AI in practice 7 min read

Where can I compare AI chatbots for customer service?

The best place to compare AI chatbots for customer service is your own ticketing system or mailbox: collect twenty to fifty real questions and put them to every candidate. A comparison site tells you which features a bot has, not whether it understands your customers. Put in that set the questions your team gets most often and the messy cases: a customer asking three things at once, a question with a typo, a complaint with emotion in it. Judge two things separately per answer: was the information correct and was the tone right. A bot that answers correctly nine times out of ten but makes something up about your warranty terms the tenth time costs you more than it delivers. Keep the set; you use it later to check every change.

What kinds of AI chatbots for customer service are there?

There are four kinds: the chat function in your existing helpdesk or webshop package, a standalone chatbot platform with its own subscription, a knowledge assistant that answers from your own documents with sources and a voice agent for the phone. The first is the quickest to switch on and fits when your conversations already live in that package; quality in Dutch varies per package. A standalone platform gives more freedom but needs a connection to your customer data, otherwise it can only talk, not help. A knowledge assistant on your own documents is the choice when most questions are about your terms, procedures or products: it first looks up the matching pieces of text and only then lets the language model answer, with the source attached. And if your customers call rather than type, a chatbot is the wrong channel.

How do you compare subscriptions for AI-based customer service solutions?

Compare subscriptions on the cost model, not on the monthly amount: per agent, per conversation, per message or per minute, plus the one-off setup. Work each model through with your own number of conversations per month and see what happens when that number doubles. A price per agent is predictable but only pays off if your team actually uses the bot. A price per conversation is fair with fluctuating volume and can climb in a busy month. Also look at three things that are not on the pricing page: where conversations are processed and whether they are used for training, whether you can export conversations when you leave and what handover to a person costs. The setup is usually the largest item and decides whether the bot is still switched on after three weeks. Do not cut it to lower the monthly fee.

Want to know which of your customers' questions a bot could handle today? Describe your bottleneck →

What is the difference between a chatbot and a knowledge assistant on your own documents?

An ordinary chatbot answers from a script or from what the language model knows in general. A knowledge assistant first looks up the matching pieces from your own handbooks, terms or protocols and only then lets the model answer, with the source attached. That difference decides whether you dare to trust the answer. With a knowledge assistant an employee or customer can see where every answer comes from. If the answer is not in your documents, it should say so instead of guessing. All the gain is in the retrieval: split documents by structure, keep provenance and remove outdated versions. From own work, not commissioned: an HR assistant that answers questions from a staff handbook with the source attached.

How do you stop a chatbot from making things up?

By letting it answer only from your own information, showing the source with every answer, instructing it explicitly to say when something is not in the documents and testing every change against thirty questions whose answers you know. A language model is built to answer fluently, not to doubt. Without a source and without a limit it answers just as convincingly about your returns policy as about something it invented. Measure two things separately: did the system retrieve the right document and was the answer correct. Then with a poor result you know immediately which side to adjust. And tidy your archive before you start: three versions of the same protocol give three answers.

Chatbot or voice agent: which fits your customer service?

Choose a chatbot if your customers mainly type, through your website or an app, and a voice agent if they call. Many organisations need both, but start with the channel where most conversations come in now and where the waiting time is. A chatbot is cheaper per conversation and easier to test. A voice agent catches what is lost today: the call outside office hours, the peak between eight and ten. Look in your phone system and your ticketing tool to see where the volume is. If you miss calls, start with the phone. If emails pile up, start with chat. The article on voice agents for customer service gives the decision rule for the phone.

What do the GDPR and the AI Act say about chatbots in customer service?

The GDPR requires a data processing agreement with the vendor and clarity about which data the chatbot sees and how long conversations are kept. Since 2 August 2026 the AI Act requires that people know they are talking to AI, so that belongs at the start of the conversation. A customer service chatbot hears names, order numbers and sometimes health or payment details. Decide per conversation type what it may ask for and what it must not record. Choose a vendor that is clear about where conversations are processed and whether they are used for training. Substantive decisions stay with a person: when in doubt, the bot hands over.

The short version

Compare with your own twenty to fifty questions, not with a feature list. Four kinds: package, standalone platform, knowledge assistant on your own documents, voice agent. Work subscriptions through on your own volume and look at setup, data processing and handover. Let the bot answer only from your information, with sources, and test every change.

Sources Regulation (EU) 2024/1689 (AI Act)

Want to know which of your customers' questions a bot could handle today?

Describe your bottleneck →