The future of customer service from handling requests to steering AI agents

I

What research measures about AI in the contact center

The reference study covers 5,172 technical-support chat agents at an American software company, progressively equipped with a generative AI tool that suggests replies in real time¹. Published in 2025 in the Quarterly Journal of Economics by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, it measures an average 15% increase in the number of requests resolved per hour¹. The working paper circulated by the NBER in April 2023 reported 14% on average and 34% for novice and low-skilled agents². The version published in the journal is now the reference.

The most instructive result is not the average. The least experienced and least skilled agents gain about 30% in requests resolved per hour, and those in the least-skilled quintile 36%¹. Agents equipped with the tool reach, after two months of tenure, the level of their unequipped colleagues with more than six months of tenure¹. Conversely, the most experienced gain little in speed and even see the quality of their exchanges drop slightly¹.

The authors offer an explanation: the tool captures the practices of the best agents and makes them available to everyone¹. In other words, AI spreads know-how that already existed in the team. That know-how therefore remains the raw material of the system.

Two other results bear directly on working conditions. Customers are more courteous to advisors who have the tool and ask less often to speak to a manager¹. Attrition among recent advisors, those with less than six months of tenure, falls by about 10 points, a drop of roughly 40% from a baseline attrition rate of 25% in that group¹. The authors suggest reading this last result with more caution than their main findings¹.

They also state an explicit limit: the study covers a single company, over the medium term, and is not designed to measure the overall effects of generative AI on jobs or wages¹. That is a useful caveat: it means drawing neither a promise nor a threat about jobs from it.

II

What Gartner predicts: more autonomy, not service without humans

On March 5, 2025, Gartner predicted that by 2029, agentic AI will autonomously resolve 80% of common customer service issues, without human intervention, leading to a 30% reduction in operational costs³. The firm distinguishes these AI agents, which can act on the customer’s behalf, from earlier tools that were limited to drafting or summarizing³.

Three months later, the same firm published a forecast that sheds light on its scope. By 2027, half of the organizations that planned to sharply reduce their customer service headcount will abandon that plan⁴. Gartner relies on a March 2025 survey of 163 customer service leaders, 95% of whom plan to keep human advisors⁴.

Later surveys point the same way. Of 321 leaders surveyed in fall 2025, only 20% say they have reduced advisor headcount because of AI, while 55% are holding headcount steady while absorbing rising contact volumes⁵. In the same survey, 42% are hiring specialized profiles, which Gartner calls AI strategists, conversational AI designers, and automation analysts⁵.

A second analysis of this survey, published in April 2026, sharpens the trend. 85% of leaders are adding tasks and responsibilities to advisor roles, and 75% are moving advisors into entirely new roles⁶. 31% have carried out or plan to carry out AI-related advisor layoffs by the first quarter of 2027, while 63% are reducing headcount gradually, through natural attrition⁶. On the customer side, Gartner cites in the same publication a separate survey, conducted in January and February 2025 among 5,801 customers in the United States, in which 54% trust a human advisor more than AI for a product or service recommendation, versus 32% who trust AI more⁶.

The overall picture is nuanced. Automation is advancing, headcounts are being reshaped, and in some organizations they are shrinking. But the available data describes a transformation of roles more than the elimination of the profession. Gartner puts it this way through Eric Keller: leaders must choose between doing the same work at lower cost or redeploying advisors to roles that AI does not replace⁶.

III

Klarna: what the company actually announced

On February 27, 2024, Klarna announced that its AI assistant, powered by OpenAI, had handled 2.3 million conversations in one month, two-thirds of its customer service chats⁷. The company reported work equivalent to that of 700 full-time agents, customer satisfaction on par with human advisors, a 25% drop in repeat inquiries, and resolution time down from 11 minutes to under 2 minutes⁷. The press release specified that customers could still choose to talk to a human advisor⁷.

In May 2025, its CEO Sebastian Siemiatkowski told Bloomberg that cost had been too predominant an evaluation criterion in this organization, resulting in lower quality⁸. According to Bloomberg, Klarna was then preparing a hiring drive so that its customers would always be able to talk to a real person, and was testing a model of advisors working remotely on a flexible basis⁸.

This episode needs to be read precisely. Klarna did not give up on AI in its customer service. The company wanted to guarantee lasting access to a human after finding that an organization driven first by cost degraded quality. That is exactly the trade-off the Gartner surveys describe.

