From chatbot to AI agent: what to expect in customer service by 2029

Chatbot or AI agent?

A chatbot answers questions, while an AI agent also carries out the steps in your systems.

A classic chatbot answers questions. It looks up an answer in a script or a knowledge base and hands the customer over when it gets stuck. An AI agent goes a step further: it understands the request, plans the steps and carries them out in your systems. It can check an order, start a refund, change a delivery address or cancel a subscription, and then confirm what it did. That shift matters for customer service, because many contacts are about getting something done, not about getting an answer.

Agents are getting better at this kind of work, at least in tests. According to Stanford HAI’s 2026 AI Index Report, AI agents went from 12% to around 66% task success on OSWorld, a test of real computer tasks. They still fail roughly 1 in 3 attempts on structured benchmarks. For customer service, that means agents are useful for clear, repeatable cases, and still need a safety net. AI in general is becoming part of normal operations fast: Stanford HAI reports that organisational adoption of AI reached 88%, although that does not mean 88% of companies use agents in customer service. We first wrote about this step in beyond chatbots: the exciting future of generative AI agents.

Most questions that reach a customer service team are already known, and agents are now smart enough to answer them faster and better. People become valuable precisely in the cases that AI cannot (yet) handle.

What do the numbers show?

According to its own figures, Klarna’s AI assistant handled 80% of customer service chats in 2025.

The best documented example comes from the payments company Klarna. Its numbers are its own, so read them as company figures, not independent research. In a press release from February 2024, one month after launch, Klarna reported that its AI assistant:

• had 2.3 million conversations, two-thirds of its customer service chats;
• did the equivalent work of 700 full-time agents;
• was on par with human agents on customer satisfaction;
• led to a 25% drop in repeat inquiries;
• helped customers resolve errands in less than 2 minutes, compared to 11 minutes before;
• worked in 23 markets, 24/7, in more than 35 languages.

A year and a half later, in its registration statement for the US stock market (Form F-1, September 2025), Klarna wrote that the assistant handled 69% of customer service chats in the twelve months ended 30 June 2025, according to its chat log data. It said the assistant did the work of over 700 full-time agents and delivered approximately $39 million in cost savings in 2024. Queries were resolved in two minutes, compared to the 12-minute average for human agents in 2024. Since launch, the assistant had 25.3 million conversations. In its annual report for 2025 (Form 20-F, February 2026), Klarna put the share at 80% of customer service chats in 2025, the equivalent work of over 850 full-time agents and approximately $59 million in cost savings; since launch the assistant had 31 million conversations. Two details in the same filing are worth noting. First, Klarna continues to offer all customers the option to talk to a human, because “certain consumers may nevertheless prefer to interact with human representatives”. It calls this a “dual-track approach” of scalable AI combined with high-quality human support. Second, the numbers are about chats. They say nothing about the hardest cases, which tend to end up with people.

Who is liable for errors?

In the Canadian case Moffatt v. Air Canada, the company was held responsible for its chatbot.

One Canadian decision points to the company. In Moffatt v. Air Canada (2024 BCCRT 149), decided on 14 February 2024, a customer had relied on the airline’s website chatbot. The chatbot suggested he could apply for a bereavement fare after his trip. Air Canada’s actual policy did not allow that. Air Canada argued it could not be held liable for information from its chatbot. The British Columbia Civil Resolution Tribunal called that “a remarkable submission” and wrote: “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.” The tribunal found that Air Canada did not take reasonable care to make sure its chatbot was accurate, and ordered it to pay a total of $812.02, including $650.88 in damages.

The amount was small, and it is one decision by a Canadian tribunal, not a rule for every country. Still, an agent that can act, not just talk, raises the stakes further: a wrong answer becomes a wrong refund, a wrong cancellation or a wrong promise.

What do EU rules require?

Customers must know they are talking to AI, and when they take out a financial service online they can ask for a human.

Several rules already apply, or will apply soon. Customers must know they are talking to AI. Under Article 50 of the AI Act, AI systems that interact directly with people must be designed so that people are informed they are interacting with an AI system, unless that is obvious from the context. That information must be clear and given at the latest at the time of the first interaction. According to the European Commission, the transparency rules come into effect in August 2026. The Commission gives chatbots as the example: humans should be made aware that they are interacting with a machine. For financial services contracted online, customers can ask for a human. A new EU directive on financial services sold at a distance, Directive (EU) 2023/2673, applies from 19 June 2026, according to the European Commission. Each country writes it into its own law. Ireland’s version, for example, S.I. No. 309/2026, gives a consumer using a trader’s online interface “a right to request and to obtain human intervention” before signing, and in justified cases afterwards.

