The median online shop automates just 16.3% of its support tickets, while the top 25% average 61.2%. That's the result of a Chatarmin analysis of 140 online shops and around 2.6 million tickets. In our experience, the biggest challenge of AI in customer service is rarely the AI itself. The gap comes from data silos, missing access to order data and trying to automate everything at once.
How far has customer service automation really come?
Not as far as most people think: the median is 16.3%. The figures come from our customer service automation benchmark. The basis: 140 shops that have been live for at least 60 days and have at least 50 resolved tickets. The per-shop values cover June 2025 to June 2026, the roughly 2.6 million resolved tickets cover December 2025 to June 2026. Data as of 1 July 2026.
Automation rate means: the share of resolved tickets in which a workflow or AI was involved in any form.
The key figures:
- Median per shop: 16.3%
- Average per shop: 25.2%
- Top 25%: 61.2% on average, starting at 37.5%
- Top 10%: 63.8% and above
- Across all tickets (December 2025 to June 2026): 33.2%
Why are the values so far apart? The median shows the shop right in the middle. The average is higher because a few shops with very high rates pull it up. If you calculate across all tickets instead of per shop, you land at 33.2%. In other words: shops with lots of tickets tend to automate more than the typical shop.
The limitation is part of the picture: the sample consists of Chatarmin customers. It is not representative of e-commerce as a whole. The study names setup, industry and ticket mix as influencing factors but doesn't explain causes. Everything you read below about reasons is therefore our assessment from practice, not a study finding.
This article is about problems and solutions. What AI can do in principle and which benefits it brings is covered in our overview of AI in customer service.
Challenges of AI in customer service: what holds automation back?
We see seven challenges again and again: chatbots with canned answers, tool sprawl, no real-time access to order data, trying to automate everything, fear of hallucinations, data protection questions and low acceptance in the team. In our experience, most of them are setup problems, not model problems.
| Challenge | How you notice it | What helps |
|---|---|---|
| Chatbot only manages standard answers | Customers type "agent" as soon as it gets specific | AI with access to order data and real actions instead of an FAQ script |
| Tool sprawl and silos | One case, several tickets in different channels | All channels in one inbox, one customer history |
| No real-time access to orders | Agents copy tracking numbers between tabs | Connect shop, ERP and shipping directly |
| Automating everything at once | Big project, little impact after months | Start with one ticket type, then expand |
| Fear of hallucinations | AI stays in test mode, nobody signs it off | Clear knowledge sources, limits and escalation rules |
| Data protection and hosting | Project is stuck with legal or the DPO | Clarify hosting, DPA and AI transparency early |
| Low acceptance in the team | Agents bypass the AI or rewrite everything | Involve the team early, make the relief measurable |
Customer service challenges without automation
Without automation, routine eats up your team's capacity and every peak becomes a bottleneck. These are the typical customer service challenges before AI even comes into play:
- Overload from routine tickets: shipping questions, invoice copies and return requests all land with people, even though the answer is already in the system
- Manual double documentation: agents enter the same information in the ticketing system, shop backend and ERP
- Peaks without a buffer: Black Friday, holiday season or sick leave hit a team that's already at capacity in normal operations
- Frustration in the team: answering the same questions all day leaves little time for the cases where good advice really matters
This is exactly where AI comes in. Why it often still doesn't take hold is covered in the following sections.
Why do so many AI rollouts in customer service fail?
In our experience: because AI is put on a foundation that can't carry it. Legacy systems, data silos and missing interfaces mean even good models only serve up canned text. Here are the seven challenges in detail.
1. The chatbot only answers standard questions
In our experience, chatbots built on isolated FAQ solutions often fall back on outdated scripts. They can answer, but they can't act.