IV

The emerging role: from handling to steering

What follows is an analysis, built on the facts above. If AI agents resolve a growing share of routine requests on their own³, human work concentrates on two areas: the conversations the AI agent should not handle alone, and the running of the AI agent itself.

What remains deeply human

Complex or rare cases first. The study by Brynjolfsson, Li, and Raymond shows that AI’s contribution is strongest on relatively rare problems, those on which advisors have the least experience¹. We can infer that infrequent situations remain those where human judgment is most called upon, with AI in support.

Next, complaints, tense situations, and decisions that commit the company: a goodwill gesture, an exception to a rule, a dispute. These moments call for empathy, accountability, and an authority the customer recognizes. Gartner’s data on customer trust in human advisors for recommendations points the same way⁶. Gartner sees this as confirmation that human involvement remains important in complex, high-stakes, or advisory exchanges⁶.

The roles that are emerging

Gartner cites three families of profiles already being hired: AI strategists, conversational AI designers, automation analysts⁵. The table below offers, as an analysis, a breakdown of these functions into concrete duties.

These titles and backgrounds are a working hypothesis, not measured data. It draws on the interpretation offered by the study’s authors: the tool spreads the practices of the best advisors¹. Those who know what a good answer looks like are therefore best placed to teach the AI agent to give it.

The augmented advisor day to day

For the advisor who stays in direct contact with customers, the change lies in the nature of the requests. Simple, repetitive questions gradually move to the AI agent. What remains are the conversations that resisted automation, often longer and more emotionally charged. The advisor receives these conversations with their history, which requires a human handoff that passes on the context instead of forcing the customer to repeat everything.

This shift has a cost if it is not supported: a day made up only of difficult cases is more tiring. The management challenge is to recognize this shift in how work is organized, in targets, and in pay, rather than measuring the augmented advisor with yesterday’s advisor’s volume metrics.

V

Turnover, working conditions, and upskilling

Turnover is a structural feature of the job. Brynjolfsson, Li, and Raymond cite industry estimates that 60% of contact center advisors leave their position every year¹. They also describe work that is often difficult, exposed to the hostility of dissatisfied customers¹.

In this context, AI can cut both ways. Well designed, it shortens the learning curve, improves the way customers treat advisors, and retains more newcomers¹. Designed solely to cut costs, it can degrade service quality, as Klarna’s CEO acknowledged⁸. The difference has less to do with the technology than with organizational choices.

Upskilling therefore becomes a central lever. An advisor who understands how the AI agent reasons, where it goes wrong, and how to correct it becomes a contributor to the quality of the system. That is also what the European regulatory framework now asks for.

VI

What the AI Act requires on AI literacy

Article 4 of the EU AI regulation, applicable since February 2, 2025, deals with AI literacy⁹. In its original version, it required providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and other people using these systems on their behalf, taking into account their technical knowledge, experience, education and training, and the context of use⁹.

The “digital omnibus” regulation on AI, Regulation (EU) 2026/1744, published in the Official Journal on July 24, 2026 and in force since July 27, 2026, rewrote this article¹⁰. Providers and deployers must now take measures to encourage the development of AI literacy among these same people, without being required to guarantee a specific level of literacy for each individual¹⁰. The requirement becomes an obligation of means. It has not been removed.

For a company deploying AI agents in its customer relationships, the consequence is concrete: it is up to you to take measures to develop AI literacy among the teams who use or supervise these systems. In practice, this means training, and keeping a record of the actions taken makes it possible to demonstrate them.

VII

What this means for your organization

Preparing your teams does not start with a tool. It starts with an explicit decision about what you want to do with the time freed up. Gartner recommends not presenting AI projects solely from a headcount-reduction angle and planning for the new roles⁵. Here is a five-step approach.

1. Map requests before mapping roles. Sort your contact reasons by two criteria: how often they occur and how much they matter to the customer. Frequent, low-stakes requests are the first candidates for the AI agent. Rare or high-stakes requests stay human, with AI in support.

2. Identify your best advisors as future designers. According to the study, it is their know-how that the tool spreads to other advisors¹. Involve them from the start in writing the rules, tone, and edge cases. It is the most direct path to the roles of conversation designer and AI quality manager.

3. Train everyone, at different levels. Article 4 of the AI Act covers your staff and the other people who use AI systems on your behalf¹⁰. A common foundation for everyone (what the AI agent does, its limits, when to step in), then an in-depth track for those who supervise or train it. Document these actions.

