AI agents for customer service: what they can do, what they should not, and how to roll them out

The chatbots of a few years ago ran on option trees: “press 1 for orders, 2 for invoices”. If the question did not fit, the customer ended up frustrated. Today’s AI agents understand what customers write in their own words, look things up in your information and systems, and can genuinely resolve a share of enquiries.

But a badly designed agent can invent a returns policy or trap an angry customer in a loop. This guide covers what an AI agent does well in customer service, what it should not do, how to design it to be reliable, what the law says and how to roll it out without taking risks.

What an AI agent in customer service is

An AI agent is an assistant built on a language model that, besides talking, looks up information and takes actions within set limits. For example, it can:

• answer questions from your documentation (opening hours, terms, how-to guides);
• check the status of an order or a ticket in your systems;
• book or change an appointment;
• classify an enquiry, summarise it and pass it to the right person with full context.

The difference from a classic chatbot is not just that it “talks better”: it can work with your real data. Which is exactly why you have to design carefully what data it can reach and what it can do.

What it does well

• Repeated questions. Opening hours, shipping terms, how to return something, how to use a product. They tend to be a large share of the volume and have clear answers.
• Status questions. “Where is my order?”, “what’s happening with my ticket?”. The agent checks the system and answers with the real data.
• Simple tasks. Rescheduling an appointment, updating a delivery address, requesting a document.
• Out-of-hours cover, even if only to handle the simple things and leave the rest ready.
• Preparing work for your team. Classifying, summarising and drafting a reply for a person to review and send. The safest way to start.

What it should not do (yet)

• Decide sensitive matters. Large refunds, exceptions to your terms, complaints with legal consequences. It can gather the information; a person decides.
• Handle a very angry customer with no way out. If the conversation gets tense, it should pass to a person as soon as possible.
• Answer beyond your information. An agent that “makes up” a policy or a deadline is worse than no agent. It must stick to your documentation and say it does not know when the answer is not there.
• Give medical, legal or financial advice unless that is exactly your service and it is supervised.

How to design it to be reliable

A reliable agent has four parts, and none of them is the AI model itself:

1. An organised knowledge base. The agent answers from your documents, retrieved according to the question (known as RAG, retrieval-augmented generation). If the documentation is out of date or contradicts itself, so will the agent. Tidying it is often the first job of the project.
2. Tools with minimal permissions. If the agent looks up orders, it should only be able to read, not change them. If it can reschedule appointments, it should not cancel them without confirmation. Any action with consequences is better with a person’s approval.
3. Clear handover rules. When the conversation passes to a person, and how (see below).
4. Logging and review. Every conversation is stored, and someone reviews a sample every week to catch errors, unanswered questions and missing documentation.

Handing over to a person

This is what most separates a good agent from a frustrating one. It should hand the conversation to a person:

• as soon as the customer asks, without making them insist;
• when it detects frustration or the customer repeats the same question;
• when the enquiry touches a sensitive subject or is beyond its scope;
• when it cannot find the answer in your information.

And it should do it well: with a summary of the conversation so the customer does not have to repeat themselves, and a clear statement of what happens next (“someone from the team will reply between 9am and 6pm”).

What the law says

• Saying it is AI, in the EU. Since 2 August 2026, the EU AI Act requires AI systems that interact with people to be designed so that people are told they are dealing with AI, unless it is obvious. The duty falls on whoever develops the system, but in practice your assistant should say so clearly from its first message.
• The right to a human, in Spain. Spain’s customer service law, Ley 10/2025, bans automated systems as the only means of customer service and requires a person to be available whenever the customer asks, at any point in the conversation. It applies to basic-service providers (utilities, transport, telecoms, financial services…) and to companies with at least 250 employees, or above €50 million in turnover or €43 million in balance sheet, which have until December 2026 to adapt.
• In the US there is no general federal rule on disclosing bots, but some states have their own. California, for example, makes it unlawful to use a bot that misleads people about its artificial identity to encourage a sale, unless it is disclosed, and Utah requires disclosure of generative AI in certain consumer interactions. Check the rules of the states where your customers are.
• Personal data. Privacy law applies. In the EU, the GDPR means a legal basis, a data processing agreement with the model and platform providers, sending only the data needed, and not leaving decisions with significant effects on a person solely to automation.
• Channel rules. If the agent answers on WhatsApp, Meta’s policy also requires clear ways to reach a person. We cover it in WhatsApp CRM integration.

How to tell whether it works

• Enquiries resolved without a person, and verified: the customer does not write again about the same thing in the following days. An enquiry the agent “closes” but the customer repeats is not resolved.
• Satisfaction after the conversation, compared with human-handled conversations.
• Handover rate and reasons, which tell you what documentation is missing or which actions are worth adding.
• Errors found in the weekly review of a sample of conversations.
• Time to first response and to resolution.

Rolling it out step by step

1. Pick one channel and one type of enquiry. For example, the frequent questions in your website chat.
2. Tidy the documentation the agent will use: one source per topic, up to date.
3. Start in copilot mode. The agent drafts replies and your team reviews them before sending. You see how it answers real enquiries with no risk to customers.
4. Let it answer on its own where it is already getting things right, handing everything else to a person.
5. Review a sample every week and extend to new topics or channels once results are stable.

Example (hypothetical). An online shop gets many questions about shipping and returns. It starts with an agent that only drafts replies for its team, based on its shipping terms and returns policy. After a few weeks of review, it switches on direct answers for order status (checked against the shop) and questions about delivery times, with handover to a person for any complaint or whenever the customer asks. Every Monday it reviews twenty random conversations and fixes the documentation where needed.

Where to start

Look at the questions you get most often and whether your documentation answers them clearly: that is the agent’s first scope. This is the kind of project we do in AI chatbots and customer service, backed by our work in AI automation and agents. If you would like us to look at your case, in the free 30-minute audit we go through your common enquiries with you and tell you what an agent could resolve and what it would cost. Write to us from the contact page.

Frequently asked questions

What is the difference between a chatbot and an AI agent?

A classic chatbot follows a fixed option tree. An AI agent understands what customers write in their own words, looks things up in your documentation and systems, and can take simple actions within the limits you set.

How do I stop the agent from making things up?

Have it answer only from your documentation, say it does not know when the answer is not there, and review a sample of conversations every week to fix errors and fill gaps.

Do I have to tell customers they are talking to AI?

In the EU, the AI Act has required since 2 August 2026 that systems talking to people tell them they are AI, unless it is obvious. Some US states have disclosure rules too. Saying so clearly from the first message is the safe default.

Must I always offer a way to reach a person?

It depends on where you operate: Spain requires it of large companies and basic-service providers, and WhatsApp requires it on its channel. Elsewhere it is not always mandatory, but it is good practice.