AI in customer service: what actually works in 2026

There are two stories about AI in customer service, and both are true. The first: a customer lands in a chatbot, types “human” three times, gets a link to the FAQ three times, and then writes an angry review. The second: a shop with two people in support walks into the office on Monday morning, and of the 60 emails from the weekend, 40 have already been answered correctly, with order number, shipment status and the right tone.

The difference between those two stories isn't the technology. It's what you let the AI do, what you forbid it from doing, and whether someone looks before it answers on its own.

This guide sorts the topic out for small and mid-sized Shopify shops in the German-speaking market. It shows which requests can be reliably automated in 2026, which can't, how to do the maths, and what to watch during the first weeks so your customers don't end up in the first story.

Where most shops stand in 2026

According to the Bitkom study 2025, 36 percent of German companies with 20 or more employees use AI, almost twice as many as the year before. Of those that do, 88 percent use it in customer contact. Customer service isn't one use case among many; it's the most common one.

The same study names legal uncertainty (53 percent) and a lack of technical know-how (53 percent) as the biggest obstacles. That matches what we hear from shop owners: the open question isn't “whether”, it's “what exactly, and what happens when it goes wrong”.

On the customer side, the picture is contradictory, and you should take that seriously. A Gartner survey from 2024 found that 64 percent of customers would prefer companies didn't use AI in customer service at all. A Gartner survey from 2026 of 3,566 customers shows the picture two years later: half now find interactions with generative AI easier, but 87 percent say a way to reach a human must be available.

Customers don't reject AI. They reject AI standing between them and a solution. That's the sentence to keep in mind for the rest of this article.

What “AI in customer service” actually means in 2026

The term is used for four very different things. When you evaluate a tool, ask first which one is meant.

• Rule-based chatbots. A widget on the website with buttons and decision trees. It can route questions, but it can't look anything up. This is the technology that earned the bad reputation.
• Generative chatbots. A language model behind the widget. Sounds more natural, but doesn't know your orders unless it's connected. It answers plausibly, not necessarily correctly.
• Copilots for staff. The AI suggests replies, a human sends them. Safe, but the time saved is limited because every case still passes through a pair of hands.
• AI agents with data access. The AI reads the request, pulls order, customer and shipment status from your systems, applies your rules and writes a finished reply. It can send automatically or present a draft, depending on the case.

Only the last category can actually answer a delivery question instead of forwarding it to you. We describe the difference in detail in AI chatbot vs. AI agent. In short: a chatbot talks, an agent looks things up.

What can be reliably automated

The best way to check is your own inbox. Take the last 100 emails and sort them roughly. In most shops the result looks similar, and the big blocks have one thing in common: the answer already exists in a system. Someone just has to look it up and put it in a sentence.

Delivery questions

“Where is my order?” is the most common request of all and the most rewarding to automate. The answer consists of three pieces of information: order found, shipment status from the carrier, expected date. An agent with Shopify access and live parcel tracking can provide it without asking anything back. According to Zendesk CX Trends 2025, 67 percent of consumers are ready to delegate tasks like order tracking to an AI assistant. Hardly anyone wants a human for this request. They want a fast, correct answer. How to reduce the request before it arrives is covered in “Where is my order?”.

Returns requests

“How do I send this back?”, “Until when can I return it?”, “Where is my refund?” These are rule questions. Once your return conditions are stored, an agent can answer them correctly and send the customer to the returns portal. The refund itself should still go through an approval. More on that in a moment.

Order status and simple changes

“Has my order shipped yet?”, “Can I still change the address?”, “Please cancel.” Here it depends on the state of the order. Not shipped yet: changing the address is easy. Already with the carrier: it needs a different answer. An agent that knows the status gives the right one of the two. One that doesn't gives a generic one.

Product and shop questions

Size, material, shipping countries, payment methods, opening hours. An agent answers these well if it has a maintained knowledge base. How to build one in an hour is in What an AI agent needs to know about your shop. Without a knowledge base, the AI guesses, and it must not.

What you shouldn't automate

This is the part many vendors leave out. There are cases where an automatic reply doesn't just fail to help; it does damage.

• Angry customers. Someone writing for the third time, in capitals, or with “lawyer” in the text doesn't want a perfectly formatted template. They want to feel that someone is taking ownership. A good agent recognises the escalation and hands the case to a human. What to do then is in Complaint management: de-escalation in 5 steps.
• Money leaving the shop. Refunds, goodwill vouchers, replacement shipments. The AI can prepare the case and make a suggestion. A human should approve it, at least until you have a feel for how good the suggestions are.
• Anything the AI isn't sure about. Two orders under the same address, a question about a product that isn't in the knowledge base, a shipment status that hints at a problem. In these cases “I'm not sure, a colleague will get back to you” is the right answer, and it should be a draft, not an automatic email.
• Legal statements. Whether a withdrawal is still possible, whether a warranty applies, what happens with a damaged parcel. The AI may repeat your stored rules. It must not invent legal advice.

How much automation makes sense overall, and why 100 percent is the wrong target, is in The 90 percent limit.

The maths: is it worth it for a small shop?

Calculate with your own numbers, not the ones from a sales deck. Here's an example with assumptions you can adjust.

