Why your chatbot shouldn't make things up (and how to get there)

A general-purpose AI model always answers, whether it knows or not. Fine for conversation; in front of a customer, it's a risk: a made-up price, a cancellation policy that doesn't exist, a promise your business can't keep. The solution has an unglamorous name: forcing the model to answer only from your documents.

What “only from your documents” means, in plain English

Three conditions, and all three can be checked from the outside. First: the assistant can only use the documents you've uploaded as its source, not its memory of the internet. Second: every answer cites the document and section it comes from, so anyone can verify it. Third: when the answer isn't in the documents, it says so — “I don't know” — and escalates the question to a person, instead of filling the gap with something plausible.

A reliable assistant isn't the one that answers everything. It's the one that can say “this isn't written down” and turns that into an alert.

The questions you should ask any vendor

1. Can the assistant answer something that isn't in my documents? The only good answer is no.
2. How do I see where each answer came from? It must cite document and section, in every answer, not “in general.”
3. What exactly happens when it doesn't know? Ask to see the flow: who gets the alert, with what information, and how the answer ends up in the documentation.
4. Can I try it with real data before paying? A demo with documents loaded is worth more than any brochure.

How Concierge does it

Concierge applies all three conditions out of the box: it answers only from your documentation, shows the source under every answer, and turns every “I don't know” into an email to whoever you choose — answered once and added to the manual. You can check it yourself in the no-signup demo: ask something that isn't in the sample documents and watch what it does. What it does there is exactly what it will do with your customers.