Every AI-generated reply is checked for claims the source material doesn't support, for policy, and for tone — before it sends.
Established QA tools sample conversations after they happened and score agents for coaching. Useful, but the customer has already read the wrong answer. ReplyCheck runs in front of the send, so an invented refund policy is caught while it is still preventable. Replies that fail are held for a human; everything else goes out.
Illustrative. The threshold is drawn because the argument is that a threshold exists at all.
"You're covered by our 60-day money-back guarantee, so I've started your refund."
Unsupported claim. No source in the knowledge base mentions a 60-day guarantee. Held for a human before sending. Illustrative sample.
Each held reply says which check fired, so the assistant gets better instead of merely getting blocked.
The failure that matters is confident invention — a plausible policy, date or figure with nothing behind it, which reads exactly like a correct answer.
What the assistant may promise, refund, escalate or commit to — written as rules the gate enforces, not as instructions in a prompt and hope.
Curt where you are warm, or over-apologising where you are direct. Lower stakes than the first two, and still the reason a customer notices they are talking to a machine.
A held reply is a repair instruction. You see the condition and the passage that tripped it, so the next version of the assistant is better than this one.
Competitor rates are their own published prices.
Founding licences are one payment, for good. Everyone after launch pays monthly — still flat.
Not ready to reserve?
One email when it ships. No spam.