Can the AI make things up?
In plain language
An AI assistant can make mistakes. That is why it is meant to answer only from your own content. If the answer comes from one of your texts, it names that text. If it finds no fitting text, it usually says so. If the language model fails, a prepared answer is sent. From the Pro package you see open questions in the report. There you add the missing knowledge.
Answer
An AI assistant can err, that cannot be ruled out. That is why we bind its answers to your own content: it answers from your website, documents and knowledge base, instead of formulating freely into the blue. If a piece of information is not on file, it says so and points onward.
In addition, handover to a human and the ongoing conversation analytics provide safety: knowledge gaps and false assumptions become visible and can be closed specifically. This keeps the information traceably bound to your sources.
The binding is not merely a request to the model: sourced statements carry the number of the passage they come from, and an answer without a single used citation is recorded as a gap. The assistant replaces figures, prices and deadlines that are not in your sources with a notice. If the language model fails, the conversation sends the stored fallback answer and not an invented one.
What stays open is not kept quiet: unanswered questions and knowledge gaps stand as separate lists in the evaluation, together with the question that triggered them. From these you extend the knowledge base in a targeted way instead of setting the assistant up again. In addition, a probe set runs daily over known questions and measures whether they still hit the right passage; an account without probe lines deliberately produces no value rather than an estimated one.
Ask it yourself, Assistant of this website
Automated assistant