How to put an AI customer service agent to work

A customer service agent can settle routine questions around the clock, but customers are wary and the business is bound by what it says. A step-by-step plan for doing it well.

Customer service is where most businesses meet AI agents first, and with good reason. The questions are frequent, many are routine and the answers usually exist already in your help pages, order system and policies. Done well, an agent answers at any hour, in several languages, and hands the hard cases to your team with the details already gathered. Done badly, it frustrates customers and makes promises your business must keep.

Start by knowing your customers' mood. Gartner surveyed 5,728 customers in December 2023 and reported in July 2024 that 64 per cent would prefer companies did not use AI for customer service. Gartner said over half would consider switching to a competitor if a company used AI for service. This does not mean you should not use an agent. It means the agent must be clearly useful, honest about what it is and quick to bring in a person.

Step one is to pick the questions. Pull a month of tickets, chats and emails and sort them by type. Look for the questions that are frequent, have a clear right answer and need information you hold in a system: where is my order, how do I return this, what are your opening hours, how do I change my booking. Leave complaints, vulnerable customers, disputes and anything involving judgement about money with people at first.

Step two is to fix the knowledge. An agent answers from what you give it. If your help pages are out of date or contradict each other, the agent will repeat the errors with confidence. The Air Canada case is the warning. In February 2024 the Guardian reported that a Canadian tribunal ordered the airline to compensate a customer misled by its chatbot about bereavement fares, rejecting Air Canada's argument that the chatbot was responsible for its own actions. Assume your business will be held to whatever the agent says.

Step three is to choose a tool. Most help desk vendors now include an AI agent. At the time of writing (November 2026), Intercom's pricing page said its Fin agent costs 0.99 US dollars per outcome, and Zendesk's pricing page said it charges per automated resolution beyond a small allowance included in each plan. Salesforce, Microsoft and others offer agents priced by conversation or credit. If you already use a help desk, start by testing the agent it offers, because it can already see your tickets and customer records.

Step four is to set the rules for handing over. Decide in advance when the agent must pass a conversation to a person: when the customer asks for one, when they are upset, when the question is outside its list, when money above a set limit is involved, or when it is not confident. Make sure the hand-over carries the conversation history, so the customer does not have to repeat themselves. A smooth hand-over is the single biggest factor in whether customers accept the agent.

Step five is to tell customers they are talking to AI. In several editions this is becoming a legal expectation as well as good practice. Australia's privacy regulator, the OAIC, says in its guidance on commercially available AI products that public-facing chatbots should make clear they are AI. Under the EU AI Act, which affects UK, US and other businesses serving EU customers, Article 50 requires people to be told they are interacting with an AI system unless it is obvious, from 2 August 2026 according to the European Commission. Elsewhere, including South Africa, Nigeria, Singapore and the Gulf, data protection laws set rules on how the customer data the agent uses is collected and shared.

Step six is to test and launch gradually. Run the agent on past questions and have your best service staff grade the answers. Then switch it on for one channel or one share of traffic. Read a sample of real conversations every day in the first weeks. Track the share of conversations it resolves, the share handed to people, customer satisfaction on both, and complaints that mention the agent.

The Klarna story is a reminder to keep checking. Klarna said in February 2024 that its AI assistant handled two-thirds of its customer service chats. By May 2025, Computing reported that the company was hiring human service staff again after its chief executive said quality had fallen. Volume handled is not the same as customers served well.

What to do on Monday

1. Export last month's customer conversations and sort them into types. Mark the five most common routine questions.
2. Check that your help pages and policies give one clear, current answer to each of those five questions.
3. Ask your current help desk vendor for a trial of its AI agent, limited to those five questions.
4. Write down the hand-over rules, including that any customer who asks for a person gets one.
5. Add a short line at the start of every AI conversation saying the customer is talking to an AI assistant.

More on this topic: AI agent guides

Sources

• Gartner: Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn't Use AI For Customer Service
• The Guardian: Air Canada ordered to pay customer who was misled by airline's chatbot
• Intercom: Intercom Pricing | Plans for every team size
• Zendesk: Zendesk Pricing Plans
• OAIC: Guidance on privacy and the use of commercially available AI products
• European Commission: AI Act | Shaping Europe's digital future
• Klarna: Klarna AI assistant handles two-thirds of customer service chats in its first month
• Computing: Meat back on the menu as Klarna rows back on AI