AI chatbots:
what actually
works.

Users want answers, not conversations.

Most people who message a chatbot have a specific question and want an answer within 30 seconds. They don't want a charming welcome, don't want to chat about their concern, don't want to click through a dialogue menu.

Takeaway: The bot should immediately ask "What are you looking for?" and answer as directly as possible. No welcome frills. No form-like questioning when one question will do.

Bots with long welcome texts ("Hi! I'm Lisa, your digital assistant at Müller GmbH, and I'm delighted to help you...") get clicked away after 5 seconds measurably more often than those that get straight to the point.

The first line decides everything.

Chatbot usage follows a brutal function: if the user's first attempt doesn't lead to a usable answer, 68% leave the bot. If the first attempt works, 84% stay and ask more questions.

Takeaway: The very first response turn has to land. It mustn't ask back, mustn't rephrase, mustn't "clarify" — it must deliver. That means: optimise the first 20 intents in full depth. Further refinements are polish after that.

In our practice that means: we invest 40% of the entire bot development time in the top-5 intents. It always pays for itself.

No escalation, no trust.

A chatbot that doesn't offer a "Let me hand you to a human" inevitably frustrates. Sooner or later there's a question it can't answer — and then there has to be a clear way out.

Takeaway: Build in three triggers for escalation:

1. User frustration: words like "human", "real person", "useless", "unhelpful" — escalate immediately.
2. Complexity: if the bot can't answer clearly twice in a row — offer escalation.
3. High stakes: always escalate medical, legal, contractual topics.

Escalation mustn't mean "Here's our email address". Better: a direct handover to a human (live chat, a callback form with a specific time window, an SMS to the boss).

FAQ is not enough training.

Many businesses think: "We have an FAQ on the website, we'll copy it into the bot, done." That gives a bot that fails on every unexpected phrasing.

Takeaway: Training must be based on real customer questions, not on what the business thinks is being asked. Our method:

• Log all incoming email and phone enquiries for 2 weeks
• Extract the 50 most common phrasings — not the "official" FAQ questions, but the word choices of real customers
• Train on these phrasings, with 3–5 variations per intent

A bot trained on real customer speech, in our tests, correctly understands over 90% of all enquiries. One that only knows FAQ text often doesn't get past 70%.

After launch comes the refining.

A chatbot isn't "built finished and shipped". It goes live and is then optimised weekly — especially in the first 6 weeks. Every "the bot didn't understand" is an optimisation signal.

Takeaway: We agree a 6-week stabilisation sprint with clients after launch. During this time we review the transcripts weekly, identify weak points and build them out. After week 6 the bot usually runs stably and only needs monthly checks.

Skip the refining after launch and you have a bot stuck at a 75% comprehension rate. Refine, and you reach 90%+ — and that is the point at which real users get real value.

Chatbots work. If you build them right.

Most of the chatbot disappointments we hear about from new clients aren't down to the tool (Claude, GPT-4 are excellent by now), but to the build process. Click a bot together from an FAQ in 2 hours and you get a bot that gathers dust in the corner. Take the 5 points above seriously and you get a tool that takes real work off your plate.

Most of our chatbot projects pay for themselves within 4–6 months — measured in staff time saved on routine enquiries. That's not AI hype, that's craft.