Almost every company with a website has had the question: should we have a chat? And right after: people, a chatbot or both? The answer depends on which questions customers ask, when they ask them and what you can afford to staff. This article covers what chat is good for, what customers actually think of chatbots, what the law has required since August 2026 and how to set up a chat that helps instead of irritating.
Which questions suit chat?
Chat suits questions that arise in the middle of something else and have a short answer: is the item in stock, what does shipping cost, where is my order, how do I withdraw from a purchase. The customer is at the checkout or on a product page and does not want to leave it to send an email. An answer within a minute keeps the purchase; no answer often loses it.
Chat suits questions less well when they need material from the customer or a judgement from you: a complaint with photos, a dispute about an invoice, a cancellation with conditions. Those get long and are better in the inbox, where the ticket can stay open for a few days. How the channels divide otherwise is covered in the article on when self-service, live chat and email fit.
The practical rule: what the help centre can already answer belongs in the chat, because a chatbot can answer it instantly there. What no help article covers belongs with a person, in the chat if it is urgent and in the inbox if it is not.
Live chat or chatbot: what do customers say?
Customers want a fast answer and are willing to take it from a bot, but only if the bot is as good as a knowledgeable person. According to Zendesk's compilation of customer service statistics, 51 per cent of consumers say they would rather deal with a bot than a person when they want immediate help, and 68 per cent think a chatbot should have the same expertise and quality as a skilled agent. The expectation of pace is high: according to Zendesk's CX Trends 2026, 74 per cent of consumers expect customer service to be available around the clock and 88 per cent expect faster responses than a year ago.
Swedish customers are more sceptical than the global average. Svenskt Kvalitetsindex, the Swedish quality index, reports that only 41 per cent feel comfortable with chatbots and AI assistants, and that 60 per cent of those interviewed would rather have customer service from real people. Phone is still the channel most consider the most important: 63 per cent, against 19 per cent for chat, according to a survey referenced by Techtidningen.
The two pictures do not contradict each other. Customers reject bad chatbots: ones that guess, that do not understand the question and that cannot be bypassed. A bot that answers "where is my order" correctly at eleven at night gets no complaints. A bot that answers a complaint question wrongly and refuses to hand over gets all of them.
How should a chatbot be built so it does not damage trust?
It should answer from your reviewed knowledge base, be able to look up the customer's own details, say when it does not know and hand over to a person without the customer starting over. Those four matter more than which language model is used.
- Grounded in the knowledge base. The bot should only state what is in your articles, terms and systems. Anything not there gets "I don't know" and an offer of a person. What that means technically and what to ask the vendor is covered in the article on chatbots that do not guess.
- Lookups against orders and customers. The most common chat questions are personal: my order, my invoice, my return. A bot without a connection to the order system can only answer in general terms, and that is when customers find it useless.
- Handover with history. When a person takes over they should see the whole conversation and what the bot looked up. The customer should not repeat themselves. Handover should take one word, and the bot should offer it after two failed attempts.
- No lock-in. A bot that makes the customer pick from five buttons before it shows a free-text field, or hides the way to a person, gets low ratings however good its answers are.
Supportifier's AI and live chat is built that way: the same knowledge base as the help centre, lookups against the customer's tickets and a colleague who can take over mid-conversation.
What does the law require of a chat?
That the customer is told they are talking to an AI, that the chat is accessible to people with disabilities and that the personal data in the conversation is handled under GDPR. The first requirement is new. Since 2 August 2026, Article 50 of the EU AI Act applies, which according to the European Commission requires AI systems that interact directly with people to be designed so that the person is informed they are dealing with an AI, unless it is obvious from the context. In practice: a clear line in the chat's first message, not a note in the terms of service. A name like "Anna" and a photo of a person do not qualify.
The accessibility requirement has applied since June 2025 for most online stores and means, among other things, that the chat must work with keyboard and screen reader and that there is an alternative for anyone who cannot use it; what is required is covered in the article on the accessibility act in customer service. The GDPR requirement is that chat logs are personal data with their own retention period, and that the vendor is a processor with an agreement stating where conversations are stored and whether they are used to train models.
How do you staff the live chat?