How you spot it:
- The bot refers every order question to the contact form
- Customers immediately ask for a human
- Your team still gets the same questions by email
What to do:
- Check whether your AI can take actions, such as looking up order status or changing an address
- Replace rigid decision trees with answers from well-maintained knowledge sources
- Read up on the difference between AI agents and chatbots before choosing your next tool
2. Legacy systems, data silos and tool sprawl
If email, WhatsApp, Instagram and live chat run in separate tools, no AI sees the whole case. If a customer writes by email first and then on WhatsApp, you end up with two tickets and often two answers.
How you spot it:
- Duplicate tickets for the same request
- Agents work from shared Outlook or Gmail inboxes
- Phone, ticketing and chat give different answers to the same question
What to do:
- Bring all channels together in one inbox with a shared customer history
- Maintain one knowledge base that all channels answer from
- Clear out legacy tools: which ones do you really use, and which only out of habit?
3. No real-time access to orders and customer history
"Where is my order?" can only be answered by someone who can see the order. Without a connection to the shop, ERP and carriers, the AI is left with a standard sentence and a link.
How you spot it:
- Agents copy tracking numbers by hand from the backend
- The AI asks for information that's already in the shop system
- Tracking sits in the ERP, but the AI only looks at the shop
What to do:
- Connect every system that holds order and shipping data, not just the shop
- Check up front where the tracking number is actually stored
- Start with the shipping question: our guide to automating WISMO requests shows how
4. Isolated automation instead of clear priorities
A common misconception: not every contact has to be automated. Trying to do everything at once leads to a big project that takes a long time to deliver anything.
How you spot it:
- The AI project has been running for months without measurable impact
- No numbers: nobody knows your own automation rate
- Nobody can say which ticket type gets automated first
What to do:
- Measure where you stand first. Without measurement you can't tell whether you're at 16% or 40%. We cover the customer service KPIs you need for this separately
- Start with the most frequent, clearly defined ticket type, then returns, invoice copies, address changes and cancellations
- Advice-heavy and emotional cases deliberately stay with the team
5. Fear of hallucinations and false promises
The fear is justified: wrong answers from an AI can cost a company dearly. In Moffatt v. Air Canada, the Civil Resolution Tribunal of British Columbia ruled in February 2024 that the airline was liable for incorrect information given by its website chatbot (decision 2024 BCCRT 149). It's a Canadian case, so have the legal position in Germany, Austria or Switzerland checked for your specific situation.
How you spot it:
- The AI stays in test mode indefinitely
- Every answer is fully checked before it's sent
- Sensitive topics such as refunds aren't covered by any rules
What to do:
- Define which sources the AI may answer from, and keep them up to date
- Define topics where the AI never decides on its own but hands over
- Let the AI ask follow-up questions on incomplete requests before it escalates
6. Data protection, hosting and AI transparency
Data protection usually slows projects down because the questions are asked too late. Where is the data stored? Which models process it? Is there a data processing agreement (DPA)?
Then there's transparency: since 2 August 2026, the AI disclosure requirement under Article 50 of the EU AI Act applies, so customers must be told they're chatting with an AI unless it's obvious.
How you spot it:
- The project has been stuck with the legal department for weeks
- Nobody can say where customer data is processed
- Your bot doesn't identify itself as an AI
What to do:
- Clarify hosting, DPA and standard contractual clauses before choosing a tool
- Check access rights: who in the team can see which customer data?
- Clearly label AI answers in the chat
7. Low acceptance in the team
AI also fails because of people who don't want to use it. Anyone who fears for their job or was never trained will work around new tools.
How you spot it:
- Agents completely rewrite AI suggestions
- Feedback on the AI only comes as complaints, not as improvements
- The team doesn't know which tickets the AI takes over
What to do:
- Communicate clearly: the AI takes over routine, not jobs
- Make experienced agents responsible for the knowledge base and workflows
- Show the team how many tickets the AI has already taken off their plate
How AI supports your team day to day instead of replacing it:
- Reply suggestions: the AI drafts a reply, the agent checks and approves it
- Handover with full context: when the AI escalates, the agent sees what has already been asked and answered
Put simply: AI supports, it doesn't replace. Teams that experience it that way actually use the suggestions.
