We build the opposite. Grounded in your content, honest about its limits, and wired to hand off to a human the moment confidence drops. What you measure at the end of month one: fewer tickets, faster resolution, and zero embarrassed-founder screenshots on Twitter.
Retrieval returned 3+ relevant chunks. Answer + citation.
Some retrieval, gaps in coverage.
Retrieval weak or ambiguous.
No relevant chunks. Out of scope.
User says 'talk to a person' or 'agent'. Never argue.
Two consecutive negative sentiment turns. Escalate quietly.
Retrieval score falls below threshold two turns in a row.
Regex + intent classifier. Humans handle money-and-anger.
Same intent asked 3× → the bot's answer isn't landing.
Something's not resolving. Bring a person in.
Logged-in enterprise user or cart above threshold. Priority queue.
Three numbers we report weekly: containment rate (% of conversations resolved without a human), CSAT on bot-only conversations, and hallucination-flag rate (answers the bot marked as low-confidence). No CSAT drop, no vanity metric wins.
The prompt is the product. If it's five lines of 'You are a helpful assistant', you don't have a chatbot — you have a leak.
Tell us your goals — we'll recommend the right technology honestly, even if it isn't this one.