The failure that costs real money is not a bot that says "I do not know". It is a bot that confidently quotes a price you do not charge, or a policy you do not have, to a customer who then holds you to it.
Here is what causes it and how to configure against it.
Why it happens
A language model produces plausible text. That is the whole mechanism. Ask it about your call-out fee and, with nothing to draw on, it will produce a number that looks like a call-out fee, because a number is what a plausible answer contains.
It is not lying. It has no concept of your prices, so it generates the shape of an answer.
The fix: grounding
Retrieval-augmented generation, in the jargon. In practice:
- Your website and documents are split into chunks and indexed.
- When a visitor asks something, the most relevant chunks are retrieved.
- The model is instructed to answer only from those chunks.
- When nothing relevant comes back, it says so.
This is standard in any serious product. If a vendor cannot explain how their bot is grounded, that is your answer.
The four settings that matter
Restrict to your content. There should be an explicit instruction that it answers from the supplied context and nothing else.
Require an admission. It must have permission to say "I do not have that information". Bots configured to always be helpful will always produce something, and something is exactly the problem.
Give it a route out. Admitting a gap is only useful if there is somewhere to go: capture the enquiry, open a ticket, fetch a human.
Surface the gaps to you. The best products log the questions they could not answer. Each one is a specific thing to add to your website, and it turns the failure into a to-do list.
The part that is your job
Grounding cannot invent content you have not written. The most common cause of a vague, unhelpful bot is a thin website.
If your site does not state your prices, your service area, or your policies, the bot cannot either. Writing that content helps your customers and your search rankings regardless of whether you deploy anything.
How to test properly
Ask it, deliberately:
- A price you do not publish
- A service you do not offer
- A location you do not cover
- Something about a competitor
Correct behaviour is a clear "I do not have that" plus a route forward. Anything invented, however plausible, is a product you should not put in front of customers.
Test it on your own content
Build one from your website and try to make it invent something. That is a more useful evaluation than any feature list.