What do the top 25% do differently?
They automate selectively and on the basis of connected data. In the benchmark, the top 25% reach 61.2% on average, and the entry point for this group is 37.5%. The top 10% start at 63.8%. The study doesn't say what's behind this. In our experience, we see four patterns at shops with high rates:
- A clear starting point: shipping and order status questions first, because they're frequent and easy to define
- Integration before AI: shop, ERP and shipping are connected before the AI goes live
- Escalation rules: for every ticket type it's defined when a human takes over
- Measurement: automation rate, response time and customer satisfaction are reviewed regularly
An example from practice: Englputzeder connected Shopify and Odoo and started with WISMO requests, CSAT surveys and automatic tagging. According to the case study, more than 40% of over 3,000 tickets a month are handled or prepared without manual intervention.
Getting from 16% to 37.5% isn't a model switch, it's process work. You'll find a step-by-step guide to implementation under customer service automation.
Does AI in customer service actually pay off?
It depends on how many recurring tickets you have and how well your systems are connected. Blanket ROI figures won't help you much. Calculate with your own numbers:
- How many tickets per month relate to shipping, returns and invoices?
- How long does an agent take for such a case today?
- What share of them can be resolved automatically with connected data?
- What do peaks like Black Friday cost you today in overtime or temporary staff?
Then there's the question: managed AI or building your own? Our view: building your own only pays off if you have your own development team that maintains integrations, the knowledge base and monitoring on an ongoing basis. For most shops, a ready-made solution with e-commerce integrations is, in our view, the faster route.
One example is armincx by Chatarmin:
- All channels in one inbox: email, WhatsApp, Instagram, Facebook and web chat come together with a shared customer history
- Actions instead of canned replies: workflows handle cancellations, address changes, invoice retrieval or returns, depending on the connected system
- E-commerce integrations: including Shopify, Shopware, JTL, Xentral and Billbee
- Data protection covered: hosting in the EU, servers in Germany. On top of that, a DPA and EU standard contractual clauses for sub-processors outside the EU
FAQ: challenges of AI in customer service
Will support staff become redundant because of AI?
No. AI takes over routine tickets and repetitive tasks, and your team gains time for complex and personal requests.
What are the disadvantages of AI in customer service?
The biggest ones are wrong answers when the AI lacks data or clear limits, and frustrated customers when a bot can neither act nor hand over. Both are setup problems you can control with connected data, clear sources and escalation rules.
What is a good automation rate in customer service?
It depends. In the Chatarmin benchmark, the median is 16.3% and a shop reaches the top 25% from 37.5%, with industry and ticket mix having a strong influence on the value.
Does it make sense to automate all customer contacts?
No. Routine and standard requests are well suited, but complex and individual cases still need people.
Our last chatbot was bad. Is it worth trying again?
It depends. If the new solution can access order data and take actions, a fresh start with a single, clearly defined ticket type is worth it.
How do I get consistent answers across phone, ticketing and chat?
With one shared knowledge base that all channels answer from. It gets easiest when as many channels as possible come together in one inbox with one customer history.
Do I have to tell customers they're chatting with an AI?
Yes. Since 2 August 2026, Article 50 of the EU AI Act requires that users are informed they're talking to an AI unless it's obvious (more on the AI disclosure requirement).
What should I consider before introducing AI in customer service?
A clean, up-to-date data basis and clearly defined support processes are the most important prerequisites. Also clarify hosting, DPA and escalation rules early on.
How do you close the gap between 16% and 61%?
By working through the challenges one by one instead of waiting for a better model. The median shop in the benchmark automates 16.3%, the top 25% reach 61.2%. In our view, the difference lies mainly in integration, clear priorities and measurement.
In short:
- Measure your rate before you change anything
- Connect your data before the AI goes live
- Start with one ticket type and expand from there
- Sort out escalation, data protection and transparency from day one
Want to see what this looks like with your shop setup? Book a 15-minute demo and we'll look at your ticket mix together.