Some decisions are high-risk. Most customer service bots fall under the transparency rules. But the AI Act treats certain uses as high-risk, such as AI that decides access to essential private and public services, for example credit scoring. According to the Commission, strict obligations for these systems apply from 2 December 2027, including human oversight and logging of activity. Staff need to understand the tools. Article 4 of the AI Act asks organisations that use AI to support the AI literacy of their staff. The European Commission explains that this has applied since 2 February 2025, and that national authorities supervise it from 2 August 2026. For customer service teams that work alongside agents, this is practical, not just legal: people who understand an agent’s limits catch its mistakes.

Customer service in 2029

The EU rules are fixed by the end of 2027; the rest is expectation and forecast.

Here it helps to separate what is fixed from what is a forecast. What is fixed. By the end of 2027, the rules above all apply in the EU: disclosure of AI, a right to request human intervention in financial services sold online, and strict obligations for high-risk uses. Any agent a company builds now has to work within them in 2029.

What the trend suggests. Agents are improving quickly in tests, and Klarna already handles a majority of its chats with AI. If both trends continue, it is reasonable to expect that by 2029 routine errands will more often be resolved end to end by an agent, while people focus on exceptions, complaints, vulnerable customers and the cases where judgement and empathy matter. That is an expectation, not a measurement. What is a forecast. Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention (press release, 5 March 2025). That is a forecast by an analyst firm, not a measurement. Customer expectations are part of the picture too. The first wave of generative AI already made automated service sound far more human, as we described in how generative AI can create customer care that’s so human-like, it’s scary. Agents add the ability to act. The wider shift is explored in how autonomous AI will reshape everything you do.

From chatbot to AI agent

Start with one type of errand, say it is AI and always offer a route to a person.

A few principles follow directly from the examples and rules above:

• Start with one type of errand.
  Pick a high-volume, low-risk case with clear rules, such as order status or address changes. Not complaints or credit decisions.
• Say it is AI, at the start.
  Not hidden in the terms, but in the first message. That is required in the EU, and it builds trust.
• Always offer a route to a person.
  Klarna does it by choice. In financial services sold online, customers have a legal right to ask for one. Make the route easy to find.
• Limit what the agent may do.
  Decide which actions it can take on its own, which need a confirmation, and which stay with people. The cost of a mistake decides, not the demo.
• Check what it says about your policies.
  The Air Canada case was about a wrong policy answer. Test the agent on your own rules, especially exceptions.
• Measure resolution, not deflection.
  A chat that ends is not a solved problem. Repeat contacts, as Klarna reported, are a better signal.
• Train the team.
  People who work with agents need to know what they do well, where they fail and how to step in. In my view, this is the most important step: your team only really gets to know an agent by trying it out, starting with that one type of errand. Give your staff freedom here, use their enthusiasm and look critically, together, at repeat contacts and mistakes.

Frequently asked questions

Will AI agents replace service staff?

At one company they already do a lot: Klarna says its assistant did the work of over 850 full-time agents in 2025, by its own figures. At the same time, Klarna still offers every customer the option to talk to a human. The likely change is in the work itself: fewer routine contacts, more complex and sensitive cases, and a new task of supervising agents.

Must you tell customers it is AI?

In the EU, yes, from August 2026, unless it is obvious from the context. Say it in the first message, not in the terms.

Is customer service AI high-risk?

Usually not. It becomes high-risk when the AI decides on access to essential services, such as credit; those obligations apply from 2 December 2027.

Where should you start?

Map your processes first, then start with one low-risk reason for contact.

Start with your processes, not with a system. List the steps of a customer contact, from recognising the question to arranging a refund, and ask per step: does AI do this, or does it belong with your staff? Then look at your top ten reasons for contact and pick the one that is most rule-based and least risky. Write down what the agent may do on its own, when it hands over to a person and how customers reach that person. Tell customers it is AI. Review the results after a month, with your service team at the table. That side of change, getting people moving, is the subject of the keynote on the organisation of the future. Agents in customer service are also one of the most visible examples of agentic AI at work, and a regular theme in the keynote on agentic AI.