Klarna’s AI agent resolves customer cases in 2 minutes, down from 11

Key Points

1. Klarna deployed an AI agent handling customer inquiries in 35 languages, resolving cases in roughly 2 minutes.
2. AI agents differ from passive software by perceiving, reasoning, acting and retaining memory across interactions.
3. Researchers stress the systems match patterns statistically rather than understand, and fail outside training conditions.

The latest:

An AI agent deployed by payments company Klarna resolves customer service inquiries in an average of 2 minutes, compared with 11 minutes for human representatives, and handles requests in 35 languages at once. The company reported satisfaction scores comparable to human agents. It is one of several deployments cited as evidence that autonomous software has moved from demonstration to daily commercial operation.

Details:

• The Klarna numbers: The payments firm’s agent accesses account information, processes refunds, and reads emotional cues in messages to adjust its tone, according to the account. Beyond the 2-minute average resolution time and 35-language coverage, customers frequently do not realize they are dealing with software until a conversation ends.
• Healthcare deployment: At Stanford University, an agent called CheXNet analyzes chest X-rays with accuracy described as exceeding that of many radiologists. It identifies pneumonia, highlights the specific regions of concern and attaches confidence levels, flagging unusual patterns for human review rather than resolving them alone.
• Financial management: Wealthfront uses AI agents to manage billions of dollars in client assets, rebalancing portfolios against individual goals, harvesting tax losses and adjusting strategy after life events clients report. One feature predicts when a client will need cash from spending patterns and shifts money into more liquid holdings.
• The scientific case: DeepMind’s AlphaFold addressed protein structure prediction, a problem that previously required years of manual work per protein. The system now predicts structures in minutes with accuracy described as remarkable, compressing a core step in drug discovery and biological research.
• How they work: Four components define an agent: perception, through text analysis, image recognition, speech recognition or data parsing; reasoning, typically via machine learning models and large language models; action, from sending emails to controlling devices; and memory, both short-term context and long-term learning.
• The stated limits: Agents are characterized as sophisticated pattern-matching systems lacking genuine comprehension, creativity and common-sense reasoning. An agent can approve or deny millions of loan applications consistently, but cannot weigh why fairness matters in lending or assess the ethics of decisions keyed to zip code data.
• Failure conditions: The systems are described as brittle: strong inside their training parameters, unreliable at the edges. A medical agent trained on adult patients could misdiagnose a child presenting the same symptoms, and a trading agent optimized for normal conditions could make catastrophic decisions during a market crash.
• Bias exposure: Because agents learn from data carrying existing social inequalities, a hiring agent trained on records from male-dominated industries may discriminate against women, and a healthcare agent trained largely on white patients may generate weaker recommendations for minority patients.
• The accountability gap: When an agent misdiagnoses a patient, makes a bad investment or wrongly denies a loan, responsibility is contested between the programmer, the deploying company and the system itself. Courts in several jurisdictions are already working through these questions.
• Existing presence: Agents already operate in streaming recommendations, email spam filtering, hospital vital-sign monitoring that flags complications earlier than traditional methods, high-frequency trading, and NASA rovers navigating Mars autonomously while Earth is 20 minutes away by radio.

Background:

Earlier tools extended physical capability — strength, sight, reach. Agents are presented as the first category to extend cognitive capacity and decision-making, shifting the user relationship from issuing commands to setting intentions that software then executes.

Between the lines:

The Klarna, Wealthfront and CheXNet deployments share a structure worth noting: each keeps a human at the decision boundary — CheXNet escalates unusual patterns for review, Wealthfront acts on life events clients themselves report. Read alongside the brittleness problem, where a child presenting adult symptoms can defeat a trained medical system, that design looks less like caution than a working requirement.

What’s next

Watch whether court rulings on liability for AI-driven decisions establish where responsibility sits, and whether deployed agents hold their reported accuracy when market or clinical conditions move outside their training data.