Teaching It to Say “I Don’t Know”
A chat assistant is only useful if it answers from your information and admits when it does not know.
- Published
- Read time
- 7 min read
Almost every business I talk to wants "AI" on their website. What they usually mean is simpler and more useful: a way to answer customers' questions at any hour, correctly, in the company's own words.
The risk is that a general AI model will happily make things up. For a mining supplier quoting products, or an agency describing its services, a confident wrong answer is worse than no answer at all.
Give it a library, not a licence to guess
The approach I use is called retrieval. Before the assistant answers, it searches the company's own documents for the passages that matter, and it is asked to answer only from those. If the answer is not there, it says so and points the person to a human.
- Collect the real material first: product sheets, service pages, FAQs, policies
- Break it into small pieces and store them so they can be searched by meaning, not just by keyword
- Answer from what is found, and nothing else
- Save the conversations, so the team can see what people actually ask and improve the content
The best assistant is a well-organised library with good manners.
What it changes for a business
Done well, this does three things. Visitors get an answer at 11pm on a Sunday instead of waiting for Monday. The team stops repeating the same replies. And the saved questions become a free map of what customers are confused about, which usually tells you what to fix on the website itself.
Where I draw the line
An assistant should never be the only way to reach a person, and it should never invent prices, promises or availability. I would rather ship a modest assistant that is always right than an impressive one that is sometimes wrong.
