AI Assistant vs Scripted Bot

"Bot" is one word covering two different tools. One follows a diagram drawn in advance; the other understands phrasing. The difference shows the moment a customer writes something nobody planned for.

Scripted bot: a tree of answers

It follows a flow a human drew. "Press 1 to book, 2 for pricing." Every branch leads to the next question.

Strengths: predictability. It will not say anything unexpected because it physically cannot say anything that was not written in. For flows involving money and commitments that matters.

Weak spot: everything outside the diagram. Asked "can I come Saturday with my kid", a script answers "sorry, I did not understand" — and the person leaves.

AI assistant: understands the phrasing

It runs on a language model: it parses the meaning of a message and answers in its own words, grounded in the company's knowledge base.

Strengths: it handles unusual questions, typos and several topics in one message. "I want something for back pain, who should I see and what does it cost" is one ordinary request to it and three separate branches to a script.

Weak spot: without constraints it can answer confidently and wrongly. That is why an assistant is never released to "talk about anything": it gets topic boundaries, a ban on quoting prices outside the price list, and an obligation to hand over to a human wherever a mistake costs money.

Which costs more to maintain

The answer here is not obvious. A script is cheaper to launch, but every change to your services means redrawing branches. Six months in, the diagram is covered in exceptions and people are afraid to touch it.

An assistant is the reverse: more expensive to launch, but changes go into the knowledge base as ordinary text. A new service is a paragraph in a document, not a new branch in a builder.

Where scripts still win

Not everything should go to AI:

• payments and order confirmation — predictability beats flexibility;
• collecting strictly defined data (order number, date, quantity) — a form is more reliable than a conversation;
• legally meaningful wording — consents, refusals, terms.

The arrangement that works is usually mixed: the assistant works out what the person wants, then hands control to a rigid scenario wherever precision matters.

How to tell a script is enough

Take a month's sample of inbound messages and count what share fits into five to eight standard questions. If nearly all of it does, a script will do the job. If every third message is phrased its own way, a script will irritate people in exactly the cases where they were ready to buy.

What it costs

An AI assistant from €1,500 per project, plus support from €500 a month. Scripted scenarios often come inside a standard CRM implementation (from €1,690) because they are built in the native flow builder.

The gain is worth calculating not from the price of the tool, but from the share of enquiries currently going unanswered for more than an hour.

How to measure whether the assistant actually works

"It feels better" is not a metric. Four numbers worth tracking from day one, all of them counted in the CRM rather than by impression.

Share of self-contained conversations. How many enquiries closed without a human. This is the headline figure: it maps directly to freed-up time.

Handover rate and its reasons. A handover is not a failure, it is a designed path. What matters is the pattern: if 40% of handovers share one topic, there is a hole in the knowledge base, and it takes half an hour to fill.

First response time. The effect shows here earliest: minutes instead of hours. Track the median rather than the mean — one overnight enquiry ruins an average.

Conversion to the target action. A booking, a request, a payment. An assistant that replies politely and moves nobody to act costs money without earning any.

What not to do. Judging the assistant by message count: the better it is, the shorter the conversation. More messages per conversation usually means the person is asking again, not that they are more engaged.

The first month's numbers will look bad, and that is normal: a knowledge base gets written from real questions, not guessed in advance.

Price ranges are our rates at the time of publication.