The standard pitch for AI in hospitality is: guests can ask anything and get an instant answer.

That sounds good until a guest asks about a restaurant you'd never recommend and gets a confident suggestion, or asks what time check-in is and gets a generic "3pm is typical" instead of the actual answer. Both failures are worse than no AI at all — they create false confidence.

For Cayo Basecamp, a guest concierge for short-term rental hosts in Belize's Cayo district, the design requirement was the opposite: the AI should know exactly what the host has approved, and nothing else.

Why capability is the wrong axis

A host running properties in San Ignacio has specific answers to specific questions. Check-in is at 3pm unless you've arranged early arrival. The driveway, not the street. For dinner, Martha's Kitchen — not the resort buffet with the tourist markup.

An AI that searches the web might find that a recommended place closed, or surface reviews that conflict with what the host actually wants to say about the property. It might hallucinate a check-in time based on industry norms. Each of those erodes exactly the thing a host is trying to create: a guest experience that feels thoughtfully curated and personally considered.

The concierge in Cayo Basecamp answers only from a guidebook the host maintains in the admin dashboard. If a question falls outside that guidebook, the AI says so — and offers to flag it for the host. That's the honest behavior. It's also the feature, not a limitation.

The gap detection loop

Failed answers are the most useful signal the system produces.

When the AI can't answer a question, it logs it. The admin dashboard has a gap panel: a ranked list of questions guests asked that the current guidebook doesn't cover, grouped by theme. The host reviews the clusters, writes or refines the relevant section, and the next guest who asks the same thing gets a real answer.

Guest asks → AI can't answer → host sees the gap → host adds content → AI answers.

The guidebook improves based on what guests actually ask, not what the host guessed they'd ask before a stay. That's a different quality of iteration.

Dogfooding before pitching

The platform is live for my own Cayo properties. That's deliberate. Before charging another host $19 a month for this, the metrics need to prove it works: activation rate (what percentage of bookings actually unlock the hub), AI questions per stay, gap panel resolution rate.

If activation is low, the onboarding flow is the problem. If the gap panel is growing faster than it's being resolved, the guidebook editor is too heavy. Both are things to fix before someone else's guests depend on it.

Dogfooding a hospitality product means staying in your own rental. I'm working toward that. The point is that the platform gets tested against real guest behavior, not a demo scenario I controlled. The gap panel tells me things I wouldn't have known to look for.