Customer service chatbots your customers actually use.

When is a customer service chatbot worth it?

Two conditions have to hold at once. The volume of enquiries has to be high enough for automation to move anything. And the enquiries have to carry a friction you can remove, which means recurring questions with answers you can maintain.

If one is missing, the effort does not pay. A bot on fifty enquiries a month costs more upkeep than it saves. A bot on a thousand enquiries whose answers exist in no maintainable state produces wrong information at scale.

The third condition is the handover. Bitkom Research asked in May 2025 how satisfied people are with service channels: 86 percent are satisfied with human contact, 50 percent with chatbots. 62 percent want to speak to a person first when they have a problem. That does not mean chatbots are useless. It means the moment the bot hands over decides how the whole thing is judged. Leave that moment undesigned and you lose exactly the customers who have a real problem.

What comes before the technology?

1. 1

   Define the use case.

   Which ten enquiries arrive every week? Your ticket system decides that, not your product catalogue.

2. 2

   Inventory the knowledge sources.

   Terms and conditions and contracts are written for legal interpretation and make poor answers until someone rewrites them for customers.

3. 3

   Settle the GDPR setup.

   Consent strategy and hosting model come before the platform choice, because they narrow it.

4. 4

   Check the integration.

   A bot that cannot open a case answers questions and resolves nothing.

5. 5

   Fix the metrics.

   The baseline is measured before go-live. Afterwards it cannot be recovered.

6. 6

   Assign responsibility.

   Prices, seasonal offers and regulations change without regard for your project plan.

7. 7

   Define the pilot scope.

   Three to five intents that work beat twenty that half work.

How do I measure success?

Start with the baseline. How many enquiries arrive today, through which channels, how long does handling take, how satisfied are people afterwards? Without those numbers there is no later comparison, only opinions.

Then pick the right measures. Resolution rate and satisfaction say more than the raw ticket count. A chatbot is first of all an additional channel, and an additional channel raises enquiry volume in the short term. Set volume reduction as the first goal and you will declare a working project a failure.

Watch the incentives you create. A measure that rewards the bot for every enquiry it does not pass on produces a bot that does not pass anything on. Check what your vendor contract measures and bills, because that shapes the reports you will read later.

And do not plan for headcount reduction. Gartner found in February 2026 that half the companies that cut service roles citing AI are expected to hire again by 2027, and that only about a fifth of those cuts were actually driven by AI. A bot that works strengthens the team. It does not replace it.

Voicebots

Same decision logic, different channel. In many industries the telephone is still the most used service channel, and high volumes of recurring enquiries make the same case there as in chat.

The demands are higher. Response times have to be short or the conversation breaks. Speech recognition has to cope with accents, background noise and half-finished sentences. And dialogue design carries more weight, because a caller cannot scroll back.

High volume, recurring standard enquiries. Telecommunications.

Tariffs, billing, outages and moves produce the same questions in large numbers, which is the clearest case for automation there is. It also produces the sharpest line between answers the AI may generate and answers that have to be written and approved.

Explanatory content with little liability attached the bot can phrase itself. Anything binding, or anything that sounds binding, is written editorially and signed off. The rule of thumb: the more a wrong answer costs, the more editorial the answer.