Six things a good chatbot should be able to do

Why does it go wrong?

Many chatbots were built as a kind of menu. They look for familiar words in your question. If they spot a word, they give the matching answer. If they spot nothing, they give a standard reply.

That works fine for the ten questions that come in most often. But people rarely ask a question exactly the way the builder imagined. And the difficult questions are precisely the ones that differ from the standard case.

The technology has moved on since then. Modern chatbots use language models. These are programs trained on huge amounts of text. They do not look at single words, but at what you mean. Even so, plenty of the older kind are still running. That is the main reason customers get stuck.

Sven Lagerweij, CEO of Duppal, puts it this way:

“Chatbots are just like people. A receptionist or a service agent also needs to be told how you want them to treat customers. It is no different with a chatbot. You can give a chatbot instructions at all kinds of levels. So what the Consumentenbond describes really does not have to happen.”

That is the heart of it. A chatbot that drives customers away is usually not broken. It is a virtual agent that was never properly installed and trained.

1. Understand what the customer means

What goes wrong

The customer asks a question in their own words. The chatbot does not recognise those words and answers something else. Or it asks the customer to “rephrase the question”. When a follow-up question comes, it starts from scratch, as if the conversation never happened.

How can it be done better

A modern chatbot looks at the intent behind the question. Whether someone writes “there is a charge I don’t recognise” or “where does this 24.95 come from”, it is the same question, and the bot sees that.

Just as important: it remembers the conversation. If you then ask “and can I get that back?”, it still knows what “that” refers to.

It also draws its answers from sources you choose yourself: your FAQs, your manuals, the data in your customer system. That keeps the answers current. Change your returns policy, and the answer changes with it.

Zoom Virtual Agent works this way. It considers the intent behind the question, remembers what the conversation is about, and draws its answers from the sources you point it to.

How to test it

Ask a question that is not in your FAQ. Use everyday language and make a typo. Then ask a follow-up question without repeating the subject.

2. Notice when the conversation is stuck

What goes wrong

The customer gets the same answer three times. They try rephrasing and get it again. This is one of the most common reasons people give up. The chatbot has no idea it is going in circles.

How can it be done better

A good chatbot keeps track of whether the conversation is making progress. It watches for signals: is someone asking the same thing again, is the tone getting irritated, has the conversation gone quiet?

When it spots such a signal, it offers a way out on its own. It does not wait for the customer to ask.

You decide how quickly that happens. For a simple question about opening hours, two attempts are fine. For a complaint about a payment, you would rather hand over straight away.

Zoom Virtual Agent picks up on these signals by itself: repeated questions, an irritated tone, a conversation that stalls. It then hands over to an agent without being asked. Where that line sits is something you set per situation.

How to test it

Ask the same question three times in different ways. If the answer does not improve and the bot offers no way out, this is missing.

3. Hand over to a person quickly

What goes wrong

This is the biggest frustration. Nearly nine out of ten people feel that chatbots have made it harder to reach someone at customer service. Sometimes you cannot even see up front that a handover is possible.

And when it does work, the next problem often starts: the agent knows nothing. The customer has to tell the whole story again.

How can it be done better

Moving to a person should be a normal step in the conversation, not an emergency brake. Two things matter here.

First: the customer knows up front that it is possible and when. For example: “You can chat with me 24 hours a day. On weekdays between 9 and 5, I can put you through to a colleague.”

Second: everything travels with the handover. The full conversation, the customer details, and whatever the bot has already worked out. The agent reads up in 10 seconds and picks up where the bot left off.

You can set your own rules for when a handover happens. For example: always when the word “complaint” appears, always for business customers, or after two failed attempts.

With Zoom, the chatbot and the contact centre are one system. Zoom Virtual Agent passes the conversation to Zoom Contact Center with everything attached. The agent has the full conversation in front of them the moment they take over. All the customer notices is that someone else has started typing.

How to test it

Ask for a human agent. Count how many steps it takes. Then check whether you have to repeat your story.

