AI can draft a guest reply without ever knowing the specifics of your property — but the more generic the draft, the less useful it is to the guest. The single biggest determinant of whether an AI-drafted guest message is actually useful is whether it was generated against a property-specific knowledge base: the room layout, the parking rules, the Wi-Fi password, the late-arrival procedure, the closest pharmacy, the appliance quirks that matter.
Below is what belongs in a property knowledge base, how to keep it current, and why the operating model has the operator maintaining the source rather than letting the AI fill in the gaps. The point is not to maximize the volume of stored information; the point is to reduce the number of moments where the AI has to guess.
What belongs in a property knowledge base
Four kinds of information show up in nearly every guest conversation, and they are the four kinds of information the property knowledge base should hold.
Arrival logistics. Address confirmation, parking instructions, the access method the property uses (smart lock, lockbox, in-person handoff, code-based), door codes and Wi-Fi credentials where they are appropriate to share in a check-in message, late-arrival procedure for guests whose travel plans slipped, and any building-specific rules about elevators, loading docks, or quiet hours.
Stay logistics. Trash and recycling procedure, where to find linens, how to operate the heating and cooling, which appliances have quirks worth documenting, bathroom fixtures that confuse guests (low water pressure in a third-floor bathroom, a kitchen disposal that needs a specific sequence), and the checkout procedure including any timed steps required by the cleaning team.
House rules and expectations. The property’s occupancy expectations, where smoking is and is not permitted, pet policy, party policy, and any noise rules the property enforces. These are not legal terms; they are the specific operational expectations the guest should know about before arrival.
Local context where appropriate. The closest grocery, the nearest pharmacy, the typical wait at the recommended breakfast spot, the public-transit option that is actually useful at this address. Optional, and bounded — the goal is guest-useful local context, not a city guide.
Where the source information comes from
The starting point is the property record the operator already maintains — the listing copy, the internal notes the team uses for turnovers, the welcome-book PDF if the property has one, and the local-knowledge tidbits that come up repeatedly in guest conversations. None of that needs to be re-entered; the operator is curating, not starting from scratch.
New entries show up in two places. The first is the moments a guest asks a question the property record does not currently answer — those moments are a signal that the source is missing something real. The second is the moments a draft reply is edited before it goes out, because the operator’s edit is usually fixing a place where the source was wrong or incomplete. Both are worth capturing back into the knowledge base after the guest has been served.
The property knowledge base is also where the operator’s style profile lives — the tone preset, the signature, the small style notes that make the AI-drafted reply read like a real version of the operator rather than a generic assistant. The same operator has a recognizable voice; the knowledge base is what carries that voice into the AI’s draft.
Version control and timely updates
A property knowledge base that drifts out of date is worse than one that is empty. A wrong Wi-Fi password is worse than no Wi-Fi password — the wrong one sends the guest through a longer, more frustrated round of troubleshooting. A wrong check-in time, a stale house rule, or a building amenity that no longer exists creates the same shape of problem.
The fix is version control, applied lightly. Each entry has an effective date. Each entry has a clearly named source (the property manager who added it, the welcome-book revision it came from). When an update is made, the previous version is preserved so the team can roll back if the update turns out to be wrong. When a piece of information becomes obsolete (the property switched from a key handoff to a smart lock), the entry is marked as historical rather than deleted, so an old draft surfaced during a search does not lead a guest astray.
A property operations calendar usually drives the timing. The trigger to update the knowledge base is the same trigger that drives any per-property change: a Wi-Fi router replacement, a cleaning team rotation, a building renovation, an ownership transition. Each of those is a natural moment to check the property record against reality and update what no longer matches.
Why inaccurate source information produces inaccurate drafts
The relationship between source and draft is straightforward. The AI’s draft is only as good as the source it reads from. If the source says the parking spot is on the north side of the building and the sign actually points to the south side, the AI will draft an answer that sends the guest to the wrong side. If the source says the kitchen disposal works a certain way and the disposal was replaced last week, the AI will draft an answer that leads the guest through a procedure that does not work.
The pattern is consistent across every kind of property detail. The wrong source produces a confidently-written draft that is confidently wrong. The guest trusts the draft more than they would trust a "let me check and get back to you" reply, which means the wrong draft causes more frustration than the right silence would have.
This is also why the operator maintains the source rather than letting the AI keep itself current. The AI cannot observe a Wi-Fi change, a lock replacement, or a new house rule on its own. The AI can flag inconsistencies when the source drifts, but the actual update is a human job — a human looking at the property, talking to the team, and writing the new entry.
What stays outside the AI’s prompt
Some property information is real and important but does not belong in the prompt the AI guest inbox reads from. Sensitive access codes, restricted-area instructions, and any information a guest should not see should stay in a separate record owned by the operator. The guest inbox never reads it.
Some information is internal team context — escalation owner per property, vendor relationships, the local locksmith on the rotation. That is information the operator needs but the guest does not, and the prompt the AI guest inbox reads from should not surface it. A separate record carries that knowledge and is read by the routing layer rather than the guest-reply layer.
The same separation applies to compliance and legal items. The property knowledge base can include the property’s compliance posture in broad terms (the jurisdiction, the operating notes the team uses to keep things current), but the rules themselves — the exact permit regime, the city’s cap mechanics, the per-quarter TOT reconciliation steps — stay under the operator’s exclusive authority, with the AI guest inbox reading only the operator-approved summary.
How the operator keeps the knowledge base healthy
The right cadence is small and frequent. After a check-in day where the team noticed a guest question the source did not answer, the source gets a new entry. After a draft was edited before it went out, the edited line is folded back into the source. After a Wi-Fi change, the password entry gets refreshed. After a building renovation, the amenities list is updated. Each of those is a small update; together they keep the source trustworthy.
A quarterly review is worth scheduling, not as a deep audit but as a surface pass. The operator reads the property record the way a new guest would read it, finds the parts that are stale, and updates them. The review is the surface that catches the slow-moving drift — the house rule that quietly stopped being enforced, the local place that closed, the amenity that has been offline for two months and is unlikely to come back.
The health metric is not how much information the knowledge base contains. It is how often a guest reply does not need to be edited before it goes out, and how often the operator’s source update catches a real guest question the AI could not answer before the update. The metric gets better when the operator’s maintenance routine is sustainable; the metric gets worse when the source drifts faster than the operator can update it. The system is doing its job when the AI drafts are right often enough that the operator spends the saved time on the conversations that actually need them.