AI CallerIt answers from your material, or it says it does not know
The failure everyone fears with AI on the phone is confident invention. The fix is not a better model, it is a hard boundary.
Three deliberate constraints
Approved sources only
You choose the documents and pages. It cannot reach past them.
A real not-found path
When the answer is not there it says so and offers a person.
Answers are traceable
Every answer records which source it came from.
Gaps are surfaced
Questions it could not answer are collected so you can fill the gap.
Nothing quoted off-book
Prices come from structured fields, not prose it might misread.
Updates apply immediately
Change a price and the next call uses it. No retraining.
Knowing when to stop is a feature
It cites what it used
Every knowledge answer records the source behind it, so you can check afterwards whether the answer was right and where it came from.
Gaps become a to-do list
Questions it could not answer are collected and ranked by how often they came up, so you know exactly what to add next.
Escalation is the other half
A boundary is only useful if there is somewhere to hand the call. Every not-found path leads to a person, never to a dead end.
You can audit it
Filter past calls by knowledge answers given, so a wrong price quoted once can be found and corrected rather than repeated.
This is why escalation rules matter as much as the knowledge base. An agent that admits the limit of what it knows is worth more than one that always has an answer.
Worth asking before you buy
What sources can it use?
How do I update it?
What happens if two sources disagree?
Can it read my website?
What languages can the knowledge be in?
Try to make it invent a price
On the demo call, push it on something it should not know. It is a good test.