Every team that has watched a chatbot invent a discount, a policy or an opening hour knows the moment: the bot sounded confident, the customer believed it, and now a human cleans up.
We built SayTrue around the opposite bet — that in customer service, the refusal is a feature. The agent answers only from knowledge the business approved. When no approved source covers a question, it says so, hands the conversation to a person, or records the request. Every answer carries a source log, so a supervisor can check which entry produced it.
This costs something: our agent answers fewer questions than an improvising one. It also means the answers it does give can be repeated with a straight face — to a guest, a manager, or an auditor.
Three honest outcomes — grounded, escalated, no source — are the whole vocabulary. The reports count them without decoration, the gap log turns the refusals into a to-do list, and the knowledge base grows exactly where customers push it.
That’s the product. The demo call is where you test it: bring the questions your business finds hardest, and watch what it does when it doesn’t know.