Monday morning at a D2C shop with 300 tickets a day. A large share of them ask about the parcel, the invoice or the returns label. For each answer your team jumps between Shopify, the shipping tool and the ERP.
Customer service automation takes these requests off your team, because the answer is already in a system. This guide shows you where to start, which rate the data from 140 shops supports and what stays with your team.
What is customer service automation?
Customer service automation means software answers recurring customer requests and completes the matching action in the shop straight away. In e-commerce it pays off first for order status, returns, cancellations, invoices and address changes, because the answer is already in a system. How much of this you can automate depends above all on the connection to your shop, ERP and shipping. Your team keeps the cases that need a judgement call: complaints, goodwill decisions and legal questions.
The term covers three levels that differ widely in effort and outcome:
- Level 1: rules and macros. The ticketing system tags emails by keyword, routes them to teams and suggests canned replies. A person still has to send the answer.
- Level 2: chatbot with a decision tree. The bot asks for the order number and shows the tracking link. As soon as the question leaves the predefined path, it lands back in the inbox.
- Level 3: AI agent connected to the shop. The agent understands free text, reads the order from Shopify or JTL and carries out the action: cancellation, new delivery address, returns label.
Common guides on the topic are written for service companies in general. For an online shop, level 3 is what counts, because the answer is usually in the order data. Our guide to AI in customer service shows how it all fits together, from the inbox to the agent.
Which requests can you automate in e-commerce?
You can mainly automate requests whose answer sits in a system. If you want to automate customer requests, start by sorting them by where the answer lives. Anything that needs a judgement call stays with a person or is only prepared. The table shows eight typical request types and what you need to connect for each.
| Request type | What runs automatically | Requirement |
|---|---|---|
| Where is my order (WISMO) | Status and tracking link without a person | Shop system plus tracking from the shop, ERP or shipping tool |
| Return and exchange | Create the return, send the label, report the status | Returns portal or ERP that can create returns |
| Cancellation and refund | Cancel the order, refund directly or as a draft for approval | Write access to the shop system |
| Invoices and receipts | Send the invoice as an attachment | Invoice in the shop, JTL, Xentral or Billbee |
| Address change | Change the delivery address while the parcel is still in the warehouse | Write access plus shipping status |
| Product advice | Answer questions on size, material and compatibility | Well-maintained product catalogue |
| Complaint about a faulty product | Collect photos, order number and fault description | Handover to the team with all the data |
| Escalated complaint | Recognise negative sentiment and hand over to your team | Handover rule based on sentiment |
The right-hand column decides your rate. An agent without access to the order can only guess on WISMO. That is why Shopify, Shopware, JTL, Xentral and Billbee connect directly to the inbox in armincx by Chatarmin. The AI looks up status, tracking and invoices itself.
Marketplaces deserve a closer look. For Amazon, armincx reads the order data and carries out AI actions. OTTO and Kaufland arrive as an email channel in the same inbox, and your team handles actions on these marketplaces itself.
How much customer service automation is realistic?
The typical shop automates 16% of its tickets, the top quarter 61%. Half of all shops are below 16.3%, and you reach the top quarter from 37.5%. That is what the Chatarmin customer service automation benchmark shows across 140 shops. The per-shop values cover June 2025 to June 2026, the overall values 2.6 million resolved tickets from December 2025 to June 2026:
- Median per shop: 16.3% of tickets automated
- Average per shop: 25.2%
- Average in the top quarter: 61.2%, entry at 37.5%
- Across all tickets: 33.2%, so one in three tickets, through AI or workflows
The spread comes from shop connections and data quality. Two shops using the same software reach very different rates. It depends on how many systems are connected and how cleanly tracking numbers and invoices are maintained. For a shop with 3,000 tickets a month, the gap between 16% and 61% is around 1,350 tickets that run without your team.
For context: AI alone fully resolves only 4.7% of all tickets. It is involved in 23.9% of tickets and closes around one in five of those without a person. Most of the rate comes from workflows that read order data and trigger actions. The data comes from shops that use armincx and reflects this customer group.
Ask about the denominator for every rate. "1 in 5 AI tickets fully resolved" and "4.7% of all tickets" describe the same performance. The month matters too: the rate ranges from 24.8% in April 2026 to 38.5% in February 2026.
How can you automate your customer service?
Start with the request that comes in most often and whose answer is in the shop, then build from there. Three to five workflows are enough to begin with. These five steps get you there:
- Count ticket categories. To automate support, you first need numbers. Tag every ticket by request type for two weeks. Afterwards you know whether WISMO makes up a fifth or half of your volume.
- Connect your systems. Connect the shop system, ERP, shipping and returns portal. The more systems are connected, the more questions the AI answers without asking back. Check where the tracking number lives. Some shops only store it in the ERP.
- Start with WISMO. The question about the parcel is clear, the answer is in the system and a wrong answer does little damage. A typical flow: someone asks about their parcel on WhatsApp, and the AI finds the order and sends the status and tracking link. You decide when a ticket goes to your team, for example when the parcel has been in transit for unusually long. The solution page on where-is-my-order requests shows what this looks like.
- Set handover rules. Define when the AI hands over to a person. Typical triggers are words like "lawyer", negative sentiment, a high order value or the third follow-up on the same case.
- Set your metrics. Measure the automation rate, resolution time and satisfaction after the ticket. Compare the values per request type before and after the launch.
Only switch on new request types once the previous one runs smoothly. A good start is three workflows: order status, returns and invoices. Cancellations and address changes come after that. This shows you at every step how much work each workflow takes over.
Rule-based automation, chatbot or AI agent: what suits your shop?
