AI-Powered Customer Support in 2026: How Call Centers Are Changing

What the evidence shows about AI in customer support: an NBER study of 5,179 agents found 14% higher productivity (34% for novices), and Klarna's AI assistant shows both the gains and the limits of automation.

PUBLISHED // UPDATED
DURATION // 08 MIN READ
AI-Powered Customer Support in 2026: How Call Centers Are Changing
FIG.01 // VISUAL DOCUMENTATION AI-Powered Customer Support in 2026: How Call Centers Are Changing

Customer support was one of the first business functions where generative AI moved from pilots into daily operations. The case is easy to make: support work is high-volume, text-heavy, and measured by clear metrics such as resolution time and issues resolved per hour. The two best-known pieces of evidence, a large academic study of call-center agents and Klarna’s AI assistant, point in a more nuanced direction than either enthusiasts or skeptics usually admit. AI helps most when it supports people, and replacing them too quickly can lower quality.

This article reviews that evidence, explains which support tasks suit automation, and offers a framework for deciding how far to go.

The research: AI as an assistant for agents

The most rigorous study so far is Generative AI at Work by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, published as an NBER working paper. The researchers studied the staggered introduction of a generative-AI conversational assistant using data from 5,179 customer support agents. The assistant suggested responses and resources to agents; it did not talk to customers directly.

Key findings:

  • Access to the tool increased productivity, measured as issues resolved per hour, by 14% on average.
  • Novice and low-skilled workers improved by 34%.
  • Experienced and highly skilled workers saw minimal impact.

The pattern matters. The assistant appears to spread the know-how of the best agents to newer ones, helping them handle cases they would otherwise escalate or solve slowly. For experienced agents, who already know the answers, the benefit was small. That suggests AI assistance is particularly valuable for onboarding, high-turnover teams, and complex product catalogs.

The deployment: Klarna’s AI assistant

Klarna offers the best-known example of AI handling customer conversations directly. In its February 2024 announcement, the company said that in its first month its OpenAI-powered assistant:

MetricKlarna’s reported figure
Conversations handled2.3 million, two-thirds of customer service chats
Equivalent workload700 full-time agents
Resolution timeUnder 2 minutes, compared with 11 minutes previously
Repeat inquiries25% drop
Coverage23 markets, 24/7, more than 35 languages
Expected impact$40 million profit improvement in 2024

These are company figures, but they show what automation can achieve for routine inquiries such as payment status, refunds, and account questions.

The course correction

In May 2025, Klarna’s chief executive Sebastian Siemiatkowski told Bloomberg that the company’s cost-cutting in customer service had gone too far. As Fortune reported, he said: “As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality,” and that investing in the quality of human support was “the way of the future.” Klarna began recruiting freelance agents so customers would always have the option of speaking to a person, while saying it remained “AI-first.”

A Klarna spokesperson told Silicon Republic that the quality problem related to the inconsistency of outsourced human agents rather than to the AI itself, and that the assistant now handled the equivalent workload of 800 full-time agents. A Forrester analyst quoted in the same article attributed the problem to “overeager” headcount reduction, noting that AI still struggles with novel scenarios, nuanced exceptions, and highly emotional conversations.

Whichever explanation is closer to the truth, the lesson is the same: automation can handle a large share of routine volume, but the cases left for people are harder, and the quality of that human layer determines customer experience.

Which support tasks suit AI

Task typeExamplesSuitable approach
Routine, policy-boundOrder status, password resets, refund eligibilityFull automation with easy escalation
Information lookupProduct specifications, troubleshooting stepsAI answers with links to sources
Complex but structuredBilling disputes, multi-step technical issuesAI assists the agent with suggested responses
Novel or emotionalComplaints, bereavement, vulnerable customersHuman-led; AI summarizes history
High-riskFraud, legal threats, safety issuesHuman-only, with AI used for triage

The NBER findings support the middle rows: suggested responses improved resolution rates, especially for less experienced staff. Klarna’s experience supports automating the top rows while protecting the bottom ones.

