AI agents in customer service: what they can realistically take over, where people are needed and where to start

An AI agent is a program based on a language model, the same kind of technology as ChatGPT. Unlike a simple chat with pre-written answers, an agent understands a question asked in the customer’s own words. It can also look into your systems itself, for example to check an order’s status in the ERP system or open a complaint ticket.

In contact centres, the word “agent” usually also means the employee on the phone. To avoid confusion, from here on we use “agent” for the program and “customer service advisor” for the person.

An agent knows nothing about your company on its own. It answers based on the documents and data it has been given access to: the price list, terms and conditions, instructions, order history. If these sources are out of date or contradictory, the agent will repeat their errors, only faster and more politely.

The simplest rule is this: the agent acts alone where the answer is unambiguous and recorded in the system. Where judgement, a decision or empathy is needed, the agent prepares material for a person.

Take a hypothetical parts distributor with 90 employees. Its six-person customer service team receives about 3,000 emails and phone calls a month. More than a third are questions about order status, delivery dates, product availability and copies of invoices – exactly what an agent can check on its own.

The table below sets out typical uses and how work is divided between the agent and people.

A language model can give an untrue answer that sounds very convincing. For a question about opening hours, that is a minor slip. For a promise of a refund or a delivery date, it is a commitment that is hard to back out of.

That is why it is worth setting the boundaries in writing before the agent talks to its first customer. A good rule: the agent does not make decisions that cost the company money or change a contract. Below are the situations in which a person should take over the conversation.

The best-documented effect concerns supporting customer service staff, not replacing them. In a study published in the Quarterly Journal of Economics in 2025, an AI assistant suggested replies to more than five thousand advisors. The number of issues resolved per hour rose by 15% on average, and new and the least experienced staff gained the most. For the most experienced, the gain was small.

The lesson for your company: an assistant agent shortens onboarding for new advisors, because it gives them access to knowledge that only the longest-serving staff had before. This is often more important than saving minutes on an individual request.

Klarna’s example shows the other side. In February 2024, the company announced that its AI assistant was handling two thirds of customer chats. In 2025, its CEO admitted that cost had been given too much weight, which lowered the quality of service. The company said it would hire advisors again, so that customers could always speak to a person.

Do not base your decision on percentages from marketing materials. Calculate using your own data from the last three months: how many requests come in, what proportion are simple questions and how long each one takes.

Below is the calculation for the hypothetical distributor from the previous section. All the values are assumptions – replace them with your own data.

In the example, the result is about 52 hours a month, roughly a third of a full-time post. That is real relief for a six-person team, but not a reason for redundancies. The time saved usually goes into handling difficult cases faster and growing sales without hiring more people.

From the value of the recovered time, deduct the costs: platform or usage fees, setting up connections to your systems and ongoing work on the agent’s knowledge. If little is left after this calculation, it is better to start by tidying up the knowledge base and simple reply templates.

Also count what does not show up in hours. The agent answers at night and at weekends, and a customer who gets their order status straight away does not call a second time. But do not add these effects to the calculation without measuring them – check them in the pilot.

Article 50 of the AI Act applies from 2 August 2026. An AI system that talks to people must tell them it is AI, unless this is obvious. A clear message at the start of the chat or phone call is enough, for example: “You are talking to a virtual assistant. You can ask for an advisor at any time.”

Be careful with emotion recognition. The AI Act treats assessing a customer’s mood from features of their voice as an emotion recognition system. A company that uses one must say so, and such a system may be classified as high-risk, with obligations from 2 December 2027. Assessing your own advisors’ emotions in this way has been prohibited since 2 February 2025.

In Poland, the AI Act is supervised by the Commission for the Development and Safety of Artificial Intelligence (KRiBSI), which can inspect companies and impose fines from 28 October 2026. The GDPR applies separately. You need a data processing agreement (DPA) with the provider, a set retention period for conversations and information for customers about who processes their data.

If you serve consumers, an agent that logs complaints must keep track of the deadline. Under Article 7a of the Polish Consumer Rights Act, the business has 14 days to respond, after which the complaint is deemed accepted. An automatic acknowledgement of receipt is not a response to the complaint.

The least risky route is to start with one channel and a narrow scope. The channel is usually an email inbox or website chat, because these are easier to control than the phone. The steps below make up a pilot lasting a few weeks.

Rollouts fail more often because of wrong assumptions than because of the technology. Here are the mistakes worth knowing about before you start: