What does an AI customer service agent actually do?
It takes over the repetitive part of the work, not the relationship. In a business of ten to fifty people, a customer request comes in by email, through the website form or over the phone, and then it waits until someone reads it, works out what is being asked, finds the answer and writes back. The agent handles those steps:
- Read and sort: it tells an order-status question from a documentation request, a complaint or a request for a quote.
- Recognize the customer: it matches the sender to your customer list or CRM.
- Find the information: order status, product sheet, terms and conditions, answers already given to the same question.
- Draft the reply: it writes it up, showing where each piece of information comes from and how confident it is.
- Flag anything unusual: an unhappy customer, an urgent request, a question it can't handle.
When the agent doesn't know, it should say so. A good starting rule: "I don't know" is an allowed answer, and the agent gives it rather than guessing. To see how this works for price requests, our page on automating quotes with an AI agent walks through the same process, from the incoming email to a proposal that is ready for approval.
Which requests can it handle alone, and which should wait for approval?
It depends on what the reply commits you to. Use the table below as a starting point and decide request by request.
| Type of request | Example | What the agent does | Who decides |
|---|---|---|---|
| Information that already exists in writing | Opening hours, documentation, tracking for a shipped order | Replies on its own, citing the source | The agent, within rules written in advance |
| A question that touches a commitment | Lead time, availability, special terms | Drafts the reply with its sources | A person approves, edits or rejects it |
| A price or discount request | Quote, goodwill gesture | Gathers the details, never sets a price | Whoever is in charge of quotes |
| A complaint or an unhappy customer | Faulty product, disputed delay | Summarizes the case and raises an alert, does not reply | A person, every time |
| A legal or sensitive question | Personal data, a dispute | Passes it on without drafting anything | The owner or their advisor |
Two things to note. First, the top row is optional: you can keep human approval on every reply at the start, then open up the simple cases once you have watched the agent work for a few weeks. Second, the boundary is written down before the agent goes live: what it does alone, what it submits for approval, what it never does. Without that list, nobody knows who sent what.
Do you have to tell customers they are talking to an AI?
Yes, as soon as they deal with it directly. Article 50 of the EU regulation on artificial intelligence (the AI Act), which has applied since 2 August 2026, requires that people interacting directly with an AI system be informed of it, unless that is obvious from the context. The information must be clear and given at the latest at the first interaction.
In practice, this separates two ways of working:
- The agent replies itself (a chat on your website, an automatic email reply): customers must know they are dealing with an AI, through a visible notice from the very first message.
- The agent drafts, a person sends: the reply goes out in the name of the person who read and approved it. This is the easiest setup to keep under control, and the one we recommend starting with.
This is not legal advice. The official text and the European Commission's FAQ are listed at the end of this article.
How many requests can an agent really take on?
The honest answer: it depends on how many of your requests are repetitive and already have a written answer. A business whose customers mostly ask about order status has very different potential from one where every request is a custom project.
There is a benchmark for large contact centers: according to a McKinsey analysis of millions of interactions across more than 30 organizations, 50 to 60 percent of customer interactions remain transactional, meaning simple requests such as checking recent transactions or paying a bill (McKinsey & Company, March 2025). Those figures come mostly from banks and telecom operators, so they say nothing about the share in your own business.
To find out for your own business, a simple count is enough: take your last fifty requests and sort them into the five rows of the table above. The share in the first row tells you what an agent could handle alone; the next two rows tell you what it could draft for you.
When are you better off without AI?
When a fixed rule does the job. An automatic acknowledgment, an up-to-date FAQ page, a reply template in your email client or a filter that sorts messages by keyword can deal with part of your requests without any AI. They are cheaper to set up and more predictable. If your customers mostly ask the same questions, a chatbot or an FAQ page is enough: see chatbot or AI agent: which one to choose.
An AI agent starts to earn its place when customers phrase the same request in many different ways, when the answer means pulling information from several tools, or when the volume leaves customers waiting. For a clear picture of what an agent can and can't do, our article on what AI agents really do in a small business covers the five tasks where the difference shows.
How do you stop it from sending the wrong thing?
With three settings, put in place before it goes live:
- Limited access: the agent reads only what it needs and writes only where it has been allowed to. The four questions to ask before connecting it to your tools are in our article on connecting an AI agent to your tools safely.
- Sourced replies: every piece of information in a draft shows where it came from. A reply without a source does not go out.
- A full record: everything the agent read, drafted, sent or put on hold stays available. When a customer disputes a reply, you know who wrote it and what it was based on.
We hold ourselves to the same rule: at Équipage IA, the agents that prepare our content and our messages submit every draft for review, then for the founder's approval, and nothing goes out without it. An agent that can't find a piece of information flags it on its task card instead of guessing.
The same principle applies after the reply: a request that went unanswered can be followed up at the right time. That is what our page on quote follow-up automation describes.
Frequently asked questions
Can an AI agent take over from the person who answers customers?
No, and that isn't the goal. It takes away the repetitive part (reading, looking things up, writing a first draft) so the person can focus on the cases that need judgment: complaints, negotiations, key accounts.
Which tools can the agent work with?
The ones you already use: your inbox, your CRM, your business software, your documents. The point is that nobody has to open yet another tool.
What happens if the agent gets it wrong?
In draft mode, the person approving the reply catches the mistake before it is sent. In direct-reply mode, the agent only handles questions whose answer is written down and sourced, and every reply is logged so it can be corrected.
Do you need an IT department to set it up?
No. Setup starts with a single task, using the tools you already have and approval rules you write down together with us.
Where to start
The simplest way is to start with one task: order-status questions, for example, or documentation requests. We look at your tools, the kinds of requests you receive and what needs to stay human, then tell you whether a fixed rule is enough, whether an agent makes sense, or whether there is nothing worth building.
Sources consulted on 3 October 2026:
- Regulation (EU) 2024/1689 (AI Act), Article 50, Transparency obligations: AI Act Service Desk, European Commission
- Transparency obligations under Article 50 of the AI Act (FAQ): European Commission, Shaping Europe's digital future
- The contact center crossroads: Finding the right mix of humans and AI, McKinsey & Company, March 2025 (consulted on 4 October 2026)