The choice depends on whether your answers need an action in the shop. For a simple invoice request, a workflow that pulls the PDF from the ERP is often enough. As soon as someone wants to cancel, exchange or change the address in free text, you need an agent with write access.
| Criterion | Rule-based automation | Chatbot with decision tree | AI agent connected to the shop |
|---|---|---|---|
| Ticket volume | up to around 100 tickets a month | a few hundred, many identical questions | from several hundred, with peaks in Q4 |
| Channels | mostly email | website chat, often a single channel | email, WhatsApp, Instagram, Facebook and live chat in one inbox |
| Integration needs | keyword rules in the ticketing system | read access to tracking | read and write access to shop, ERP and shipping |
| Effort | hours for rules and canned replies | maintaining every single path | setting up the connections and training the AI |
| Limit | a person still writes the answer | breaks off as soon as the question leaves the path | goodwill, legal and escalations go to people |
Ideally, you combine all three. The page on invoice requests shows how a workflow sends invoices. Rules filter out spam and supplier mail, the workflow handles invoices, and the AI agent takes on everything that triggers an action in the shop. What matters is that all three run in the same inbox, so no ticket gets handled twice.
If you recognise your shop in the right-hand column, take a look at your own tickets. Book a demo of armincx and go through your request types with us, so you can see which workflows suit your shop system.
Which requests should you not automate?
Automate everything with a clear answer and hand over everything that needs a judgement call. Four cases belong with your team:
- Complaints with a risk of escalation. Anyone threatening a bad review or a lawyer wants a person. When sentiment is negative, the AI hands over to your team.
- Legal questions. Withdrawal after the cancellation period, warranty claims, requests for stored data. Every word counts here, and a wrong promise costs money.
- Goodwill above a fixed limit. Set a goodwill limit, for example €30. Refunds above it go to your team as a draft and are only triggered after approval.
- The first reply to VIP customers. Someone ordering for the fifth time notices the difference between a canned reply and a personal message.
If you don't trust the AI yet, start with drafts. With AI drafts in armincx, the AI writes the answer and your team checks and sends it. You switch on autopilot per request type once the drafts are right. Within a few weeks you learn where the AI answers reliably and where it still needs your team's help. The page on cancellations and refunds shows how they run as a workflow.
Sensitive product ranges also need their own rule. For supplements, cosmetics or medical devices, every question about effect, dosage or tolerance belongs with a person, even if the AI had an answer ready.
Complaints about faulty products sit in between. The AI collects photos, order number and fault description, and your team makes the decision. This saves the follow-up questions that otherwise drag a complaint out.
What does customer service automation cost and when does it pay off?
The calculation starts with the time your team spends on each standard request today. An example: 1,000 WISMO tickets a month at four minutes each add up to almost 67 hours of work. Every ticket a workflow takes over saves those minutes. If it takes over half, that is a good 33 hours a month, almost a full working week. For your own calculation you need three values:
- Volume per request type: from your ticket count over the past weeks
- Minutes per ticket: time five tickets per type and take the average
- Cost per hour: salary plus on-costs for your team
Multiply the three values and you know what each request type costs per month.
On the other side are set-up, licence and AI usage. Automated customer service pays off as soon as the time saved exceeds these costs. Weigh both sides against each other per request type. This shows you the volume at which a workflow pays for itself and which requests are better left with your team.
Automated tickets also close faster. Workflow tickets close after a median of 1.0 days, manually handled ones after 3.0 days. In peak season around Black Friday, the automation rate rises to 35.2%, right when your team needs the most relief. Both values come from the Chatarmin AI customer service benchmark across 131 shops.
What are the legal rules for automated customer service?
Since 2 August 2026, your customers must be able to tell that they are writing to an AI. This is set out in Article 50 of the EU AI Act. The simplest way is a short note at the start of the chat, such as "You're chatting with our AI assistant".
Order data is personal data. You need a data processing agreement under Art. 28 GDPR with every provider that processes it. Clarify these three questions with every provider before you start:
- Where is the data stored? Servers in the EU make your data protection review much easier.
- Which language model generates the answers? Ask to be shown whether customer data is used to train third-party models.
- Who sees what? Define which order data the AI may read and which actions it may carry out.
FAQ on customer service automation
What is customer service automation?
Customer service automation means software answers recurring requests and carries out actions such as cancellations or returns labels directly in the shop. In e-commerce this mainly covers order status, returns, cancellations, invoices and address changes.
What percentage of customer service can be automated?
In the Chatarmin benchmark of 140 shops, the typical shop automates 16% of its tickets and the top quarter 61% on average. The difference comes down to shop connections and data quality.
What does customer service automation cost?
The costs are made up of set-up, licence and AI usage and depend on ticket volume and connections. Set them against the handling time each automated ticket saves.
Which tools do I need?
You need an inbox that brings all channels together and a connection to your shop system, ERP and shipping. Only access to order data turns canned replies into complete answers.
Does automation replace employees?
No. It takes standard questions such as order status and invoices off your team, leaving more time for complaints, goodwill and advice.
How do I get started with automation?
Count your ticket categories for two weeks and start with the most frequent request whose answer is in the system. That is often the question about the parcel.
Conclusion: where to start this week
Customer service automation pays off where the answer is in the system and the action is clear. The top quarter of shops automates 61% of its tickets, the typical shop 16%. You close the gap through connections to your shop, ERP and shipping. Every connected system makes further request types automatable and gives your team time for complaints, advice and goodwill.
Start with WISMO, then add invoices and returns, for example through returns management in armincx. Leave complaints, goodwill and legal questions to your team. Your next step: count your ticket categories this week and mark the three whose answer is already in your shop system today.