Metrics that reveal quality problems

Automation rates and handle time are easy to improve and easy to misread. A balanced scorecard should include:

  • First-contact resolution: were issues actually solved, or did customers come back?
  • Repeat contact rate: Klarna highlighted a 25% drop in repeat inquiries; a rise is an early warning sign.
  • Escalation quality: when AI hands off, does the agent receive a useful summary?
  • Customer satisfaction by issue type: averages can hide poor experiences in complex cases.
  • Time to reach a person: customers who want a human should be able to reach one quickly.

Lessons from regulators and courts

Two public cases show why accuracy and access to humans matter as much as automation rates.

A chatbot’s answer can bind the company. In 2024, a Canadian tribunal ordered Air Canada to compensate a customer who had relied on its website chatbot’s incorrect description of the airline’s bereavement-fare policy, as reported by The Guardian. Air Canada argued that the chatbot was a “separate legal entity” responsible for its own actions. The tribunal rejected that argument, writing that the chatbot “is still just a part of Air Canada’s website” and that the airline “is responsible for all the information on its website.” The award was small, about C$650.88 plus interest and fees, but the principle is significant: companies cannot disclaim what their AI assistants tell customers.

Blocking access to humans can create legal risk. The U.S. Consumer Financial Protection Bureau’s June 2023 report Chatbots in consumer finance warned that financial institutions have strong incentives to replace in-person, phone, and live-chat support with chatbots, and that “deficient chatbots that prevent access to live, human support can lead to law violations, diminished service, and other harms.”

Practical implications for any support organization:

  • Ground answers in current policy documents and review them when policies change, so the assistant cannot describe outdated or invented rules.
  • Show sources to customers where possible, such as a link to the policy page the answer came from.
  • Make the route to a human obvious, especially for complaints, disputes, refunds, and vulnerable customers.
  • Log conversations so disputed answers can be reviewed.
  • Test edge cases regularly, including unusual policies like bereavement fares, cancellations, and exceptions, where assistants are most likely to be wrong.

Agent-assist versus customer-facing bots

The research and the legal cases point toward a useful distinction. In the Brynjolfsson, Li, and Raymond study, the AI suggested responses to human agents, who decided what to send. The authors report that, beyond the productivity gain, AI assistance improved customer sentiment, increased employee retention, and may have helped workers learn, with a human still accountable for every reply.

Customer-facing bots remove that human checkpoint, which is where the Air Canada error occurred. Many support teams therefore start with agent-assist tools, which carry lower risk and build a tested knowledge base, and expand self-service only for question types where the assistant has proved accurate.

Practical guidance

  1. Start with agent assistance. The NBER evidence shows clear productivity gains without exposing customers to AI errors directly.
  2. Automate routine intents first. Choose high-volume questions with clear policy answers.
  3. Always offer a path to a person. Klarna’s revised approach makes this explicit.
  4. Keep and invest in experienced agents. They handle the hardest cases and their knowledge is what AI tools learn from.
  5. Ground answers in your own documentation. Linking AI responses to help-center articles reduces fabricated answers.
  6. Review conversations regularly. Sample AI-handled conversations weekly, especially those marked resolved.

The workforce question

Support automation raises real questions about jobs. The research suggests the near-term effect is uneven: newer agents benefit most from AI assistance, while routine volume shrinks. Organizations can use that to reduce burnout and improve training, or to cut headcount quickly. Klarna’s experience suggests that cutting too far, too fast, risks lower quality and a costly reversal.

Bottom line

The strongest evidence points to a hybrid model. AI assistants raised support productivity by 14% on average and 34% for novices in a study of more than 5,000 agents, and automated assistants can handle most routine conversations, as Klarna showed. The same Klarna case also shows the limits: when cost becomes the main goal, quality suffers. Companies that automate routine work, equip agents with AI, and guarantee access to skilled people are most likely to see lasting benefits.

Sources