Staff the live chat during the hours the questions come, let the bot take the rest, and never show "we are online" when nobody is. A live chat that answers after twenty minutes is worse than no chat, because the customer has waited instead of emailing.
Start by reading when chat questions actually arrive. For most online stores it is evenings and weekends, when customer service is not at work. That is the argument for the bot: it takes the common questions when nobody else can, and leaves the unusual ones as tickets for the next morning with a clear message about when the customer will get an answer.
During staffed hours, expect one person to handle two to three simultaneous chats if the questions are short and the bot has already looked up the order, fewer if the questions require doing something in another system. Set a queue with an honest waiting time, and let the customer leave an email address and move on if the wait is more than a few minutes. That is better for the customer and for CSAT than holding them.
What should you measure in the chat?
Measure the share of conversations the bot resolves without a person, the share handed over, the time to first reply from a person, the customer rating per conversation and the share who leave the chat without an answer. Those five show whether the chat helps or just moves questions.
| Measure | What it tells you | What to do if it is poor |
|---|---|---|
| Resolved by the bot | How much of the questions the knowledge base covers | Write the missing articles, connect more lookups |
| Handed over to a person | Where the bot runs out | Read the conversations before handover, fill the gaps |
| Time to first reply from a person | Whether staffing is enough | Move hours, show an honest waiting time |
| Rating per conversation | Whether the customer got help | Compare bot and people separately |
| Abandoned conversations | Whether the customer gave up | Check the queue, lock-ins and buttons |
Always compare the bot's and the people's ratings separately. A bot that scores 4.2 out of 5 on order questions and 2.1 on complaints should not answer complaints, and that only shows if you split the measures.
What to do
- List the twenty most common questions arriving by email and phone today. Mark which have an answer in the help centre. That is the bot's first job.
- Write the missing articles before you switch the chat on. A bot without material guesses or answers "I don't know" to everything.
- Connect the order system so the bot can answer "where is my order" with the customer's own status. Without it the chat is an FAQ with a delay.
- Set the handover rules: after two failed attempts, on the word "human", and always for complaints, cancellations and grievances.
- Write the first message so it says it is an AI, what it can help with and how the customer reaches a person. That satisfies the AI Act and sets expectations the right way.
- Decide staffed hours from when the questions arrive, and show them in the chat. Outside them the bot answers and takes tickets.
- Follow the five measures every week for the first month, and remove the bot from the question types where it scores low. In Supportifier this is controlled per question type in the knowledge base, so the bot answers alone where it is good and hands over where it is not.
Common questions
Must a chatbot say it is an AI?
Yes. Since 2 August 2026, Article 50 of the AI Act requires that anyone chatting with an AI is informed of it, unless it is obvious. The information must come at the latest at first contact, clearly and in the chat itself. A human name and a portrait on the bot make the requirement harder to meet, not easier.
What happens if the bot answers wrongly?
The company is responsible for what the bot says, just as for what an employee says. That is why the bot should only answer from reviewed material and hand over when unsure. Log all conversations, read the handed-over ones every week and correct the knowledge base when an error is found; the error then disappears in all channels at once.
Does a small online store need live chat with people at all?
Not necessarily. A store with a few hundred orders a week often manages with a chatbot that answers from the help centre and looks up orders, and creates a ticket in the inbox when it cannot answer. Live chat with people pays off when questions arrive during office hours and affect purchases, for example product questions at the checkout.
Should the chat be open around the clock?
The bot yes, the people no. Customers expect to be able to get help at any time, but they do not expect a person at three in the morning. What they do not accept is a chat that looks open and then does not answer. Always show whether a person is available right now, and what happens to the question if not.
Sources
- 92 customer service statistics you need to know in 2026 — Zendesk, 2026
- Contextual Intelligence Becomes the New Standard for Exceptional Customer Experience in 2026 — Zendesk, CX Trends 2026
- AI och kundnöjdhet: är botten nådd? — Svenskt Kvalitetsindex, read 2026
- Kundo: "Kundservice har fått högre prioritet i många organisationer" — Techtidningen, read 2026
- Transparency obligations under Article 50 of the AI Act — European Commission, read 2026
Martin Carlsson