I run a short-term rental portfolio alongside the consulting work. At the time of this project it was 12 listings: five in Florida, four in the Tennessee Smokies, and three in Arizona. Guest messaging runs through Hospitable, which sits on top of the booking platforms and handles the automated messages every stay gets.
The goal was to eventually let an AI draft guest replies. I had one rule going in. Hospitable sends a message to the guest the moment you hit send, with no draft step and no undo, so nothing would go out without me approving it first. And before I'd approve anything, I wanted the AI answering from what we had actually told guests over the years. A generic template wasn't good enough.
What was breaking
Nothing was on fire. The problem was quieter than that. Guest questions arrive at bad times: a late check-in on a Friday night, a hot tub question during dinner, a parking question while I'm on a client call. The answers lived in my head and in old message threads. Hospitable has a Knowledge Hub feature for exactly this, and ours was empty for every property.
The typing was never the cost. The interruptions were, along with the fact that if I was unreachable, nobody else could answer.
What we built
Step one: connect. Hospitable publishes an MCP server, which is a way for an AI assistant to call an application's API directly. I connected it read-only to start: properties, reservations, calendars, and the full message history for every reservation.
Step two: pilot one property and lock the format. I picked one listing and built its FAQ with the AI, then adjusted until it read right. The format that stuck has three parts:
- A quick-facts table at the top: check-in and checkout times, wifi, parking, pool, pets, trash day, and the other things that get asked every week.
- Grouped questions and answers, written in my voice, the way I'd actually reply.
- An escalation tag on every item: "safe to auto-answer" or "always escalate."
Once the format was approved, every other property used the same structure.
Step three: read everything. The tempting shortcut is to sample a few dozen threads per property. I decided against it. If this knowledge base is going to answer guests without me, it needs to have seen the odd questions, and the odd questions are the ones a sample misses. So the rule was exhaustive: every thread, every property.
That's a lot of API calls. Hospitable allows roughly 100 per minute, so the read ran as a paced background job, with the work split across several AI subagents so no single one had to hold a whole property's history at once. The first pass produced a FAQ file for all 12 properties, but the six busiest were under-read because of the rate limit. A second pass went back and read every remaining thread for those six: 1,574 threads in total, 362 at the busiest and 168 at the quietest of the group. Each of those FAQs came out at 45 to 65 items. The knowledge base ended up complete for 11 of the 12.
The twelfth had no message history in the system at all, which meant it was either newly onboarded or not syncing. Its FAQ was built from the listing text alone and flagged as unverified. That flag paid for itself. The listing turned out to quote two different late-checkout fees and two different smoking fees, and it described a shared resort pool as private. Fixing the listing came before letting anything quote from it.
What the messages actually contained
This was the surprise. Most of the message volume was us. Booking confirmation, pre-arrival details, check-in instructions, the trash reminder, checkout instructions, the review request. Every stay generates the same handful of automated messages, and they dominate every thread.
The guests' own questions were sparse. A thread might have eight automated messages and one real question. But that one question was what we were reading for. It's where the answer we'd never written down lived.
Two lessons came out of that. First, if you feed all of your message history to an AI without separating your own templates from the guests' questions, you'll train it to sound like a confirmation email. Second, reading is the cheap part. The rare, specific question is the valuable part, and you only find it by reading everything.
The read surfaced other things too. Recurring maintenance items that guests had mentioned more than once across different stays, which went onto a standing fix list. A couple of listing details that didn't match what we'd been telling guests. That wasn't the goal, but it was worth knowing.
How it runs day to day
Today the FAQs live as files in a repository, reviewed by me, and are being loaded into each property's Knowledge Hub in Hospitable so the answers are available to anyone on the team, with or without an AI in the loop.
Alongside them is a response rules file: a short list of the question types the agent is allowed to answer, and the rule that everything else escalates to me. Anything involving money, damage, a complaint, or safety escalates no matter how simple it looks.
The agent also works from a hard allowlist of active property IDs. We've since sold or stopped managing some of the original 12, and their data still lingers in the account. The agent can't draft or send for any property that isn't on the list.
The next stage is the one the FAQ was built for: detect a new unanswered guest message, draft a reply from that property's FAQ, show it to me, and send on approval. That isn't switched on yet. Two things come first: a reliable scheduled trigger, and proof that the connection survives an unattended run before anything auto-sends.
What changed
- Every answer we've ever given a guest, across 11 properties, is written down, reviewed, and in our voice.
- Every answer is tagged as safe to automate or not, before any automation exists. That decision gets made calmly instead of at 11pm.
- Listing errors that could have led to a wrong quote got fixed at the source.
- A maintenance list that had only ever existed as scattered guest comments now exists as a list.
What this means for you
If you answer the same questions over text, email, or a messaging platform, your sent history is already a knowledge base. It just isn't organized. The process here works for any business with message history: rentals, clinics, salons, service companies, anyone with a shared inbox.
I wrote a step-by-step version in how to turn your message history into a FAQ your team or an AI can answer from. If you'd rather talk it through, book a free 30-minute audit and we'll look at where your questions come from and what could answer them.