AI Agents for Customer Emails: The Dos and Don'ts

Monday morning, 47 new emails. Half of them ask about delivery times, opening hours, the status of an order or whether an appointment is still available. Each answer is quick to write, but together they take hours. And while the team answers standard questions, the enquiries that are about real new business are left waiting.

An AI agent can take a lot of this off your hands. This article explains what such an agent actually does, when a person should approve the reply, what data protection law and the EU AI Act require and how the connection to your existing systems works.

The problem: the same 20 questions, every day

In many businesses, customer enquiries repeat themselves a lot. Typical examples:

• Retail and online shops: Where is my order? Can I exchange this? Is the item back in stock?
• Service providers: Roughly what does it cost? When do you have an appointment? Which documents do you need?
• Hotels and restaurants: Is the room free in May? Are dogs allowed? Is there parking?

The answers usually already exist somewhere: in the inventory system, the calendar, the price list or old emails. The problem is not missing knowledge but the effort of gathering it every time and writing it up politely.

What an AI agent actually does in your inbox

An AI agent is a program that uses a language model to complete tasks in several steps on its own, for example reading an email, looking up data and preparing a reply. In the inbox, it typically works like this:

1. Read and classify: the agent recognises what the email is about, such as order status, appointment request, complaint or new project enquiry.
2. Look things up: it fetches the relevant information from your systems, for example the delivery status from the shop or free slots from the calendar.
3. Prepare a reply: it writes a draft in your company's tone, based on your guidelines and text templates.
4. Forward: anything it cannot answer reliably goes to the responsible person with a short summary.

What an agent should not do matters just as much: invent prices, make promises nobody has checked, or answer legally sensitive questions such as warranty claims and cancellations on its own. You set these limits during setup.

Draft with approval or fully automatic: when to use which

The safest way to start is "AI writes, a person approves". The agent places a finished draft in the inbox, a team member reads it, adjusts it if necessary and sends it with one click. That saves writing time, and nobody has to worry about a wrong answer going out.

Fully automatic replies, without human review, are only suitable for cases that are clear and low-risk:

• Opening hours, directions, parking
• Acknowledgements of receipt with a realistic response time
• Order status, when the data comes directly from the shop

Anything involving money, deadlines, complaints or personal circumstances should be checked by a person. A good solution allows both: you start with approval for everything and only switch individual categories to automatic once the drafts have been reliable for weeks.

Data protection and the EU AI Act: what customers need to know

Customer emails contain personal data. The rules of the GDPR therefore apply, and since 2 August 2026 there are additional transparency obligations under the EU AI Act. The key points, in plain language:

• Data processing agreement: if an external AI service processes emails, you need a data processing agreement with that provider under Art. 28 GDPR. Reputable providers supply one.
• Storage location and training: choose providers and settings so that your data is processed in the EU where possible and not used to train other companies' models. Check this in the contract terms.
• Data minimisation: the agent should only see the data it needs for the reply, such as the order status, not the entire customer history.
• Privacy policy: add that and how you use AI to handle enquiries.
• No important decisions by the AI alone: Art. 22 GDPR restricts decisions made solely by automated means that significantly affect people. A refusal of a goodwill request should therefore be decided by a person.

On the EU AI Act: since 2 August 2026, the transparency obligations under Article 50 apply. AI systems that interact directly with people must be designed so that people are informed they are dealing with an AI, unless this is obvious. Whether this applies to drafts that an employee reviews and sends personally depends on the setup. For fully automatic replies, a clear note such as "This reply was generated automatically" is the safe choice. In addition, Article 4 requires you to take measures to support the AI literacy of the staff who work with the system. Find out more on our page about the EU AI Act. This section is not legal advice.

Connecting your CRM, shop or booking system

An agent is only as good as the data it can access. To write "Your order is on its way, the tracking number is …", it needs a connection to the shop. For appointment requests it needs the calendar. Typical connections:

• Inbox: Microsoft 365 or Outlook and Gmail offer official interfaces through which an agent can read emails and create drafts. Your team keeps working in the inbox they know.
• Shop and inventory: for example Shopify, WooCommerce or weclapp for order and stock data.
• CRM: so that enquiries are attached to the right customer and nothing is created twice.
• Calendars and booking systems: for free slots and reservations.

To connect the systems we often use n8n, a tool for workflows between programs that can also run on your own servers. Find out more on our page about n8n automation.

How to introduce the agent step by step

An AI agent in the inbox is not a switch you flip once. A step-by-step rollout has proven itself:

1. Collect enquiries: export the emails from recent weeks and sort them roughly by topic. This shows you which categories really are frequent.
2. Write down reply rules: for each category, note which information belongs in the reply, where it comes from and when a person has to take over.
3. Test phase with drafts: the agent only writes drafts. Your team rates them and notes what was wrong or unsuitable.
4. Refine: texts, tone and limits are adjusted until the drafts fit reliably.
5. Expand: only then are new categories added or individual simple cases answered automatically.

This way your team gets to know the system, and you keep control over every reply that leaves your company.

Costs and getting started

A sensible start is small: one or two enquiry categories, drafts with approval, a connection to one system. At EA Digital Solutions, AI integration starts from €2,500. If several systems such as the shop, CRM and calendar are connected as well, we charge system integration from €4,500 (as of October 2026). On top of that come running costs for the AI service, which depend on the number of emails.

On funding: aws (Austria Wirtschaftsservice, the federal promotional bank) supports AI and digitalisation projects through time-limited calls. At the time of our research in October 2026, no AI call was open. You always apply before the project starts. An overview is available under AI funding in Austria.

Example: an online retailer with many questions about order status starts with exactly this one category. The agent reads the order number from the email, fetches the status from the shop and places a draft in the inbox. Only once that runs reliably are exchange requests added.

Your next step

In the free process check you send us three typical enquiries, without real customer data of course. We show you how an agent would answer them, which systems it needs and what to consider regarding data protection. Our office is in Graz, and remote collaboration works just as well.