The chatbot was live. Testing showed what was wrong.

The situation

The Groundlens chatbot was live and sounded convincing in conversation. That made the problem hard to spot. Testing found incorrect details about prices, timing, and how our offers related to each other. A connected tool could also create a CRM record and send messages even though the visitor was only talking to the chatbot.

This was not client work. It was our own build. The question was practical: how can a website chatbot be useful and natural without guessing about business facts or quietly taking action in the background?

Three versions, three different failures

The first build: The chatbot could answer freely and naturally. Testing found incorrect business facts, and its lead-capture tool was not bounded safely enough. Fluent language had created trust the system had not earned.

The deterministic rebuild: We removed almost all model authority over facts and answers. The system became more predictable, but it felt like a form. It was more accurate and less useful at the same time.

Version 8.1: We separated language from authority. The model may draft short, natural replies. Fixed code decides which Groundlens facts are inserted, which links and actions are allowed, and what happens when something fails.

Sounding natural is not evidence that an answer is correct.

What runs today

English and German use the same n8n path. GPT-4.1-mini drafts the reply. The browser sends up to eight recent messages and bounded conversation state. n8n keeps no separate conversation memory. A visitor can start over at any time.

The chatbot has no tools. It does not collect leads, write to HubSpot or another CRM, send email or Slack messages, or change customer or production systems. Real contact and qualification use separate forms.

OpenAI processes the current message and the recent conversation context sent with it. Visitors see that notice before sending and are told not to enter personal data, credentials, or confidential information. Completed n8n execution payloads for this chatbot are deleted after two hours by a scoped timer. The separate privacy notice describes the other provider retention rules.

What we tested

The final version passed 22 of 22 workflow tests and 66 of 66 website tests. We also checked the live English desktop version across two conversation turns and the live German version at a mobile viewport. Both completed without browser or console errors.

Those results apply to the tested paths. They are not a guarantee for every possible question, and they do not prove conversion, revenue, or time savings. The evidence is narrower: the system has named boundaries, those boundaries can be tested, and the live version matches the build we verified.

Why this matters beyond a chatbot

The same problem appears wherever AI works with business facts or actions. A model can answer a customer question, summarize an order, or suggest the next step in a recurring task. The wording can remain flexible. Prices, approvals, permitted actions, and failure paths need a clear source and fixed boundaries. The chatbot is simply a visible example of that separation.

What remains human

A person still defines the business facts, reviews changes, and decides what the system may do at all. Sales conversations, unusual requests, and decisions with consequences remain outside the chatbot. When the system cannot answer safely, a clear boundary is better than a confident guess.

The live chatbot is available on groundlens.io if you want to try it. If you are considering AI for recurring work in your own business, the separate fit check starts with the work rather than a preferred tool.