AI agents for customer service sort each request, look up the data for the case and draft the reply using the company's rules. A person decides whatever those rules do not cover: sensitive complaints, exceptions, compensation and any reply that commits the company.
This article is part of AI-First hybridisation consultancy.
In short
- The agent does the repetitive part: reading, sorting, finding data and drafting the reply.
- The person decides exceptions, sensitive complaints and anything that commits the company.
- The client has to know when they are dealing with an AI and be able to reach a person.
- The service rules are written before the agent goes in.
- You start with a person approving every reply.
What an AI agent does in customer service
An AI agent in customer service reads the request, sorts it, looks up the data for the case and drafts a reply using the company's rules. It has little in common with an FAQ chatbot, which only answers what it has written down. The agent works with the specific case: this client, this order, this invoice.
Dividing the work this way is company hybridisation with AI applied to one area: the agent does what repeats and a person decides what matters.
Customer service is a good place to start because most requests look alike. It is also a delicate one, because every reply reaches someone outside the company.
What is hybridised: the path of a request
This table follows a request from the moment it comes in until it is closed:
| Step | The agent does it | The person does it |
|---|---|---|
| Intake | Reads the message and sorts it by type and by how pressing it is | Nothing, unless the agent is unsure |
| Data | Finds the order, the invoice or the client's history | Nothing |
| Reply | Drafts a reply with the rule that applies | Approves it, corrects it or rewrites it |
| Exception | Sets the case aside and explains which rule is missing | Decides what is done |
| Close | Records what was answered and flags what is pending | Reviews what repeats and proposes new rules |
The step that frees the most hours is usually the second. Looking things up in three systems before you can answer is what takes the most time in a service team, and it hardly ever needs judgement.
What a person always decides
Some replies do not go out until someone has read them:
- Complaints where the client is angry or is threatening to leave.
- Compensation, refunds or concessions that fall outside the rule.
- Anything that changes the terms of a contract.
- Decisions with legal effects on the client. Article 22 of the GDPR protects people from decisions based solely on automated processing.
- Cases nobody has seen before.
In all of them the agent is still useful: it gathers the data, sums up the history and proposes a way forward. The decision belongs to the person.
What the client has to know
The client has to know when they are dealing with an AI. Article 50 of the EU AI Act requires that anyone interacting with an AI system is told so, unless it is obvious, and it applies from 2 August 2026.
In design terms, that comes down to two things. The agent's first message says it is an AI, in one plain sentence. And at any point there is a clear way to ask for a person, one that really works and that someone answers.
Hiding it goes badly. The client works it out in the end, and from then on distrusts the correct answers too.
The rules before the agent
A service agent without rules answers politely and with whatever seems plausible. For it to answer the way the company does, someone first has to write down how the company answers. These are the rules used most:
- Response times: how soon each type of request is answered.
- Tone: how to greet, how to apologise and which words are not used.
- Data: what the client can be told and what has to be checked first.
- Limits: what the agent may offer without asking, and what it may not.
- Escalation: in which cases the request is set aside, and who receives it.
Those rules live in a Book of Criteria, with versions and reviewed by a person. We explain it in what a Book of Criteria is and why your AI agents need one.
Three ways to start, from least to most
You do not have to choose between doing everything by hand and leaving it all to the agent. There are three steps:
- The agent drafts and the person sends. Every reply goes through someone, who now starts from a draft with the data already gathered.
- The agent answers whatever has a clear rule and sets the rest aside. The person sees what was set aside and a sample of the rest.
- The agent answers and also proposes rules based on what it sets aside. The person approves the rules.
Always start with the first. You move up a step when corrections become rare, and back down if they rise again. How to design those checkpoints is in human oversight of AI: how to design it in sales and operations.
How it is measured
Four figures, taken before you start and repeated afterwards, tell you whether the agent helps:
- Time to first reply.
- Requests the agent sets aside, and how many were right to be set aside.
- Replies the person corrects, with the reason.
- Cases that are reopened because the reply solved nothing.
The last figure guards against a common mistake: answering faster and worse. How to take these measurements before you start is in how to measure a process before automating it.
Common mistakes
Service projects with agents that go wrong tend to trip over the same things:
- Starting with complaints, which is where there are fewest rules and an error costs most.
- Leaving the rules inside the agent's instructions, where the service team can neither read nor change them.
- Measuring only speed, and celebrating a response time that hides more reopened cases.
- Taking the person out of the process before knowing how many replies they were correcting.
Where to start
With a single type of request, the most frequent and the dullest: the status of an order, an invoice that has not arrived, a change of details. Write its rules, measure how it is handled today and set the agent to drafting replies.
In the Legathum AI-First Hybridisation Method, that first process comes out of the diagnosis, and during implementation the first result in production arrives within 2 days, measured against the baseline. The work, area by area, is that of the AI-First hybridisation consultancy. When every area works this way, the company is an AI-first company.
The legal references in this article are EU law. Other jurisdictions have their own rules.
Frequently asked questions
What does an AI agent do in customer service?
It reads and sorts each request, looks up the data for the case, drafts the reply using the company's rules and sets aside whatever those rules do not cover.
How is it different from a chatbot?
A chatbot answers frequent questions with what it has written down. The agent works with the specific case: it checks the order or the invoice and drafts a reply for that client.
What should an agent not decide?
Sensitive complaints, compensation outside the rule, changes to terms and any decision with legal effects on the client. A person decides those.
Do you have to tell the client they are dealing with an AI?
Yes. Article 50 of the EU AI Act requires that they are told, unless it is obvious, from 2 August 2026. It is best said in the first message, with a way to reach a person.
Which type of request do you start with?
The most frequent and most repetitive one, with the agent drafting and a person approving every reply.
Where do I start?
Write the rules for one type of request and measure how it is handled today. If you want to look at it with your own company's case, the first step is a 15-minute meeting with Deibson Silva.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 50 (transparency obligations). Consolidated text, European Commission AI Act Service Desk.
- AI Act, Article 113 (entry into force and application). AI Act Service Desk.
- European Data Protection Board, Guidelines on automated individual decision-making and profiling (GDPR, Article 22).