4. Revisit your metrics. If the AI agent handles simple requests, volume handled per advisor stops being a good indicator. Measure resolution of complex cases, the quality of human handoffs, satisfaction after escalation, and contribution to improving the AI agent.

5. Design the human handoff as a journey in its own right. Define when the AI agent hands over, with what information, and who decides. A well-designed handoff protects both the customer, who does not have to repeat themselves, and the advisor, who receives a complete file.

These steps rest on a simple principle: your teams stay in control of what the AI agent says and does. That is what distinguishes a managed transformation from cost-cutting imposed on you.

VIII

Where QLOA stands

IX

Frequently asked questions

What is the future of customer service with AI?

Gartner predicts that by 2029, agentic AI will autonomously resolve 80% of common customer service issues³. Its surveys also indicate that most leaders plan to keep human advisors and that many are expanding their responsibilities toward higher-value tasks⁴⁶. The emerging future of customer service is therefore hybrid: automation of routine requests and advisors refocused on what requires judgment.

Will AI replace customer service advisors?

The available data does not describe a total replacement. According to a Gartner survey conducted in fall 2025 among 321 leaders, 20% have reduced advisor headcount because of AI, and 55% have kept it steady while absorbing more volume⁵. Some organizations are shrinking their teams, often through natural attrition⁶. The profession is being transformed more than it is disappearing.

What are the new customer relations jobs?

Gartner notes that 42% of organizations are hiring specialized profiles: AI strategists, conversational AI designers, automation analysts⁵. On analysis, we can add AI agent supervision, conversation analysis, and answer quality control. Experienced advisors, whose practices AI spreads¹, are well placed to move into these functions.

How much productivity does AI bring to a contact center?

The study by Brynjolfsson, Li, and Raymond published in the Quarterly Journal of Economics in 2025 measures an average 15% increase in issues resolved per hour across 5,172 advisors¹. The gain reaches about 30% for the least experienced and lowest-skilled, and remains small for the most experienced¹. The study covers a single company.

Does the AI Act require training advisors on AI?

It imposes an obligation of means, which in practice means training. Article 4 of the EU AI regulation, applicable since February 2, 2025, was rewritten by Regulation (EU) 2026/1744, in force since July 27, 2026⁹¹⁰. Companies that deploy AI systems must take measures to encourage AI literacy among their staff, without having to guarantee a specific level for each person¹⁰.

Why did Klarna announce it was hiring human advisors?

In May 2025, Klarna’s CEO explained to Bloomberg that cost had weighed too heavily in the organization of its customer service, at the expense of quality⁸. According to Bloomberg, the company was preparing a hiring drive to guarantee its customers the possibility of talking to a real person, while continuing to use AI⁸. In February 2024, its AI assistant was already handling two-thirds of conversations⁷.

X

Sources

1. 1

   Erik Brynjolfsson, Danielle Li, Lindsey Raymond, “Generative AI at Work”, The Quarterly Journal of Economics, vol. 140, no. 2, pp. 889-942, 2025. https://doi.org/10.1093/qje/qjae044 (author version: arxiv.org)

2. 2

   Erik Brynjolfsson, Danielle Li, Lindsey Raymond, “Generative AI at Work”, NBER Working Paper No. 31161, April 2023. nber.org

3. 3

   Gartner, “Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029”, March 5, 2025. gartner.com

4. 4

   Gartner, “Gartner Predicts 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI”, June 10, 2025. gartner.com

5. 5

   Gartner, “Gartner Survey Finds Only 20% of Customer Service Leaders Report AI-Driven Headcount Reduction”, December 2, 2025. gartner.com

6. 6

   Gartner, “Gartner Survey Finds 85% of Service and Support Leaders are Expanding Human Agent Responsibilities Despite Expectations of Mass AI Layoffs”, April 28, 2026. gartner.com

7. 7

   Klarna, “Klarna AI assistant handles two-thirds of customer service chats in its first month”, February 27, 2024. klarna.com

8. 8

   Bloomberg, “Klarna Slows AI-Driven Job Cuts With Call for Real People”, May 8, 2025. bloomberg.com

9. 9

   Regulation (EU) 2024/1689 of the European Parliament and of the Council of June 13, 2024 laying down harmonised rules on artificial intelligence, Article 4 and Article 113. eur-lex.europa.eu

10. 10

    Regulation (EU) 2026/1744 of the European Parliament and of the Council of July 8, 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of the harmonised rules on artificial intelligence (Digital Omnibus on AI), OJ L of July 24, 2026, Article 1, point 5, and Article 4. data.europa.eu