AI Chatbot for Your Website: What It Must Do, What Is Nonsense

Most website chatbots are expensive and useless. The technology has been ready for a long time. The bots simply know nothing.

I have tested dozens of chatbots over the past two years. On client sites, on our own shops, at competitors. At nano, mate and MUSTAX we run an AI layer in support today that answers around 65 % of enquiries fully automatically. So I can tell you fairly precisely what a chatbot has to deliver and what is pure sales talk.

The distinction almost nobody makes

Before we talk about capabilities: "chatbot" is an umbrella term for three very different things.

Most vendors sell you number two and call it number three. That difference is the entire value. If you want the boundaries cleanly, read AI agent, chatbot or workflow.

When a chatbot genuinely helps

A chatbot pays off under three conditions. Miss one and it gets expensive.

First: you have volume. Under 50 enquiries a week you do not need a bot. A good FAQ page and a contact form are enough. The build simply does not pay for itself.

Second: your enquiries repeat. For us around 40 % of all enquiries are the one question: "Where is my parcel?" That is the case a bot is built for. One question, one data source, one clear outcome. How we solved it is in Automating WISMO enquiries.

Third: your customers ask outside your working hours. A large share of our enquiries arrives in the evening and at weekends. An answer at 10 pm is worth more than a perfect answer on Tuesday at 11.

If all three apply, a chatbot is one of the best investments in the shop. If two are missing, keep your money.

When it only annoys

There are situations where a chatbot actively does damage. I see three of them constantly.

The bot as a bouncer. You hide the support email address so everyone has to pass through the bot first. That is a barrier rather than automation. Customers notice immediately and get angry before they have even described their problem.

The bot on emotional cases. The gift did not arrive in time. The customer is furious. A politely worded template makes the situation worse. When sentiment is negative we escalate to a human immediately, with no detour.

The bot without knowledge. It answers every question politely and says nothing. That is the most common case out there. The customer asks three questions, gets three paraphrases of the shipping page and writes an email anyway. You have spent money and added a step.

Must-have 1: It has to know your order data

This is the most important capability, and it is missing from most offers.

A chatbot with no connection to your shop system can answer exactly one kind of question: the general kind. Shipping costs, return window, materials. Those answers are already on your website. The bot saves you almost no time there.

The value appears the moment it is allowed to look at the order. Here is how it works for us: the customer gives an order number or email, the bot pulls the order via the Shopify API, queries tracking status at the carrier and replies with the real state including the expected delivery date.

That is the difference between "your parcel is on its way" and "your parcel has been sitting in the Hamburg distribution centre since yesterday, delivery expected tomorrow". Only the second answer ends the conversation.

Ask every vendor directly: which systems does the bot read live? If the answer is "our knowledge base", you are buying a better FAQ page.

Must-have 2: A clean handover to humans

Every bot hits its limits. The only question is what happens then.

A bad handover means the customer gets a contact form and types their issue a second time. A good handover means the case lands with a human including the full conversation, the order number and the recognised issue, and the customer learns when to expect a reply.

We have fixed triggers for the handover:

The last point is the hardest and the most important. A bot that cannot say "I do not know" will invent answers. That is more dangerous than no answer at all, because an invented delivery date costs you real money later.

Must-have 3: Language and tone

This looks like cosmetics. It decides whether your customer believes the bot.

Real German, written as German. Many tools are built in English and translate. You recognise the result by sentences like "I appreciate your patience", rendered word for word into German. German customers read that as fake.

Your shop's form of address. German has an informal "du" and a formal "Sie", and your bot has to match whichever your shop uses. A break mid-journey looks cheap.

Short answers. The most common flaw in AI replies is length. Nobody reads five paragraphs in a chat window. We capped our answers hard: status, date, next step. Done.

Languages you actually serve. If you ship to Austria and Switzerland, German is enough. If you ship to the Netherlands, the bot needs Dutch, otherwise it produces support work instead of saving it.

Must-have 4: You have to see what it does

A chatbot without a log is a black box in your customer contact.

We log every automated reply and review them monthly. In the beginning we reviewed weekly. Most of our improvements came out of those reviews rather than out of the prompt.

At minimum you need: every conversation readable afterwards, a rate for handovers to humans, and an alert when the bot is unsure more often than usual. Without those three you first notice something is broken in your revenue.

The typical bad buys

Four patterns I see again and again.

The widget with no connection. 300 euros a month per brand for a bot that reads out your FAQ. Across all three brands that would have been around 900 euros monthly for us. Our own AI layer costs roughly 50 euros a month for all three together, plus about 15 euros of server. The arithmetic behind that is in What automation really costs.

The bot before the process work. If your delivery times are unclear and your returns rules contradict each other, the bot just makes the chaos visible faster. Sort the process first, then automate.

Everything at once. A bot meant to answer every question from day one answers none of them well. We started with a single question and only expanded months later.

Buying with no exit. Your answer logic and your conversation histories belong to you. Ask before signing how you export both. Anyone without a clear answer to that is selling you a dependency.

The limits

I make my living from automation. This section still belongs here.

A chatbot does not reduce your enquiries. It reduces the time per enquiry. Often the number of contacts even goes up, because the barrier is lower. That is fine, and it is a different calculation from the one vendors show you.

It has an error rate. Ours is not zero either. That is why our bot may inform but may not move money. Refunds and address changes always go through a human.

It needs upkeep. New products, changed shipping rules, a new carrier: every change in the business is a change to the bot. Budget a few hours a month. If you do not have them, leave it.

It does not replace good product advice. "Which toothbrush suits sensitive gums?" is a question we answer ourselves at nano, deliberately. Those conversations sell, and they build trust. You do not hand that to a machine.

If your case falls into this list, automation may not be worth it at all. Damian wrote up the longer version of that question: When automation makes no sense.

Conclusion

A good chatbot is access to your data with a text interface in front of it. The chat window is the smallest part.

The checklist is short. Does it know real order data? Does it hand over cleanly to humans? Does it sound like your shop? Can you read back what it did? Four times yes and it is worth it. One no on the first question is enough to decline the offer.

For us that is exactly why around 65 % of enquiries run without us and the remaining 35 % get more attention than they did before. What the complete setup behind it looks like is in the guide to customer service automation.

If you are unsure whether it adds up for your shop, drop us a line at Flowhouse. You get an assessment, even if it says you should work on your FAQ first.