.3 million conversations, two-thirds of the company's total support volume, doing work the company said was equivalent to 700 full-time agents. Average resolution time fell from 11 minutes to about two. It was, briefly, the cleanest case study anywhere for the idea that AI had simply solved this job.

Call center agents are rated High risk, with a horizon toward 2027, the most urgent classification this series applies. Fifteen months after the launch, Klarna's CEO Sebastian Siemiatkowski told reporters the company had gone too far, that cost had been too dominant a factor in how the rollout was designed, and that the result was lower quality. Customer satisfaction with the AI assistant had fallen sharply, formal complaints had piled up, and Klarna began rehiring human agents, building a hybrid model where AI handles routine volume and people handle the disputes, fraud claims, and hardship cases the AI kept getting wrong.

That reversal doesn't undo the underlying risk. It sharpens what the risk actually is: this job's most standardized, high-volume work is genuinely being automated at scale, right now. What Klarna's course correction shows is where the automation stops being safe to run without a person watching it.

Key Points

  • Call center agents are rated High risk with a Low-term horizon toward 2027, and unlike most professions in this series, a flagship full-automation attempt has already been tried, at scale, and partially reversed.
  • Klarna's AI assistant handled 2.3 million conversations in its first month in 2024, equivalent to 700 agents, but by 2025 the company's CEO admitted the rollout prioritized cost over quality and began rehiring humans for complex cases.
  • Qualtrics research found roughly 20% of customers reporting no benefit from AI customer service, a failure rate about four times higher than other AI applications, driven substantially by poor handoffs to human agents.
  • Salesforce cut about 4,000 customer support roles in 2025 while reporting a roughly even split between AI and human-handled interactions, alongside a 17% reduction in support costs.
  • Gartner surveys found 64% of customers would prefer companies not use AI for customer service at all, and 87% say they must retain access to a human agent regardless of how AI is deployed.

What a Call Center Agent Actually Does

The job spans a wide range of complexity behind a deceptively uniform title: answering routine questions, account changes, order status, billing lookups, is high-volume and repetitive. Handling a fraud dispute, a hardship request, or an angry customer with a genuinely unusual problem requires judgment, empathy, and the authority to make an exception a script doesn't cover. Most call centers run both kinds of work through the same queue, which is exactly what makes this job harder to fully automate than its most repetitive slice suggests.

The Klarna Reversal

Klarna's 2024 launch was framed, correctly, as a genuine engineering achievement: an AI system handling the bulk of a major fintech's support volume, faster than humans, at a cost the company estimated saving tens of millions of dollars a year. The 2025 walk-back is just as instructive. Customer satisfaction with the AI assistant reportedly fell by double digits, and complaints accumulated over billing and refund handling, exactly the categories requiring judgment about an individual's specific circumstances rather than a lookup against a policy document. Siemiatkowski's public admission, that evaluating the rollout primarily on cost produced a lower-quality outcome, is a rare instance of a company naming the tradeoff explicitly rather than letting quarterly earnings quietly absorb it.

THE HANDOFF PROBLEM

Qualtrics research found roughly 20% of customers getting no benefit at all from an AI customer service interaction, a failure rate about four times higher than other AI applications the same research tracked, and identified the handoff from bot to human as a recurring point of failure: customers forced to repeat information, agents receiving an escalated case with no context. Air Canada learned the liability version of this the hard way in 2024, when a tribunal held the airline responsible for false refund information its chatbot gave a customer, rejecting the company's argument that the bot was a separate entity.

What AI Is Doing Elsewhere

Klarna's experience hasn't slowed the broader deployment. Salesforce cut roughly 4,000 customer support roles in 2025, with CEO Marc Benioff saying the company needed "less heads" thanks to its AI agent platform, reporting close to an even split between AI-handled and human-handled interactions and a 17% reduction in support costs. Sierra, an enterprise voice-AI startup founded by former Salesforce co-CEO Bret Taylor, reached roughly $100 million in annualized revenue within seven quarters and was valued at $15.8 billion in a 2026 funding round.

The more durable pattern, though, may be assist rather than replacement. Cresta and Observe.AI sell real-time guidance to human agents rather than autonomous bots, and their published customer results, average handle time cut by roughly 15%, with some deployments reporting cuts as high as 40%, alongside measurable gains in customer satisfaction, describe a model where AI makes each agent faster without removing the agent. That is a smaller headline than "AI replaces the call center," and it may be closer to what actually survives contact with customers who have a genuinely hard problem.

What Customers Actually Want

The customer-preference data cuts hard against full automation. Gartner found 64% of customers saying they'd prefer companies not use AI for customer service at all, and a separate survey found 87% saying they must retain access to a human agent no matter how a company deploys AI. Trust splits along similar lines: in one survey, 54% of customers said they trust a human agent's recommendation more than an AI's, against 32% who trust the AI. None of this stops companies from deploying AI. It does explain why the deployments that stick tend to preserve a human escalation path rather than eliminate it.

The Labor Market Reality

The U.S. Bureau of Labor Statistics projects a 5% decline in customer service representative employment from 2025 to 2035, against a base of about 2.8 million jobs, with a median wage of $20.59 an hour. Even with that decline, the agency projects roughly 289,500 openings a year, almost entirely from turnover in an occupation that has always had high churn. This is a profession shrinking at the margin, not disappearing, and the openings number suggests the immediate experience for most current agents is more likely to be reduced hours or non-renewal than a mass layoff event.

How to Use AI as a Call Center Agent Now

For routine volume: assume AI will keep absorbing the most standardized interactions, account lookups, status checks, simple billing questions. Resisting that is not a viable position for an individual agent or a company.

For complex cases: the Klarna and Qualtrics data both point the same direction, escalations, disputes, and anything emotionally charged are where AI-only handling fails most visibly. Building expertise in exactly that category is the most defensible position in this profession right now.

For agent-assist tools: platforms like Cresta and Observe.AI make an individual agent measurably faster and more consistent without removing them from the interaction. Adopting these tools well is a more available career strategy than competing with a fully autonomous system for the same call.

What I Think

The 2027 horizon and High-risk rating hold up, and Klarna's course correction doesn't really contradict that; it refines it. The routine slice of this job is being automated on the timeline the rating implies. What the reversal shows is that companies optimizing purely for cost reduction on the complex slice run into customer backlash and legal liability fast enough to force a correction, which is a real constraint on how far full automation goes even when the technology is capable enough to attempt it.

What I'd watch is whether other companies learn from Klarna's specific mistake, treating quality and escalation design as first-class requirements rather than an afterthought, or whether they repeat it at smaller scale without the public reckoning. Either way, the job that survives this transition looks less like today's call center and more like a smaller team of people handling the cases that were always the actual reason a human needed to be on the line.

"Klarna proved the technology could do 700 people's jobs. It also proved, within a year, which parts of those 700 jobs it couldn't do without making customers angry enough to complain to a regulator."