Case study
Answers in your voice
The problem
The owner was answering the same customer questions all day, and the answers had to be exact — theirs, not an approximation. When they were busy or offline, people waited, and a wait is where you lose them. A generic chatbot would have been worse than nothing: a confident wrong answer puts words in your mouth.
What we built
An assistant that answers instantly in the owner's own words, built only from what they had actually said. Its real work is knowing when not to answer: every reply clears a confidence bar, and a near-miss escalates to a person instead of shipping as a fact. It repeats the owner, or it stays quiet. It never improvises.
What it did over ten weeks
- Wrong-answer rate
- 33% → 5%
- Questions handled
- 667
- Answered with no AI guess
- 29%
- A correction goes live
- Same day
Measured on every answer a customer rated
From 87 people, around the clock
Returned word for word in the owner's own answers
With tests so the mistake can't come back
How we did it
- A knowledge base built only from the owner's own words, so every answer traces back to something they said
- A confidence bar every answer clears before it sends
- Near-misses escalate to a person instead of guessing
- Every answer rated by the customer, so quality is a number we watch, not a hope