Language models are trained to produce fluent, plausible text. Ask a general chatbot about your business and it may produce a plausible price that you never charged. In AI this is called a hallucination. For a business it is simply a wrong answer given with confidence.
Ground every answer in your own information
The most reliable fix is to give the assistant your information and instruct it to answer only from that. This is called grounding, or retrieval: before replying, the system finds the relevant parts of your content and the model writes its answer from them.
Selo Assist works this way. It answers from your website pages, files, FAQs and notes, and nothing else.
Make "I don't know" an acceptable answer
An assistant should be allowed, and expected, to say it does not have the information. The right follow-up is to offer a person, not to guess.
In Selo Assist, when the answer is not in your information, the assistant tells the customer and offers to connect them with your team, and the chat is flagged in your inbox.
Check the sources behind answers
Before going live, test the assistant with real questions and look at where each answer came from. Selo Assist's test chat shows a confidence level and the sources behind each reply.
If an answer is wrong, the fix is usually in the source: an outdated page, two documents that disagree, or a missing fact.
Keep your information consistent
If your price list says one thing and an old blog post says another, any assistant can be confused. Remove outdated content from what you give it, and re-sync your website after changes.
Guard against tricks
Some visitors try to make chatbots say silly things or ignore their rules. A business assistant should resist instructions hidden in customer messages. Selo Assist has protections against attempts to make it ignore its rules.