Suppose you get 300 support requests a month. Suppose around 200 of them fall into the categories above: delivery question, return, order status, product question. Suppose you need four minutes per reply on average, including looking things up. That's about 13 hours a month for replies that always follow the same pattern.

If an agent answers a large share of those without you and presents the rest as finished drafts you approve in 30 seconds, a fraction of that time remains. What you gain isn't primarily money. It's time for the 100 other requests where something actually went wrong, and the peace of not having to open the inbox at the weekend.

On the cost side, look at the pricing model. With Humane you pay per ticket, not per seat: the Starter plan includes 100 tickets, then €0.49 per ticket, less in larger plans. Every feature is included in every plan. Why we do it this way is in Why we bill per ticket. For your calculation it means costs grow with volume, not with team size, and you pay less in quiet months.

The Salesforce State of Service Report 2025, based on 6,500 service professionals, currently sees 30 percent of cases handled by AI and expects half by 2027. Those figures come from large organisations and don't transfer to your shop. But they show the direction: the share that works without a human is growing, and it's growing through the standard cases.

The difference between draft and automatic

If you take only one thing from this article, make it this: a good AI agent has two modes, and you decide per category which one applies.

In draft mode the agent writes the reply, but it doesn't go out. It waits in the dashboard, you read it, change it or approve it. This is the mode for the start and for anything sensitive.

In automatic reply mode the reply goes straight to the customer without you seeing it. This is the mode for categories where, after a few weeks, you know the agent gets them right: delivery questions with a clear status, returns questions with clear rules.

Here's how it works at Humane: the first week is a training week, everything lands as a draft. After that you enable automation per category. Only what's been approved and where the agent is confident goes out automatically. Anything unclear stays a draft. If you correct a draft, the correction becomes a rule for all following cases. Which tickets go out without approval in practice and which never do is described in Auto-send: which tickets Humane answers without approval.

The important thing about this setup: you don't hand over control on day one, but category by category, after you've seen what happens.

How to recognise a good tool

When comparing vendors, these are the questions that make the difference. Most can be answered during a trial period.

1. Can the AI look things up in Shopify, or does it just talk? Let it answer a real delivery question. If the reply contains no status and no date, it can't.
2. What happens when it's unsure? Send a question about an order that doesn't exist. The right reaction is a draft or an honest follow-up question, not a made-up answer. More in Hallucinations in customer service.
3. Does the reply sound like your shop? An agent that writes “Dear valued customer” when you've been informal for five years turns every reply into an email that feels foreign. How tone works is in Tone of voice: how an AI sounds like your brand.
4. Does it learn from corrections? If you have to make the same change three times, it doesn't.
5. Where is the data stored, and is there a data processing agreement? Customer emails contain names, addresses, order histories. Without a data processing agreement you can't hand them to a service provider. What to check is in AI support and GDPR.
6. How is it billed? Per seat, per feature, per ticket? Run your volume for a quiet month and a busy one.

Getting started: the first four weeks

A realistic sequence that works for most shops.

Week 1: connect and watch. Connect the shop, answer a few questions about it, set up the knowledge base. With Humane this usually takes under 15 minutes, and training starts automatically afterwards. Everything stays in draft mode. You read every draft and correct what doesn't fit. This is the most important week, because every correction becomes a rule.

Week 2: enable the first category. Usually delivery questions. You've watched for a week how the agent answers them. If the drafts have gone out unchanged for several days, you switch on automation for that category. Everything else stays a draft.

Week 3: returns and order status. Same logic. Check that the rules apply, then enable. Refunds stay behind approval.

Week 4: measure. Three numbers are enough: share of requests answered automatically. Share of drafts you approved without changes. Follow-up questions per order. If the third number rises, the replies weren't good enough, and you switch the affected category back to draft.

What you shouldn't do in this period: enable everything at once because the first three drafts looked good. The mistakes happen in the cases that are rare.

What has changed in 2026

Two things are different from two years ago, and both matter for small shops.

First: the technology is no longer the bottleneck. An agent that reads Shopify, queries the carrier and writes in your tone isn't an enterprise solution any more; it's set up in minutes. The bottleneck now is the preparation: clear rules, a maintained knowledge base, a decision about what may go out automatically.

Second: customer expectations have shifted. According to Zendesk CX Trends 2025, 61 percent of consumers expect AI interactions to feel tailored to them, and 64 percent are more likely to trust AI agents that come across as friendly and empathetic. Customers no longer compare your automatic reply with “no reply at all”, but with the best automatic reply they got this week.

For you that means: an AI that knows the order number, the status and your tone is now the baseline, not an extra. One that can't stands out.

What you can do now

• Count your last 100 support emails and sort them into delivery question, return, order status, product question, complaint, other. That's your basis for every decision.
• Write down your rules: return period, goodwill, what happens with “delivered but not here”, how you address customers. One page is enough.
• Start in draft mode and stay there until you've gone a week without having to change anything.
• Enable automation per category, not globally. Start with delivery questions.
• Keep a path to a human open, visible and without hurdles. That isn't a weakness of automation; it's the condition for customers accepting it.

The full guide to support processes, beyond AI, is in Customer service for online shops: the complete guide.

If you want to see what your own requests look like as drafts: try it free for 14 days, no credit card, and in the first week nothing goes out without you.