4. Be honest that it is a chatbot

What goes wrong

Chatbots are given human names like Dan or Anna. They pretend to type. There is a green dot next to them, just like a colleague who is online. One in ten people only found out during the conversation that there was no human on the other end.

Often, the opening line clearly states that this is a chatbot. But that message disappears as soon as the conversation starts.

How can it be done better

Use a notice that remains visible throughout the conversation. Not a single line at the top that scrolls away, but something that remains in view.

This is no longer a matter of taste. The European AI Act requires companies to inform people when they are interacting with a computer, and it is not obvious. On modern platforms, this is a setting you switch on. Nothing needs to be built for it.

In Zoom Virtual Agent, this is a setting you switch on. The notice then remains visible throughout the conversation. Nothing needs to be built for it, and you meet the legal requirement straight away.

A human name is fine, by the way. What matters is that there is no doubt about what sits behind that name.

How to test it

Scroll back up halfway through the conversation. Can you still see that you are talking to a chatbot? Also ask the bot directly whether it is human. A well-configured bot answers that honestly and immediately.

5. Say when an answer might be wrong

What goes wrong

Chatbots that use language models sometimes get things wrong. That comes with the technology. The problem is that very few companies say so. In the Consumentenbond study, only seven out of a hundred companies warned that answers might contain errors.

So customers take those answers at face value. For a question about opening hours, that is annoying. For a question about what an insurance policy covers, it can do real damage.

How can it be done better

There are three ways, and they work best together.

• Add a short note to AI answers saying the answer was generated by AI and cannot be relied on as binding.
• Show where the answer came from, for example, with a link to the terms and conditions. Then the customer can check it themselves.
• Set a lower limit. Systems can estimate how confident an answer is. If that confidence falls below your limit, the bot does not give the answer but instead hands over. Better no answer than a wrong one.

Zoom Virtual Agent can show a warning and a source with every answer. It can also indicate how confident it is in each answer. If that confidence drops below the limit you have set, the bot does not answer but hands over instead.

Companies that are open about this are actually rated more positively by customers. Being honest about the limits builds trust.

How to test it

Ask about an edge case. If you get a confident answer with no warning and no source, this is not set up.

6. Let the conversation be saved

What goes wrong

At a third of the companies studied, you cannot save the chat. If you are promised something or given advice, you cannot find it again later. Call a week afterward, and you have nothing to point to.

How can it be done better

Make sure the customer can take the conversation with them. A button to save it or send the transcript by email automatically. Store it in your own customer system as well, so a colleague can look it up later.

Zoom Virtual Agent saves and transcribes conversations as standard. You can have the transcript emailed to the customer automatically, or stored in your own customer system.

This pays off for you too. Saved conversations show where customers get stuck, which questions keep coming up, and which answers aren’t working. Without that data, you cannot improve your chatbot. You simply do not know what is going wrong.

How to test it

Look in the chat window for an option to save or email the conversation. Is it hidden behind three dots? Then it exists, but nobody will find it.

Run the check yourself

You can find out where your own chatbot stands in ten minutes. Work through the six points:

Ask a question in your own words, with a typo in it.
Ask the same question three different ways.
Ask for a human agent.
Scroll back: does it still say you are talking to a bot?
Check whether there is a warning or a source with the answer.
Try to save the conversation.

Every point where you get stuck is a point where your customers get stuck every day.

In closing

Criticism of chatbots is almost never about the technology.

It is about companies that turned on a chatbot and never looked at it again.

That is good news. A problem in the technology is something you have to wait out. A problem in the setup is something you can fix yourself.

The research figures in this article come from the September 2026 issue of Consumentengids, the magazine of the Dutch Consumers’ Association (“Chatbot als klantenservice”). The association surveyed more than 10,000 panel members and examined the customer service of one hundred companies. The features described are available in platforms such as Zoom Virtual Agent combined with Zoom Contact Center. The quotation has been translated from Dutch.