Can AI replace a call centre agent

TL;DR

Can AI Replace a Call Centre Agent? What the Evidence Says

  • The best field evidence shows AI as an assistant lifting agent productivity ~14% - with the biggest gains (~35%) going to the newest agents, and turnover falling 8.6 points.
  • Customers keep voting for humans: 64% would prefer no AI in customer service; 53% would consider switching over it.
  • Klarna ran the full replacement experiment in public - 700 agents' work claimed in 2024, humans rehired in 2025.
  • The realistic model is triage: machines take the routine, humans take the difficult - which changes who the job suits.

There are two camps in this argument, and both arrived before the evidence did.

One camp says the call centre agent is finished - that a technology which answers instantly, never sleeps and costs cents per contact must eventually replace a human who does none of those things. The other says AI is a gimmick that collapses the moment a customer is angry, confused or in trouble. Both camps can point at anecdotes. Since 2024, though, the question has acquired something better than anecdotes: a large randomised field study in a top economics journal, repeated customer surveys at scale, and one company generous enough to run the full replacement experiment in public.

Can AI replace call centre agents?

For routine contacts, increasingly yes; for the job as a whole, the evidence says no. Gartner predicts agentic AI will resolve 80% of common customer service issues by 2029, yet its 2026 survey found 85% of service leaders expanding human agent responsibilities rather than eliminating them. The strongest field evidence - a study of 5,172 agents published in the Quarterly Journal of Economics - found AI worked best as an assistant, lifting productivity by around 14% while the humans kept the conversation. Klarna, which claimed its AI replaced 700 agents in 2024, was rehiring humans by 2025.

The rest of this article walks through each strand, because the details are where the useful conclusions sit.

The field study: AI behind the agent

How much more productive are call centre agents with AI?

The landmark study - Brynjolfsson, Li and Raymond, published in the Quarterly Journal of Economics in 2025 - tracked 5,172 customer support agents at a Fortune 500 software firm as a generative AI assistant was rolled out. Issues resolved per hour rose 13.8% on average, but the gain was concentrated among newer and lower-skilled agents, who improved by around 35%, while top performers saw little change. Turnover among agents with AI access fell by 8.6 percentage points, and customers escalated to supervisors less often.

Three details of this study reward attention. First, the AI never spoke to a customer. It suggested responses; the agent decided. The productivity gain came from a machine putting the accumulated phrasing of the best agents at the fingertips of everyone else, which is why the newest agents gained most - the assistant compressed months of tacit learning into their first weeks.

Second, the humans got happier, or at least stayed longer. An 8.6 percentage point drop in turnover is enormous in a sector where attrition runs at 30-40% and replacing one agent costs 30-50% of a salary. Customer sentiment improved too, and requests for a manager fell.

Third, the study measured a version of AI that made agents better rather than fewer. That distinction - augmentation versus replacement - turns out to be the entire argument.

+13.8% Issues resolved per hour when 5,172 agents got a generative AI assistant - with novice agents gaining around 35% Brynjolfsson, Li & Raymond, QJE 2025
64% Customers who would prefer companies did not use AI for customer service; 53% would consider switching to a competitor over it Gartner, July 2024
-8.6pts Fall in agent turnover when AI assistance was introduced - in a sector that loses roughly a third of its people a year Generative AI at Work, NBER w31161

The customers: a stubborn preference

Do customers prefer AI or human agents?

Customers keep telling surveys they want humans. Gartner's December 2023 survey of 5,728 customers found 64% would prefer companies did not use AI in customer service, and 53% would consider switching to a competitor if a company planned to. Their top concerns were reaching a human becoming harder, AI displacing jobs, and wrong answers. In Gartner's later survey of 5,801 US customers, 54% trusted a human agent more than AI for recommendations against 32% the other way. Preference is strongest where stakes and emotion are high.

It is worth being precise about what customers are objecting to, because it is not speed or automation as such. Nobody campaigns against the bot that resets a password in eleven seconds. The stated concerns are about the failure modes: being trapped in a loop with no route to a person, being given a confident wrong answer, and the wider unease about jobs. A company that automates the routine and makes the human easy to reach when it matters is not really the thing 64% of customers are objecting to. A company that uses AI as a wall between customers and payroll is.

Nobody campaigns against the bot that resets a password in eleven seconds. The objection is to the wall.

The experiment: Klarna, there and back

Klarna deserves its own section because it is the cleanest natural experiment the industry has. In February 2024 the Swedish fintech announced that its OpenAI-powered assistant was handling two-thirds of customer service chats - 2.3 million in the first month - and doing the work of 700 agents. The announcement was treated, reasonably, as the replacement thesis proven.

Fifteen months later the company announced it was recruiting humans again. CEO Sebastian Siemiatkowski's explanation was unusually direct for a chief executive unwinding his own strategy: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." The company began building a flexible human support pool and said that investing in the quality of human support is the way of the future.

The reversal does not mean the AI failed at what it did. It resolved millions of routine contacts and still does. What failed was the totalising version of the thesis - the claim that resolving routine contacts is the whole job. Klarna rediscovered, at scale and in public, the thing the QJE study had measured quietly: the technology's best position is behind the person, not instead of them.

What is actually left for the humans

What can AI not do in a contact centre?

The parts of the job that decide whether a customer stays. AI handles high-volume, low-ambiguity contacts well, but it does not read distress reliably, exercise judgement when systems disagree, take responsibility for an exception, or build the trust that retention conversations run on - which is why Gartner found 75% of service organisations moving agents into complex, judgement-heavy and advisory roles rather than out of the door. Klarna's CEO drew the same conclusion in public after the company's all-in AI experiment produced what he called lower quality service.

Follow that redistribution to its end point and the shape of the future agent job comes into focus. The queue that reaches a human is stripped of its easy calls. What remains is concentrated difficulty: the emotional, the ambiguous, the high-stakes, the angry. Gartner's 2026 survey found service organisations redeploying people precisely there - and 63% reducing any surplus headcount through natural attrition rather than layoffs, which tells you they expect to need the people they keep.

The fit question the industry has not asked yet

Here is the implication that gets missed while the argument stays stuck on replacement. If the routine layer goes to the machines, the human job intensifies. All-difficult queues mean more sustained emotional labour per hour, more judgement calls, fewer scripted recoveries. Some agents are built for exactly that work and are wasted on password resets. Others chose the job partly because the routine calls gave the day a rhythm they could manage.

Same job title, two different futures - and which future an agent gets depends on preferences no CV or interview reveals. This is what Sariio MAPS reads: each person's working preferences - contact load, pace, structure, autonomy - from a ten-minute survey retaken at least twice a year, with team-level maps a leader can put next to the redesigned work. Centres planning their AI transition on seat counts alone are planning half of it. The other half is knowing which people the remaining work will energise, and which it will grind - while there is still time to move people rather than lose them.

Can AI replace a call centre agent? It can replace the part of the job that was least human to begin with. What it leaves behind is the most human work the industry has ever asked its people to do - and that raises the bar on knowing your people, not the machines.

The companion pieces: the South African jobs picture is in will AI take call centre jobs in South Africa; the customer-side view is in how can AI improve customer experience; the retention economics are in why call centre agents really quit.

Sources