A guidebook the chatbot reads from — not a chatbot guessing about your house.
SimpStay is our first AI-native product. Hosts composed a structured guidebook in plain language; an agent answered guest questions 24/7 by quoting from it. Nothing the host hadn’t written. Built and shipped end-to-end.
Short-stay hosts answer the same fifty questions every check-in. Wi-Fi password, AC instructions, where the nearest 7-Eleven is, what time recycling collection runs. Most hosts answer those questions ten times a week, in three time zones, on no sleep.
We didn’t build a chatbot. We built a structured guidebook the host edits in plain language, and layered an agent that retrieves from it — the property, the neighbourhood, and the guest’s actual reservation. The agent doesn’t guess. It quotes the host.
On the engineering side: a web authoring tool for the host. A retrieval layer over the guidebook plus reservation context. A guest-facing chat that opens from the booking link. An eval suite the host can run before publishing — every change tested against fifty stock questions before going live.
What we learned shipping this: AI products live or die in the boring layer. The retrieval logic, the eval cadence, the human escape hatch — that’s where the work is. The model swaps out every six months. The product is everything around it.