Group Bookings, Accessibility and the Requests You Must Not Automate

The most important configuration decision in a guest messaging deployment is not what the assistant answers. It is what it must never answer on its own. Most properties skip the decision, and the one incident that ends a deployment is almost always on that list.

The principle

Automate where a wrong answer is cheap and a fast answer is valuable. Escalate where a wrong answer causes harm that no speed repays.

Check-in time, parking, breakfast hours, Wi-Fi: wrong is mildly annoying, fast is genuinely useful. Automate.

Whether the room a guest with a mobility need has booked is actually step-free: wrong is a guest standing at the door with luggage and no room they can use. Fast does not help. Escalate.

The list

Every property's list differs at the edges. The core is the same almost everywhere.

Accessibility requests. The knowledge base knows which rooms are step-free. It does not know whether one is free on the guest's dates, whether the lift is working this week, or whether the guest's specific need - a roll-in shower, a hearing loop, a room near the lift - matches what the room actually has. An assistant answering probably yes to a guest who needs certainly yes is the most damaging thing it can do. A person checks, holds and confirms.

Complaints. Tone at a difficult moment is where automation is weakest and where the guest remembers most. The assistant should recognise frustration and hand over, with the original message preserved so the person picking up sees exactly what was said.

Group bookings above a threshold. Above whatever number the property sets - often ten or twenty - pricing, holds and terms need a person. The assistant takes the details and promises a call, rather than quoting.

Anything with legal or financial consequence. Cancellation disputes, damage charges, refund requests, invoice queries. A confident wrong answer here becomes a commitment.

Medical and safety. Allergies, medical equipment, emergencies. Route immediately, every time.

Anything the property has decided is sensitive locally. Every property has one or two - a neighbourhood question, a policy with history. The owner knows which.

Why it is a management decision

This list is not a configuration. It is a statement of what the property is willing to have said in its name without a person checking. That is a management call, it takes twenty minutes, and it should be made by whoever would answer for the incident.

Once made, the assistant honours it: recognises the topic, tells the guest clearly that a person will handle it and when, and routes the full request to whoever owns that topic on a channel they watch. The guest is handed to someone, not dropped.

Writing the rule so it survives

Complex things go to the manager fails twice. Nothing defines complex, and the manager is not on shift at 11pm on a Saturday.

A rule that survives names four things: the topic, the owner as a role, the channel that role watches during the hours they watch it, and what happens if nobody responds within a set time. Accessibility requests go to the duty manager on the operations WhatsApp group; if unacknowledged in fifteen minutes, to the general manager.

Test it deliberately

Before go-live, send the assistant each item on the list, phrased as a guest would. A frustrated message. A request for a step-free room. A booking for eighteen. A question about a refund.

Watch what happens - not the reply, the path. Where did the notification land, who saw it, how long until a person acted. If any of them got an answer instead of a handover, the list is not yet enforced.

The counter-intuitive result

Properties that write this list tend to automate more, not less. Having decided what the assistant will never touch, they can let it handle everything else with confidence - and the share of conversations resolved without a person rises, because staff stop second-guessing every reply.

The deployments that end badly are the ones where nothing was excluded, the assistant handled everything with equal confidence, and the one incident that mattered was the one it should have handed over.

What to do

Write the list. Name an owner and a channel for each item. Send the assistant each one and watch the path. Twenty minutes to write, an hour to test, and it is the difference between a deployment that runs for years and one that ends with a single conversation.