Why We Replaced Our AI Chatbot With Seven Scripted Questions

We build AI tools for a living. Last week we deleted the AI from our own website.

The configurator on contriboot.dev used to be a real LLM chat — streaming responses, conversation memory, technical-level detection, the works. It was genuinely well built. We replaced it with seven pre-written questions and a contact form.

Here's why, and why you might want to do the same before you bolt a chatbot onto your site.

//The job was never "conversation"

An intake chat has one job: figure out what the visitor needs, qualified enough that the follow-up is useful. That's it.

When we looked at our real sessions, the LLM wasn't discovering anything a fixed question couldn't. Every productive conversation converged on the same information:

• What kind of problem is this?
• What's broken or manual today?
• What tools are involved?
• Who's going to use the fix?
• How urgent is it?
• Is there budget behind it?

That's not a conversation. That's a checklist wearing a conversation costume.

//What the LLM was actually costing us

Token spend on every visitor. Including the bored ones, the bots, and the competitors poking around. A public LLM endpoint is a metered faucet anyone can turn on.

Unpredictable answers. Ours was well-prompted and stayed on script most of the time. "Most of the time" is doing heavy lifting in that sentence. Every public LLM surface is a place where your business can say something you didn't approve.

A gate in front of the value. To protect the spend, we required an email before the chat started. So the visitor paid the toll before they got anything. Backwards.

Latency theater. Streaming tokens look impressive, but the visitor is waiting on a network round trip for a question we already knew we were going to ask.

//What the scripted version does better

The new flow asks the same seven questions every time, in the same order, with quick-tap answers where the options are known. It still feels like a terminal conversation — typing effect, message bubbles, the whole thing. Visitors don't experience a form. They experience a fast chat that never rambles.

The differences underneath:

• Zero per-visitor cost. It's a state machine. It runs in the browser.
• Contact info moves to the end. You answer the questions first, then tell us where to send the quote. Finishing the flow is the commitment, not starting it.
• Every brief is complete. No conversation that wanders off before asking about timeline. Seven questions in, seven answers out, every single time.
• Nothing to jailbreak. There is no model. The worst a malicious visitor can do is answer questions weirdly.

The briefs land in our queue, we review every one personally, and the reply is a custom quote — usually same business day.

//When you actually need the LLM

This isn't an anti-AI take. We deploy LLMs in client builds constantly. The rule we use:

Scripted wins when the information you need is known in advance. Intake, qualification, booking, support triage with a known issue tree. If you can write the questions down, write them down.

LLM wins when the input space is genuinely open. Support across a large product surface, document analysis, research, anything where the next question depends on understanding a long messy answer.

Most website chatbots are doing the first job with the second tool. That's paying inference prices for a checklist.

//The uncomfortable question

If your chatbot vendor charges per conversation, ask yourself what percentage of those conversations follow a path you could have written on a whiteboard in twenty minutes.

If the answer is "most of them," you don't have an AI problem. You have seven questions and a form.

Want to see the result? Run through ours. It takes about two minutes, there's no account, and a human reads every brief.