Who Owns the Knowledge Base When It Goes Stale?

An AI assistant is accurate on the day it launches. That is the last day anyone can take its accuracy for granted.

What follows is not dramatic. No system fails. Nothing throws an error. The property changes in a hundred small ways and the knowledge base does not, and one day a guest is told confidently about a service that was withdrawn in the spring.

How drift actually happens

Nobody decides to let a knowledge base go stale. It happens through ordinary operational change, where each individual item is too small to trigger anyone's sense that something needs updating.

The shuttle now runs on a different schedule. The breakfast supplier changed and one of the dietary options went with it. A renovation closed the smaller pool for a season. Prices went up in January. The spa stopped taking walk-ins. The number for lost property changed when the desk moved.

Every one of those was communicated internally. None of them felt like a technology change, so none reached the assistant.

The result is an assistant that is broadly right and specifically wrong, which is the most dangerous accuracy profile there is. Guests trust it because it is usually correct, then act on the part that is not.

The ownership question

The reason drift persists is almost always that no one owns the content. The assistant was implemented by a project - a vendor, an IT contact, whoever ran the rollout - and when the project ended, ownership of the answers ended with it.

Three roles need naming, and at a small property they are the same person:

Content owner. Keeps the answers true. Needs to be close to guests, which usually means front office rather than IT or marketing, and needs authority to change an answer without convening anyone.

Escalation owner. Watches what the assistant hands over during their shift, on a channel they actually use.

Boundary owner. Decides what the assistant is allowed to say at all - which topics it must never answer automatically regardless of confidence. This is a management decision, not an operational one.

The point is not the org chart. It is that each of these is currently owned by someone specific or by nobody, and nobody is the common case.

Where the update signal already exists

The most efficient mechanism is not a new process. It is attaching the assistant to processes that already run.

Anything that already has a checklist when it changes - prices, opening hours, seasonal schedules, service changes, contact details - should have the assistant added to that checklist. You are not asking anyone to remember something new. You are adding one line to something they already do.

Where no checklist exists, the fallback is a standing monthly review, which works but relies on someone caring about it in six months' time. Checklists survive staff turnover; good intentions do not.

Read the escalations

The other source of truth is free and largely ignored: the conversations the assistant could not handle, and the ones where a guest corrected it.

Those records are the highest-signal feedback you have about your own operation. An escalation is a guest telling you, in their own words, that something they expected to be answerable was not. A correction is a guest telling you something has changed and you did not notice.

A monthly pass over both, taking twenty minutes, catches most drift before it reaches a complaint. Look for:

• Repeated escalations on one topic. That is a gap in the knowledge base, and a frequent one.
• Guests correcting the assistant. Something changed and nobody updated it.
• Questions the assistant answered but hedged. Hedging usually means the underlying answer is unclear internally, not just to the system.

That last one is worth dwelling on. When an assistant gives a vague answer, the cause is frequently that the property itself does not have a settled position. The assistant is not failing; it is reflecting an ambiguity that was always there and that staff had been papering over individually.

The uncomfortable benefit

A knowledge base is an unusually honest audit of what a property actually knows about itself.

Building one surfaces questions nobody had a firm answer to. Maintaining one surfaces every change that was never properly communicated. Reading its escalations surfaces the gap between what guests ask and what the operation is prepared to answer.

Most properties find that uncomfortable at first and useful within a month. The assistant is not the only thing that was working from out-of-date information - it is just the only thing that says so out loud, at scale, to guests, until someone fixes it.